diff --git a/docs/source/en/_toctree.yml b/docs/source/en/_toctree.yml index 2faabfec30ce..01c431118ffe 100644 --- a/docs/source/en/_toctree.yml +++ b/docs/source/en/_toctree.yml @@ -252,6 +252,8 @@ title: SD3ControlNetModel - local: api/models/controlnet_sparsectrl title: SparseControlNetModel + - local: api/models/controlnet_union + title: ControlNetUnionModel title: ControlNets - sections: - local: api/models/allegro_transformer3d @@ -366,6 +368,8 @@ title: ControlNet-XS - local: api/pipelines/controlnetxs_sdxl title: ControlNet-XS with Stable Diffusion XL + - local: api/pipelines/controlnet_union + title: ControlNetUnion - local: api/pipelines/dance_diffusion title: Dance Diffusion - local: api/pipelines/ddim diff --git a/docs/source/en/api/models/controlnet_union.md b/docs/source/en/api/models/controlnet_union.md new file mode 100644 index 000000000000..9c0d86984549 --- /dev/null +++ b/docs/source/en/api/models/controlnet_union.md @@ -0,0 +1,35 @@ + + +# ControlNetUnionModel + +ControlNetUnionModel is an implementation of ControlNet for Stable Diffusion XL. + +The ControlNet model was introduced in [ControlNetPlus](https://github.com/xinsir6/ControlNetPlus) by xinsir6. It supports multiple conditioning inputs without increasing computation. + +*We design a new architecture that can support 10+ control types in condition text-to-image generation and can generate high resolution images visually comparable with midjourney. The network is based on the original ControlNet architecture, we propose two new modules to: 1 Extend the original ControlNet to support different image conditions using the same network parameter. 2 Support multiple conditions input without increasing computation offload, which is especially important for designers who want to edit image in detail, different conditions use the same condition encoder, without adding extra computations or parameters.* + +## Loading + +By default the [`ControlNetUnionModel`] should be loaded with [`~ModelMixin.from_pretrained`]. + +```py +from diffusers import StableDiffusionXLControlNetUnionPipeline, ControlNetUnionModel + +controlnet = ControlNetUnionModel.from_pretrained("xinsir/controlnet-union-sdxl-1.0") +pipe = StableDiffusionXLControlNetUnionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", controlnet=controlnet) +``` + +## ControlNetUnionModel + +[[autodoc]] ControlNetUnionModel + diff --git a/docs/source/en/api/pipelines/controlnet_union.md b/docs/source/en/api/pipelines/controlnet_union.md new file mode 100644 index 000000000000..147b2cd3e0d9 --- /dev/null +++ b/docs/source/en/api/pipelines/controlnet_union.md @@ -0,0 +1,35 @@ + + +# ControlNetUnion + +ControlNetUnionModel is an implementation of ControlNet for Stable Diffusion XL. + +The ControlNet model was introduced in [ControlNetPlus](https://github.com/xinsir6/ControlNetPlus) by xinsir6. It supports multiple conditioning inputs without increasing computation. + +*We design a new architecture that can support 10+ control types in condition text-to-image generation and can generate high resolution images visually comparable with midjourney. The network is based on the original ControlNet architecture, we propose two new modules to: 1 Extend the original ControlNet to support different image conditions using the same network parameter. 2 Support multiple conditions input without increasing computation offload, which is especially important for designers who want to edit image in detail, different conditions use the same condition encoder, without adding extra computations or parameters.* + + +## StableDiffusionXLControlNetUnionPipeline +[[autodoc]] StableDiffusionXLControlNetUnionPipeline + - all + - __call__ + +## StableDiffusionXLControlNetUnionImg2ImgPipeline +[[autodoc]] StableDiffusionXLControlNetUnionImg2ImgPipeline + - all + - __call__ + +## StableDiffusionXLControlNetUnionInpaintPipeline +[[autodoc]] StableDiffusionXLControlNetUnionInpaintPipeline + - all + - __call__ diff --git a/src/diffusers/__init__.py b/src/diffusers/__init__.py index db46dc1d8801..e4f8b1bff7af 100644 --- a/src/diffusers/__init__.py +++ b/src/diffusers/__init__.py @@ -91,6 +91,7 @@ "CogView3PlusTransformer2DModel", "ConsistencyDecoderVAE", "ControlNetModel", + "ControlNetUnionModel", "ControlNetXSAdapter", "DiTTransformer2DModel", "FluxControlNetModel", @@ -376,6 +377,9 @@ "StableDiffusionXLControlNetPAGImg2ImgPipeline", "StableDiffusionXLControlNetPAGPipeline", "StableDiffusionXLControlNetPipeline", + "StableDiffusionXLControlNetUnionImg2ImgPipeline", + "StableDiffusionXLControlNetUnionInpaintPipeline", + "StableDiffusionXLControlNetUnionPipeline", "StableDiffusionXLControlNetXSPipeline", "StableDiffusionXLImg2ImgPipeline", "StableDiffusionXLInpaintPipeline", @@ -583,6 +587,7 @@ CogView3PlusTransformer2DModel, ConsistencyDecoderVAE, ControlNetModel, + ControlNetUnionModel, ControlNetXSAdapter, DiTTransformer2DModel, FluxControlNetModel, @@ -846,6 +851,9 @@ StableDiffusionXLControlNetPAGImg2ImgPipeline, StableDiffusionXLControlNetPAGPipeline, StableDiffusionXLControlNetPipeline, + StableDiffusionXLControlNetUnionImg2ImgPipeline, + StableDiffusionXLControlNetUnionInpaintPipeline, + StableDiffusionXLControlNetUnionPipeline, StableDiffusionXLControlNetXSPipeline, StableDiffusionXLImg2ImgPipeline, StableDiffusionXLInpaintPipeline, diff --git a/src/diffusers/models/__init__.py b/src/diffusers/models/__init__.py index 65e2418ac794..65b6acd094d1 100644 --- a/src/diffusers/models/__init__.py +++ b/src/diffusers/models/__init__.py @@ -44,6 +44,7 @@ ] _import_structure["controlnets.controlnet_sd3"] = ["SD3ControlNetModel", "SD3MultiControlNetModel"] _import_structure["controlnets.controlnet_sparsectrl"] = ["SparseControlNetModel"] + _import_structure["controlnets.controlnet_union"] = ["ControlNetUnionModel"] _import_structure["controlnets.controlnet_xs"] = ["ControlNetXSAdapter", "UNetControlNetXSModel"] _import_structure["controlnets.multicontrolnet"] = ["MultiControlNetModel"] _import_structure["embeddings"] = ["ImageProjection"] @@ -100,6 +101,7 @@ ) from .controlnets import ( ControlNetModel, + ControlNetUnionModel, ControlNetXSAdapter, FluxControlNetModel, FluxMultiControlNetModel, diff --git a/src/diffusers/models/controlnets/__init__.py b/src/diffusers/models/controlnets/__init__.py index 3e4b3561e839..c558c40be375 100644 --- a/src/diffusers/models/controlnets/__init__.py +++ b/src/diffusers/models/controlnets/__init__.py @@ -15,6 +15,7 @@ SparseControlNetModel, SparseControlNetOutput, ) + from .controlnet_union import ControlNetUnionInput, ControlNetUnionInputProMax, ControlNetUnionModel from .controlnet_xs import ControlNetXSAdapter, ControlNetXSOutput, UNetControlNetXSModel from .multicontrolnet import MultiControlNetModel diff --git a/src/diffusers/models/controlnets/controlnet_union.py b/src/diffusers/models/controlnets/controlnet_union.py new file mode 100644 index 000000000000..076629200eac --- /dev/null +++ b/src/diffusers/models/controlnets/controlnet_union.py @@ -0,0 +1,917 @@ +# Copyright 2024 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from dataclasses import dataclass +from typing import Any, Dict, List, Optional, Tuple, Union + +import torch +from torch import nn + +from ...configuration_utils import ConfigMixin, register_to_config +from ...image_processor import PipelineImageInput +from ...loaders.single_file_model import FromOriginalModelMixin +from ...utils import logging +from ..attention_processor import ( + ADDED_KV_ATTENTION_PROCESSORS, + CROSS_ATTENTION_PROCESSORS, + AttentionProcessor, + AttnAddedKVProcessor, + AttnProcessor, +) +from ..embeddings import TextImageTimeEmbedding, TextTimeEmbedding, TimestepEmbedding, Timesteps +from ..modeling_utils import ModelMixin +from ..unets.unet_2d_blocks import ( + CrossAttnDownBlock2D, + DownBlock2D, + UNetMidBlock2DCrossAttn, + get_down_block, +) +from ..unets.unet_2d_condition import UNet2DConditionModel +from .controlnet import ControlNetConditioningEmbedding, ControlNetOutput, zero_module + + +@dataclass +class ControlNetUnionInput: + """ + The image input of [`ControlNetUnionModel`]: + + - 0: openpose + - 1: depth + - 2: hed/pidi/scribble/ted + - 3: canny/lineart/anime_lineart/mlsd + - 4: normal + - 5: segment + """ + + openpose: Optional[PipelineImageInput] = None + depth: Optional[PipelineImageInput] = None + hed: Optional[PipelineImageInput] = None + canny: Optional[PipelineImageInput] = None + normal: Optional[PipelineImageInput] = None + segment: Optional[PipelineImageInput] = None + + def __len__(self) -> int: + return len(vars(self)) + + def __iter__(self): + return iter(vars(self)) + + def __getitem__(self, key): + return getattr(self, key) + + def __setitem__(self, key, value): + setattr(self, key, value) + + +@dataclass +class ControlNetUnionInputProMax: + """ + The image input of [`ControlNetUnionModel`]: + + - 0: openpose + - 1: depth + - 2: hed/pidi/scribble/ted + - 3: canny/lineart/anime_lineart/mlsd + - 4: normal + - 5: segment + - 6: tile + - 7: repaint + """ + + openpose: Optional[PipelineImageInput] = None + depth: Optional[PipelineImageInput] = None + hed: Optional[PipelineImageInput] = None + canny: Optional[PipelineImageInput] = None + normal: Optional[PipelineImageInput] = None + segment: Optional[PipelineImageInput] = None + tile: Optional[PipelineImageInput] = None + repaint: Optional[PipelineImageInput] = None + + def __len__(self) -> int: + return len(vars(self)) + + def __iter__(self): + return iter(vars(self)) + + def __getitem__(self, key): + return getattr(self, key) + + def __setitem__(self, key, value): + setattr(self, key, value) + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class QuickGELU(nn.Module): + """ + Applies GELU approximation that is fast but somewhat inaccurate. See: https://github.com/hendrycks/GELUs + """ + + def forward(self, input: torch.Tensor) -> torch.Tensor: + return input * torch.sigmoid(1.702 * input) + + +class ResidualAttentionMlp(nn.Module): + def __init__(self, d_model: int): + super().__init__() + self.c_fc = nn.Linear(d_model, d_model * 4) + self.gelu = QuickGELU() + self.c_proj = nn.Linear(d_model * 4, d_model) + + def forward(self, x: torch.Tensor): + x = self.c_fc(x) + x = self.gelu(x) + x = self.c_proj(x) + return x + + +class ResidualAttentionBlock(nn.Module): + def __init__(self, d_model: int, n_head: int, attn_mask: torch.Tensor = None): + super().__init__() + self.attn = nn.MultiheadAttention(d_model, n_head) + self.ln_1 = nn.LayerNorm(d_model) + self.mlp = ResidualAttentionMlp(d_model) + self.ln_2 = nn.LayerNorm(d_model) + self.attn_mask = attn_mask + + def attention(self, x: torch.Tensor): + self.attn_mask = self.attn_mask.to(dtype=x.dtype, device=x.device) if self.attn_mask is not None else None + return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0] + + def forward(self, x: torch.Tensor): + x = x + self.attention(self.ln_1(x)) + x = x + self.mlp(self.ln_2(x)) + return x + + +class ControlNetUnionModel(ModelMixin, ConfigMixin, FromOriginalModelMixin): + """ + A ControlNetUnion model. + + Args: + in_channels (`int`, defaults to 4): + The number of channels in the input sample. + flip_sin_to_cos (`bool`, defaults to `True`): + Whether to flip the sin to cos in the time embedding. + freq_shift (`int`, defaults to 0): + The frequency shift to apply to the time embedding. + down_block_types (`tuple[str]`, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`): + The tuple of downsample blocks to use. + only_cross_attention (`Union[bool, Tuple[bool]]`, defaults to `False`): + block_out_channels (`tuple[int]`, defaults to `(320, 640, 1280, 1280)`): + The tuple of output channels for each block. + layers_per_block (`int`, defaults to 2): + The number of layers per block. + downsample_padding (`int`, defaults to 1): + The padding to use for the downsampling convolution. + mid_block_scale_factor (`float`, defaults to 1): + The scale factor to use for the mid block. + act_fn (`str`, defaults to "silu"): + The activation function to use. + norm_num_groups (`int`, *optional*, defaults to 32): + The number of groups to use for the normalization. If None, normalization and activation layers is skipped + in post-processing. + norm_eps (`float`, defaults to 1e-5): + The epsilon to use for the normalization. + cross_attention_dim (`int`, defaults to 1280): + The dimension of the cross attention features. + transformer_layers_per_block (`int` or `Tuple[int]`, *optional*, defaults to 1): + The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for + [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`], + [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`]. + encoder_hid_dim (`int`, *optional*, defaults to None): + If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim` + dimension to `cross_attention_dim`. + encoder_hid_dim_type (`str`, *optional*, defaults to `None`): + If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text + embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`. + attention_head_dim (`Union[int, Tuple[int]]`, defaults to 8): + The dimension of the attention heads. + use_linear_projection (`bool`, defaults to `False`): + class_embed_type (`str`, *optional*, defaults to `None`): + The type of class embedding to use which is ultimately summed with the time embeddings. Choose from None, + `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`. + addition_embed_type (`str`, *optional*, defaults to `None`): + Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or + "text". "text" will use the `TextTimeEmbedding` layer. + num_class_embeds (`int`, *optional*, defaults to 0): + Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing + class conditioning with `class_embed_type` equal to `None`. + upcast_attention (`bool`, defaults to `False`): + resnet_time_scale_shift (`str`, defaults to `"default"`): + Time scale shift config for ResNet blocks (see `ResnetBlock2D`). Choose from `default` or `scale_shift`. + projection_class_embeddings_input_dim (`int`, *optional*, defaults to `None`): + The dimension of the `class_labels` input when `class_embed_type="projection"`. Required when + `class_embed_type="projection"`. + controlnet_conditioning_channel_order (`str`, defaults to `"rgb"`): + The channel order of conditional image. Will convert to `rgb` if it's `bgr`. + conditioning_embedding_out_channels (`tuple[int]`, *optional*, defaults to `(48, 96, 192, 384)`): + The tuple of output channel for each block in the `conditioning_embedding` layer. + global_pool_conditions (`bool`, defaults to `False`): + """ + + _supports_gradient_checkpointing = True + + @register_to_config + def __init__( + self, + in_channels: int = 4, + conditioning_channels: int = 3, + flip_sin_to_cos: bool = True, + freq_shift: int = 0, + down_block_types: Tuple[str, ...] = ( + "CrossAttnDownBlock2D", + "CrossAttnDownBlock2D", + "CrossAttnDownBlock2D", + "DownBlock2D", + ), + only_cross_attention: Union[bool, Tuple[bool]] = False, + block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280), + layers_per_block: int = 2, + downsample_padding: int = 1, + mid_block_scale_factor: float = 1, + act_fn: str = "silu", + norm_num_groups: Optional[int] = 32, + norm_eps: float = 1e-5, + cross_attention_dim: int = 1280, + transformer_layers_per_block: Union[int, Tuple[int, ...]] = 1, + encoder_hid_dim: Optional[int] = None, + encoder_hid_dim_type: Optional[str] = None, + attention_head_dim: Union[int, Tuple[int, ...]] = 8, + num_attention_heads: Optional[Union[int, Tuple[int, ...]]] = None, + use_linear_projection: bool = False, + class_embed_type: Optional[str] = None, + addition_embed_type: Optional[str] = None, + addition_time_embed_dim: Optional[int] = None, + num_class_embeds: Optional[int] = None, + upcast_attention: bool = False, + resnet_time_scale_shift: str = "default", + projection_class_embeddings_input_dim: Optional[int] = None, + controlnet_conditioning_channel_order: str = "rgb", + conditioning_embedding_out_channels: Optional[Tuple[int, ...]] = (48, 96, 192, 384), + global_pool_conditions: bool = False, + addition_embed_type_num_heads: int = 64, + num_control_type: int = 6, + num_trans_channel: int = 320, + num_trans_head: int = 8, + num_trans_layer: int = 1, + num_proj_channel: int = 320, + ): + super().__init__() + + # If `num_attention_heads` is not defined (which is the case for most models) + # it will default to `attention_head_dim`. This looks weird upon first reading it and it is. + # The reason for this behavior is to correct for incorrectly named variables that were introduced + # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131 + # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking + # which is why we correct for the naming here. + num_attention_heads = num_attention_heads or attention_head_dim + + # Check inputs + if len(block_out_channels) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}." + ) + + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types) + + # input + conv_in_kernel = 3 + conv_in_padding = (conv_in_kernel - 1) // 2 + self.conv_in = nn.Conv2d( + in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding + ) + + # time + time_embed_dim = block_out_channels[0] * 4 + self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift) + timestep_input_dim = block_out_channels[0] + self.time_embedding = TimestepEmbedding( + timestep_input_dim, + time_embed_dim, + act_fn=act_fn, + ) + + if encoder_hid_dim_type is not None: + raise ValueError(f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None.") + else: + self.encoder_hid_proj = None + + # class embedding + if class_embed_type is None and num_class_embeds is not None: + self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim) + elif class_embed_type == "timestep": + self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim) + elif class_embed_type == "identity": + self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim) + elif class_embed_type == "projection": + if projection_class_embeddings_input_dim is None: + raise ValueError( + "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set" + ) + # The projection `class_embed_type` is the same as the timestep `class_embed_type` except + # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings + # 2. it projects from an arbitrary input dimension. + # + # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations. + # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings. + # As a result, `TimestepEmbedding` can be passed arbitrary vectors. + self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim) + else: + self.class_embedding = None + + if addition_embed_type == "text": + if encoder_hid_dim is not None: + text_time_embedding_from_dim = encoder_hid_dim + else: + text_time_embedding_from_dim = cross_attention_dim + + self.add_embedding = TextTimeEmbedding( + text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads + ) + elif addition_embed_type == "text_image": + # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much + # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use + # case when `addition_embed_type == "text_image"` (Kandinsky 2.1)` + self.add_embedding = TextImageTimeEmbedding( + text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim + ) + elif addition_embed_type == "text_time": + self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift) + self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim) + + elif addition_embed_type is not None: + raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.") + + # control net conditioning embedding + self.controlnet_cond_embedding = ControlNetConditioningEmbedding( + conditioning_embedding_channels=block_out_channels[0], + block_out_channels=conditioning_embedding_out_channels, + conditioning_channels=conditioning_channels, + ) + + task_scale_factor = num_trans_channel**0.5 + self.task_embedding = nn.Parameter(task_scale_factor * torch.randn(num_control_type, num_trans_channel)) + self.transformer_layes = nn.ModuleList( + [ResidualAttentionBlock(num_trans_channel, num_trans_head) for _ in range(num_trans_layer)] + ) + self.spatial_ch_projs = zero_module(nn.Linear(num_trans_channel, num_proj_channel)) + self.control_type_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift) + self.control_add_embedding = TimestepEmbedding(addition_time_embed_dim * num_control_type, time_embed_dim) + + self.down_blocks = nn.ModuleList([]) + self.controlnet_down_blocks = nn.ModuleList([]) + + if isinstance(only_cross_attention, bool): + only_cross_attention = [only_cross_attention] * len(down_block_types) + + if isinstance(attention_head_dim, int): + attention_head_dim = (attention_head_dim,) * len(down_block_types) + + if isinstance(num_attention_heads, int): + num_attention_heads = (num_attention_heads,) * len(down_block_types) + + # down + output_channel = block_out_channels[0] + + controlnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1) + controlnet_block = zero_module(controlnet_block) + self.controlnet_down_blocks.append(controlnet_block) + + for i, down_block_type in enumerate(down_block_types): + input_channel = output_channel + output_channel = block_out_channels[i] + is_final_block = i == len(block_out_channels) - 1 + + down_block = get_down_block( + down_block_type, + num_layers=layers_per_block, + transformer_layers_per_block=transformer_layers_per_block[i], + in_channels=input_channel, + out_channels=output_channel, + temb_channels=time_embed_dim, + add_downsample=not is_final_block, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + resnet_groups=norm_num_groups, + cross_attention_dim=cross_attention_dim, + num_attention_heads=num_attention_heads[i], + attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel, + downsample_padding=downsample_padding, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention[i], + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + self.down_blocks.append(down_block) + + for _ in range(layers_per_block): + controlnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1) + controlnet_block = zero_module(controlnet_block) + self.controlnet_down_blocks.append(controlnet_block) + + if not is_final_block: + controlnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1) + controlnet_block = zero_module(controlnet_block) + self.controlnet_down_blocks.append(controlnet_block) + + # mid + mid_block_channel = block_out_channels[-1] + + controlnet_block = nn.Conv2d(mid_block_channel, mid_block_channel, kernel_size=1) + controlnet_block = zero_module(controlnet_block) + self.controlnet_mid_block = controlnet_block + + self.mid_block = UNetMidBlock2DCrossAttn( + transformer_layers_per_block=transformer_layers_per_block[-1], + in_channels=mid_block_channel, + temb_channels=time_embed_dim, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + output_scale_factor=mid_block_scale_factor, + resnet_time_scale_shift=resnet_time_scale_shift, + cross_attention_dim=cross_attention_dim, + num_attention_heads=num_attention_heads[-1], + resnet_groups=norm_num_groups, + use_linear_projection=use_linear_projection, + upcast_attention=upcast_attention, + ) + + @classmethod + def from_unet( + cls, + unet: UNet2DConditionModel, + controlnet_conditioning_channel_order: str = "rgb", + conditioning_embedding_out_channels: Optional[Tuple[int, ...]] = (16, 32, 96, 256), + load_weights_from_unet: bool = True, + ): + r""" + Instantiate a [`ControlNetUnionModel`] from [`UNet2DConditionModel`]. + + Parameters: + unet (`UNet2DConditionModel`): + The UNet model weights to copy to the [`ControlNetUnionModel`]. All configuration options are also + copied where applicable. + """ + transformer_layers_per_block = ( + unet.config.transformer_layers_per_block if "transformer_layers_per_block" in unet.config else 1 + ) + encoder_hid_dim = unet.config.encoder_hid_dim if "encoder_hid_dim" in unet.config else None + encoder_hid_dim_type = unet.config.encoder_hid_dim_type if "encoder_hid_dim_type" in unet.config else None + addition_embed_type = unet.config.addition_embed_type if "addition_embed_type" in unet.config else None + addition_time_embed_dim = ( + unet.config.addition_time_embed_dim if "addition_time_embed_dim" in unet.config else None + ) + + controlnet = cls( + encoder_hid_dim=encoder_hid_dim, + encoder_hid_dim_type=encoder_hid_dim_type, + addition_embed_type=addition_embed_type, + addition_time_embed_dim=addition_time_embed_dim, + transformer_layers_per_block=transformer_layers_per_block, + in_channels=unet.config.in_channels, + flip_sin_to_cos=unet.config.flip_sin_to_cos, + freq_shift=unet.config.freq_shift, + down_block_types=unet.config.down_block_types, + only_cross_attention=unet.config.only_cross_attention, + block_out_channels=unet.config.block_out_channels, + layers_per_block=unet.config.layers_per_block, + downsample_padding=unet.config.downsample_padding, + mid_block_scale_factor=unet.config.mid_block_scale_factor, + act_fn=unet.config.act_fn, + norm_num_groups=unet.config.norm_num_groups, + norm_eps=unet.config.norm_eps, + cross_attention_dim=unet.config.cross_attention_dim, + attention_head_dim=unet.config.attention_head_dim, + num_attention_heads=unet.config.num_attention_heads, + use_linear_projection=unet.config.use_linear_projection, + class_embed_type=unet.config.class_embed_type, + num_class_embeds=unet.config.num_class_embeds, + upcast_attention=unet.config.upcast_attention, + resnet_time_scale_shift=unet.config.resnet_time_scale_shift, + projection_class_embeddings_input_dim=unet.config.projection_class_embeddings_input_dim, + controlnet_conditioning_channel_order=controlnet_conditioning_channel_order, + conditioning_embedding_out_channels=conditioning_embedding_out_channels, + ) + + if load_weights_from_unet: + controlnet.conv_in.load_state_dict(unet.conv_in.state_dict()) + controlnet.time_proj.load_state_dict(unet.time_proj.state_dict()) + controlnet.time_embedding.load_state_dict(unet.time_embedding.state_dict()) + + if controlnet.class_embedding: + controlnet.class_embedding.load_state_dict(unet.class_embedding.state_dict()) + + controlnet.down_blocks.load_state_dict(unet.down_blocks.state_dict(), strict=False) + controlnet.mid_block.load_state_dict(unet.mid_block.state_dict(), strict=False) + + return controlnet + + @property + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors + def attn_processors(self) -> Dict[str, AttentionProcessor]: + r""" + Returns: + `dict` of attention processors: A dictionary containing all attention processors used in the model with + indexed by its weight name. + """ + # set recursively + processors = {} + + def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]): + if hasattr(module, "get_processor"): + processors[f"{name}.processor"] = module.get_processor() + + for sub_name, child in module.named_children(): + fn_recursive_add_processors(f"{name}.{sub_name}", child, processors) + + return processors + + for name, module in self.named_children(): + fn_recursive_add_processors(name, module, processors) + + return processors + + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor + def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]): + r""" + Sets the attention processor to use to compute attention. + + Parameters: + processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`): + The instantiated processor class or a dictionary of processor classes that will be set as the processor + for **all** `Attention` layers. + + If `processor` is a dict, the key needs to define the path to the corresponding cross attention + processor. This is strongly recommended when setting trainable attention processors. + + """ + count = len(self.attn_processors.keys()) + + if isinstance(processor, dict) and len(processor) != count: + raise ValueError( + f"A dict of processors was passed, but the number of processors {len(processor)} does not match the" + f" number of attention layers: {count}. Please make sure to pass {count} processor classes." + ) + + def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor): + if hasattr(module, "set_processor"): + if not isinstance(processor, dict): + module.set_processor(processor) + else: + module.set_processor(processor.pop(f"{name}.processor")) + + for sub_name, child in module.named_children(): + fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor) + + for name, module in self.named_children(): + fn_recursive_attn_processor(name, module, processor) + + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor + def set_default_attn_processor(self): + """ + Disables custom attention processors and sets the default attention implementation. + """ + if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()): + processor = AttnAddedKVProcessor() + elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()): + processor = AttnProcessor() + else: + raise ValueError( + f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}" + ) + + self.set_attn_processor(processor) + + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attention_slice + def set_attention_slice(self, slice_size: Union[str, int, List[int]]) -> None: + r""" + Enable sliced attention computation. + + When this option is enabled, the attention module splits the input tensor in slices to compute attention in + several steps. This is useful for saving some memory in exchange for a small decrease in speed. + + Args: + slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`): + When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If + `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is + provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim` + must be a multiple of `slice_size`. + """ + sliceable_head_dims = [] + + def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module): + if hasattr(module, "set_attention_slice"): + sliceable_head_dims.append(module.sliceable_head_dim) + + for child in module.children(): + fn_recursive_retrieve_sliceable_dims(child) + + # retrieve number of attention layers + for module in self.children(): + fn_recursive_retrieve_sliceable_dims(module) + + num_sliceable_layers = len(sliceable_head_dims) + + if slice_size == "auto": + # half the attention head size is usually a good trade-off between + # speed and memory + slice_size = [dim // 2 for dim in sliceable_head_dims] + elif slice_size == "max": + # make smallest slice possible + slice_size = num_sliceable_layers * [1] + + slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size + + if len(slice_size) != len(sliceable_head_dims): + raise ValueError( + f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different" + f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}." + ) + + for i in range(len(slice_size)): + size = slice_size[i] + dim = sliceable_head_dims[i] + if size is not None and size > dim: + raise ValueError(f"size {size} has to be smaller or equal to {dim}.") + + # Recursively walk through all the children. + # Any children which exposes the set_attention_slice method + # gets the message + def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]): + if hasattr(module, "set_attention_slice"): + module.set_attention_slice(slice_size.pop()) + + for child in module.children(): + fn_recursive_set_attention_slice(child, slice_size) + + reversed_slice_size = list(reversed(slice_size)) + for module in self.children(): + fn_recursive_set_attention_slice(module, reversed_slice_size) + + def _set_gradient_checkpointing(self, module, value: bool = False) -> None: + if isinstance(module, (CrossAttnDownBlock2D, DownBlock2D)): + module.gradient_checkpointing = value + + def forward( + self, + sample: torch.Tensor, + timestep: Union[torch.Tensor, float, int], + encoder_hidden_states: torch.Tensor, + controlnet_cond: Union[ControlNetUnionInput, ControlNetUnionInputProMax], + control_type: torch.Tensor, + conditioning_scale: float = 1.0, + class_labels: Optional[torch.Tensor] = None, + timestep_cond: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + guess_mode: bool = False, + return_dict: bool = True, + ) -> Union[ControlNetOutput, Tuple[Tuple[torch.Tensor, ...], torch.Tensor]]: + """ + The [`ControlNetUnionModel`] forward method. + + Args: + sample (`torch.Tensor`): + The noisy input tensor. + timestep (`Union[torch.Tensor, float, int]`): + The number of timesteps to denoise an input. + encoder_hidden_states (`torch.Tensor`): + The encoder hidden states. + controlnet_cond (`Union[ControlNetUnionInput, ControlNetUnionInputProMax]`): + The conditional input tensors. + control_type (`torch.Tensor`): + A tensor of shape `(batch, num_control_type)` with values `0` or `1` depending on whether the control + type is used. + conditioning_scale (`float`, defaults to `1.0`): + The scale factor for ControlNet outputs. + class_labels (`torch.Tensor`, *optional*, defaults to `None`): + Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings. + timestep_cond (`torch.Tensor`, *optional*, defaults to `None`): + Additional conditional embeddings for timestep. If provided, the embeddings will be summed with the + timestep_embedding passed through the `self.time_embedding` layer to obtain the final timestep + embeddings. + attention_mask (`torch.Tensor`, *optional*, defaults to `None`): + An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask + is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large + negative values to the attention scores corresponding to "discard" tokens. + added_cond_kwargs (`dict`): + Additional conditions for the Stable Diffusion XL UNet. + cross_attention_kwargs (`dict[str]`, *optional*, defaults to `None`): + A kwargs dictionary that if specified is passed along to the `AttnProcessor`. + guess_mode (`bool`, defaults to `False`): + In this mode, the ControlNet encoder tries its best to recognize the input content of the input even if + you remove all prompts. A `guidance_scale` between 3.0 and 5.0 is recommended. + return_dict (`bool`, defaults to `True`): + Whether or not to return a [`~models.controlnet.ControlNetOutput`] instead of a plain tuple. + + Returns: + [`~models.controlnet.ControlNetOutput`] **or** `tuple`: + If `return_dict` is `True`, a [`~models.controlnet.ControlNetOutput`] is returned, otherwise a tuple is + returned where the first element is the sample tensor. + """ + if not isinstance(controlnet_cond, (ControlNetUnionInput, ControlNetUnionInputProMax)): + raise ValueError( + "Expected type of `controlnet_cond` to be one of `ControlNetUnionInput` or `ControlNetUnionInputProMax`" + ) + if len(controlnet_cond) != self.config.num_control_type: + if isinstance(controlnet_cond, ControlNetUnionInput): + raise ValueError( + f"Expected num_control_type {self.config.num_control_type}, got {len(controlnet_cond)}. Try `ControlNetUnionInputProMax`." + ) + elif isinstance(controlnet_cond, ControlNetUnionInputProMax): + raise ValueError( + f"Expected num_control_type {self.config.num_control_type}, got {len(controlnet_cond)}. Try `ControlNetUnionInput`." + ) + + # check channel order + channel_order = self.config.controlnet_conditioning_channel_order + + if channel_order != "rgb": + raise ValueError(f"unknown `controlnet_conditioning_channel_order`: {channel_order}") + + # prepare attention_mask + if attention_mask is not None: + attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0 + attention_mask = attention_mask.unsqueeze(1) + + # 1. time + timesteps = timestep + if not torch.is_tensor(timesteps): + # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can + # This would be a good case for the `match` statement (Python 3.10+) + is_mps = sample.device.type == "mps" + if isinstance(timestep, float): + dtype = torch.float32 if is_mps else torch.float64 + else: + dtype = torch.int32 if is_mps else torch.int64 + timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device) + elif len(timesteps.shape) == 0: + timesteps = timesteps[None].to(sample.device) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timesteps = timesteps.expand(sample.shape[0]) + + t_emb = self.time_proj(timesteps) + + # timesteps does not contain any weights and will always return f32 tensors + # but time_embedding might actually be running in fp16. so we need to cast here. + # there might be better ways to encapsulate this. + t_emb = t_emb.to(dtype=sample.dtype) + + emb = self.time_embedding(t_emb, timestep_cond) + aug_emb = None + + if self.class_embedding is not None: + if class_labels is None: + raise ValueError("class_labels should be provided when num_class_embeds > 0") + + if self.config.class_embed_type == "timestep": + class_labels = self.time_proj(class_labels) + + class_emb = self.class_embedding(class_labels).to(dtype=self.dtype) + emb = emb + class_emb + + if self.config.addition_embed_type is not None: + if self.config.addition_embed_type == "text": + aug_emb = self.add_embedding(encoder_hidden_states) + + elif self.config.addition_embed_type == "text_time": + if "text_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`" + ) + text_embeds = added_cond_kwargs.get("text_embeds") + if "time_ids" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`" + ) + time_ids = added_cond_kwargs.get("time_ids") + time_embeds = self.add_time_proj(time_ids.flatten()) + time_embeds = time_embeds.reshape((text_embeds.shape[0], -1)) + + add_embeds = torch.concat([text_embeds, time_embeds], dim=-1) + add_embeds = add_embeds.to(emb.dtype) + aug_emb = self.add_embedding(add_embeds) + + control_embeds = self.control_type_proj(control_type.flatten()) + control_embeds = control_embeds.reshape((t_emb.shape[0], -1)) + control_embeds = control_embeds.to(emb.dtype) + control_emb = self.control_add_embedding(control_embeds) + emb = emb + control_emb + emb = emb + aug_emb if aug_emb is not None else emb + + # 2. pre-process + sample = self.conv_in(sample) + + inputs = [] + condition_list = [] + + for idx, image_type in enumerate(controlnet_cond): + if controlnet_cond[image_type] is None: + continue + condition = self.controlnet_cond_embedding(controlnet_cond[image_type]) + feat_seq = torch.mean(condition, dim=(2, 3)) + feat_seq = feat_seq + self.task_embedding[idx] + inputs.append(feat_seq.unsqueeze(1)) + condition_list.append(condition) + + condition = sample + feat_seq = torch.mean(condition, dim=(2, 3)) + inputs.append(feat_seq.unsqueeze(1)) + condition_list.append(condition) + + x = torch.cat(inputs, dim=1) + for layer in self.transformer_layes: + x = layer(x) + + controlnet_cond_fuser = sample * 0.0 + for idx, condition in enumerate(condition_list[:-1]): + alpha = self.spatial_ch_projs(x[:, idx]) + alpha = alpha.unsqueeze(-1).unsqueeze(-1) + controlnet_cond_fuser += condition + alpha + + sample = sample + controlnet_cond_fuser + + # 3. down + down_block_res_samples = (sample,) + for downsample_block in self.down_blocks: + if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention: + sample, res_samples = downsample_block( + hidden_states=sample, + temb=emb, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + ) + else: + sample, res_samples = downsample_block(hidden_states=sample, temb=emb) + + down_block_res_samples += res_samples + + # 4. mid + if self.mid_block is not None: + sample = self.mid_block( + sample, + emb, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + ) + + # 5. Control net blocks + controlnet_down_block_res_samples = () + + for down_block_res_sample, controlnet_block in zip(down_block_res_samples, self.controlnet_down_blocks): + down_block_res_sample = controlnet_block(down_block_res_sample) + controlnet_down_block_res_samples = controlnet_down_block_res_samples + (down_block_res_sample,) + + down_block_res_samples = controlnet_down_block_res_samples + + mid_block_res_sample = self.controlnet_mid_block(sample) + + # 6. scaling + if guess_mode and not self.config.global_pool_conditions: + scales = torch.logspace(-1, 0, len(down_block_res_samples) + 1, device=sample.device) # 0.1 to 1.0 + scales = scales * conditioning_scale + down_block_res_samples = [sample * scale for sample, scale in zip(down_block_res_samples, scales)] + mid_block_res_sample = mid_block_res_sample * scales[-1] # last one + else: + down_block_res_samples = [sample * conditioning_scale for sample in down_block_res_samples] + mid_block_res_sample = mid_block_res_sample * conditioning_scale + + if self.config.global_pool_conditions: + down_block_res_samples = [ + torch.mean(sample, dim=(2, 3), keepdim=True) for sample in down_block_res_samples + ] + mid_block_res_sample = torch.mean(mid_block_res_sample, dim=(2, 3), keepdim=True) + + if not return_dict: + return (down_block_res_samples, mid_block_res_sample) + + return ControlNetOutput( + down_block_res_samples=down_block_res_samples, mid_block_res_sample=mid_block_res_sample + ) diff --git a/src/diffusers/pipelines/__init__.py b/src/diffusers/pipelines/__init__.py index 6d3a20511696..5f2ae02023f4 100644 --- a/src/diffusers/pipelines/__init__.py +++ b/src/diffusers/pipelines/__init__.py @@ -162,6 +162,9 @@ "StableDiffusionXLControlNetImg2ImgPipeline", "StableDiffusionXLControlNetInpaintPipeline", "StableDiffusionXLControlNetPipeline", + "StableDiffusionXLControlNetUnionPipeline", + "StableDiffusionXLControlNetUnionInpaintPipeline", + "StableDiffusionXLControlNetUnionImg2ImgPipeline", ] ) _import_structure["pag"].extend( @@ -495,6 +498,9 @@ StableDiffusionXLControlNetImg2ImgPipeline, StableDiffusionXLControlNetInpaintPipeline, StableDiffusionXLControlNetPipeline, + StableDiffusionXLControlNetUnionImg2ImgPipeline, + StableDiffusionXLControlNetUnionInpaintPipeline, + StableDiffusionXLControlNetUnionPipeline, ) from .controlnet_hunyuandit import ( HunyuanDiTControlNetPipeline, diff --git a/src/diffusers/pipelines/controlnet/__init__.py b/src/diffusers/pipelines/controlnet/__init__.py index b1671050c93f..a49dccf235a3 100644 --- a/src/diffusers/pipelines/controlnet/__init__.py +++ b/src/diffusers/pipelines/controlnet/__init__.py @@ -1,80 +1,86 @@ -from typing import TYPE_CHECKING - -from ...utils import ( - DIFFUSERS_SLOW_IMPORT, - OptionalDependencyNotAvailable, - _LazyModule, - get_objects_from_module, - is_flax_available, - is_torch_available, - is_transformers_available, -) - - -_dummy_objects = {} -_import_structure = {} - -try: - if not (is_transformers_available() and is_torch_available()): - raise OptionalDependencyNotAvailable() -except OptionalDependencyNotAvailable: - from ...utils import dummy_torch_and_transformers_objects # noqa F403 - - _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) -else: - _import_structure["multicontrolnet"] = ["MultiControlNetModel"] - _import_structure["pipeline_controlnet"] = ["StableDiffusionControlNetPipeline"] - _import_structure["pipeline_controlnet_blip_diffusion"] = ["BlipDiffusionControlNetPipeline"] - _import_structure["pipeline_controlnet_img2img"] = ["StableDiffusionControlNetImg2ImgPipeline"] - _import_structure["pipeline_controlnet_inpaint"] = ["StableDiffusionControlNetInpaintPipeline"] - _import_structure["pipeline_controlnet_inpaint_sd_xl"] = ["StableDiffusionXLControlNetInpaintPipeline"] - _import_structure["pipeline_controlnet_sd_xl"] = ["StableDiffusionXLControlNetPipeline"] - _import_structure["pipeline_controlnet_sd_xl_img2img"] = ["StableDiffusionXLControlNetImg2ImgPipeline"] -try: - if not (is_transformers_available() and is_flax_available()): - raise OptionalDependencyNotAvailable() -except OptionalDependencyNotAvailable: - from ...utils import dummy_flax_and_transformers_objects # noqa F403 - - _dummy_objects.update(get_objects_from_module(dummy_flax_and_transformers_objects)) -else: - _import_structure["pipeline_flax_controlnet"] = ["FlaxStableDiffusionControlNetPipeline"] - - -if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: - try: - if not (is_transformers_available() and is_torch_available()): - raise OptionalDependencyNotAvailable() - - except OptionalDependencyNotAvailable: - from ...utils.dummy_torch_and_transformers_objects import * - else: - from .multicontrolnet import MultiControlNetModel - from .pipeline_controlnet import StableDiffusionControlNetPipeline - from .pipeline_controlnet_blip_diffusion import BlipDiffusionControlNetPipeline - from .pipeline_controlnet_img2img import StableDiffusionControlNetImg2ImgPipeline - from .pipeline_controlnet_inpaint import StableDiffusionControlNetInpaintPipeline - from .pipeline_controlnet_inpaint_sd_xl import StableDiffusionXLControlNetInpaintPipeline - from .pipeline_controlnet_sd_xl import StableDiffusionXLControlNetPipeline - from .pipeline_controlnet_sd_xl_img2img import StableDiffusionXLControlNetImg2ImgPipeline - - try: - if not (is_transformers_available() and is_flax_available()): - raise OptionalDependencyNotAvailable() - except OptionalDependencyNotAvailable: - from ...utils.dummy_flax_and_transformers_objects import * # noqa F403 - else: - from .pipeline_flax_controlnet import FlaxStableDiffusionControlNetPipeline - - -else: - import sys - - sys.modules[__name__] = _LazyModule( - __name__, - globals()["__file__"], - _import_structure, - module_spec=__spec__, - ) - for name, value in _dummy_objects.items(): - setattr(sys.modules[__name__], name, value) +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_flax_available, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["multicontrolnet"] = ["MultiControlNetModel"] + _import_structure["pipeline_controlnet"] = ["StableDiffusionControlNetPipeline"] + _import_structure["pipeline_controlnet_blip_diffusion"] = ["BlipDiffusionControlNetPipeline"] + _import_structure["pipeline_controlnet_img2img"] = ["StableDiffusionControlNetImg2ImgPipeline"] + _import_structure["pipeline_controlnet_inpaint"] = ["StableDiffusionControlNetInpaintPipeline"] + _import_structure["pipeline_controlnet_inpaint_sd_xl"] = ["StableDiffusionXLControlNetInpaintPipeline"] + _import_structure["pipeline_controlnet_sd_xl"] = ["StableDiffusionXLControlNetPipeline"] + _import_structure["pipeline_controlnet_sd_xl_img2img"] = ["StableDiffusionXLControlNetImg2ImgPipeline"] + _import_structure["pipeline_controlnet_union_inpaint_sd_xl"] = ["StableDiffusionXLControlNetUnionInpaintPipeline"] + _import_structure["pipeline_controlnet_union_sd_xl"] = ["StableDiffusionXLControlNetUnionPipeline"] + _import_structure["pipeline_controlnet_union_sd_xl_img2img"] = ["StableDiffusionXLControlNetUnionImg2ImgPipeline"] +try: + if not (is_transformers_available() and is_flax_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_flax_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_flax_and_transformers_objects)) +else: + _import_structure["pipeline_flax_controlnet"] = ["FlaxStableDiffusionControlNetPipeline"] + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .multicontrolnet import MultiControlNetModel + from .pipeline_controlnet import StableDiffusionControlNetPipeline + from .pipeline_controlnet_blip_diffusion import BlipDiffusionControlNetPipeline + from .pipeline_controlnet_img2img import StableDiffusionControlNetImg2ImgPipeline + from .pipeline_controlnet_inpaint import StableDiffusionControlNetInpaintPipeline + from .pipeline_controlnet_inpaint_sd_xl import StableDiffusionXLControlNetInpaintPipeline + from .pipeline_controlnet_sd_xl import StableDiffusionXLControlNetPipeline + from .pipeline_controlnet_sd_xl_img2img import StableDiffusionXLControlNetImg2ImgPipeline + from .pipeline_controlnet_union_inpaint_sd_xl import StableDiffusionXLControlNetUnionInpaintPipeline + from .pipeline_controlnet_union_sd_xl import StableDiffusionXLControlNetUnionPipeline + from .pipeline_controlnet_union_sd_xl_img2img import StableDiffusionXLControlNetUnionImg2ImgPipeline + + try: + if not (is_transformers_available() and is_flax_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_flax_and_transformers_objects import * # noqa F403 + else: + from .pipeline_flax_controlnet import FlaxStableDiffusionControlNetPipeline + + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/src/diffusers/pipelines/controlnet/pipeline_controlnet_union_inpaint_sd_xl.py b/src/diffusers/pipelines/controlnet/pipeline_controlnet_union_inpaint_sd_xl.py new file mode 100644 index 000000000000..0465391d7305 --- /dev/null +++ b/src/diffusers/pipelines/controlnet/pipeline_controlnet_union_inpaint_sd_xl.py @@ -0,0 +1,1801 @@ +# Copyright 2024 Harutatsu Akiyama, Jinbin Bai, and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable, Dict, List, Optional, Tuple, Union + +import numpy as np +import PIL.Image +import torch +import torch.nn.functional as F +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionModelWithProjection, +) + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import ( + FromSingleFileMixin, + IPAdapterMixin, + StableDiffusionXLLoraLoaderMixin, + TextualInversionLoaderMixin, +) +from ...models import AutoencoderKL, ControlNetModel, ControlNetUnionModel, ImageProjection, UNet2DConditionModel +from ...models.attention_processor import ( + AttnProcessor2_0, + XFormersAttnProcessor, +) +from ...models.controlnets import ControlNetUnionInput, ControlNetUnionInputProMax +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_invisible_watermark_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import is_compiled_module, randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput + + +if is_invisible_watermark_available(): + from diffusers.pipelines.stable_diffusion_xl.watermark import StableDiffusionXLWatermarker + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + from diffusers import StableDiffusionXLControlNetUnionInpaintPipeline, ControlNetUnionModel, AutoencoderKL + from diffusers.models.controlnets import ControlNetUnionInputProMax + from diffusers.utils import load_image + import torch + import numpy as np + from PIL import Image + + prompt = "A cat" + # download an image + image = load_image( + "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/in_paint/overture-creations-5sI6fQgYIuo.png" + ).resize((1024, 1024)) + mask = load_image( + "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/in_paint/overture-creations-5sI6fQgYIuo_mask.png" + ).resize((1024, 1024)) + # initialize the models and pipeline + controlnet = ControlNetUnionModel.from_pretrained( + "brad-twinkl/controlnet-union-sdxl-1.0-promax", torch_dtype=torch.float16 + ) + vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16) + pipe = StableDiffusionXLControlNetUnionInpaintPipeline.from_pretrained( + "stabilityai/stable-diffusion-xl-base-1.0", + controlnet=controlnet, + vae=vae, + torch_dtype=torch.float16, + variant="fp16", + ) + pipe.enable_model_cpu_offload() + controlnet_img = image.copy() + controlnet_img_np = np.array(controlnet_img) + mask_np = np.array(mask) + controlnet_img_np[mask_np > 0] = 0 + controlnet_img = Image.fromarray(controlnet_img_np) + union_input = ControlNetUnionInputProMax( + repaint=controlnet_img, + ) + # generate image + image = pipe(prompt, image=image, mask_image=mask, control_image_list=union_input).images[0] + image.save("inpaint.png") + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://arxiv.org/pdf/2305.08891.pdf). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +class StableDiffusionXLControlNetUnionInpaintPipeline( + DiffusionPipeline, + StableDiffusionMixin, + StableDiffusionXLLoraLoaderMixin, + FromSingleFileMixin, + IPAdapterMixin, + TextualInversionLoaderMixin, +): + r""" + Pipeline for text-to-image generation using Stable Diffusion XL. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion XL uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([` CLIPTextModelWithProjection`]): + Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->unet->vae" + + _optional_components = [ + "tokenizer", + "tokenizer_2", + "text_encoder", + "text_encoder_2", + "image_encoder", + "feature_extractor", + ] + _callback_tensor_inputs = [ + "latents", + "prompt_embeds", + "negative_prompt_embeds", + "add_text_embeds", + "add_time_ids", + "negative_pooled_prompt_embeds", + "add_neg_time_ids", + "mask", + "masked_image_latents", + ] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + tokenizer_2: CLIPTokenizer, + unet: UNet2DConditionModel, + controlnet: ControlNetUnionModel, + scheduler: KarrasDiffusionSchedulers, + requires_aesthetics_score: bool = False, + force_zeros_for_empty_prompt: bool = True, + add_watermarker: Optional[bool] = None, + feature_extractor: Optional[CLIPImageProcessor] = None, + image_encoder: Optional[CLIPVisionModelWithProjection] = None, + ): + super().__init__() + + if not isinstance(controlnet, ControlNetUnionModel): + raise ValueError("Expected `controlnet` to be of type `ControlNetUnionModel`.") + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + unet=unet, + controlnet=controlnet, + scheduler=scheduler, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.register_to_config(requires_aesthetics_score=requires_aesthetics_score) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.mask_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor, do_normalize=False, do_binarize=True, do_convert_grayscale=True + ) + self.control_image_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True, do_normalize=False + ) + + add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available() + + if add_watermarker: + self.watermark = StableDiffusionXLWatermarker() + else: + self.watermark = None + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt + def encode_prompt( + self, + prompt: str, + prompt_2: Optional[str] = None, + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: Optional[str] = None, + negative_prompt_2: Optional[str] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + pooled_prompt_embeds: Optional[torch.Tensor] = None, + negative_pooled_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale) + else: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] + text_encoders = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # textual inversion: process multi-vector tokens if necessary + prompt_embeds_list = [] + prompts = [prompt, prompt_2] + for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, tokenizer) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + + # We are only ALWAYS interested in the pooled output of the final text encoder + pooled_prompt_embeds = prompt_embeds[0] + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + # "2" because SDXL always indexes from the penultimate layer. + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + + uncond_tokens: List[str] + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = [negative_prompt, negative_prompt_2] + + negative_prompt_embeds_list = [] + for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = tokenizer( + negative_prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + # We are only ALWAYS interested in the pooled output of the final text encoder + negative_pooled_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) + + if self.text_encoder_2 is not None: + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + if self.text_encoder_2 is not None: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.controlnet.pipeline_controlnet_sd_xl.StableDiffusionXLControlNetPipeline.check_image + def check_image(self, image, prompt, prompt_embeds): + image_is_pil = isinstance(image, PIL.Image.Image) + image_is_tensor = isinstance(image, torch.Tensor) + image_is_np = isinstance(image, np.ndarray) + image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image) + image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor) + image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray) + + if ( + not image_is_pil + and not image_is_tensor + and not image_is_np + and not image_is_pil_list + and not image_is_tensor_list + and not image_is_np_list + ): + raise TypeError( + f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}" + ) + + if image_is_pil: + image_batch_size = 1 + else: + image_batch_size = len(image) + + if prompt is not None and isinstance(prompt, str): + prompt_batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + prompt_batch_size = len(prompt) + elif prompt_embeds is not None: + prompt_batch_size = prompt_embeds.shape[0] + + if image_batch_size != 1 and image_batch_size != prompt_batch_size: + raise ValueError( + f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}" + ) + + def check_inputs( + self, + prompt, + prompt_2, + image, + mask_image, + strength, + num_inference_steps, + callback_steps, + output_type, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + controlnet_conditioning_scale=1.0, + control_guidance_start=0.0, + control_guidance_end=1.0, + callback_on_step_end_tensor_inputs=None, + padding_mask_crop=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + if num_inference_steps is None: + raise ValueError("`num_inference_steps` cannot be None.") + elif not isinstance(num_inference_steps, int) or num_inference_steps <= 0: + raise ValueError( + f"`num_inference_steps` has to be a positive integer but is {num_inference_steps} of type" + f" {type(num_inference_steps)}." + ) + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if padding_mask_crop is not None: + if not isinstance(image, PIL.Image.Image): + raise ValueError( + f"The image should be a PIL image when inpainting mask crop, but is of type" f" {type(image)}." + ) + if not isinstance(mask_image, PIL.Image.Image): + raise ValueError( + f"The mask image should be a PIL image when inpainting mask crop, but is of type" + f" {type(mask_image)}." + ) + if output_type != "pil": + raise ValueError(f"The output type should be PIL when inpainting mask crop, but is" f" {output_type}.") + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + # Check `image` + is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance( + self.controlnet, torch._dynamo.eval_frame.OptimizedModule + ) + if ( + isinstance(self.controlnet, ControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, ControlNetModel) + ): + self.check_image(image, prompt, prompt_embeds) + elif ( + isinstance(self.controlnet, ControlNetUnionModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, ControlNetUnionModel) + ): + self.check_image(image, prompt, prompt_embeds) + + else: + assert False + + # Check `controlnet_conditioning_scale` + if ( + isinstance(self.controlnet, ControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, ControlNetModel) + ): + if not isinstance(controlnet_conditioning_scale, float): + raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.") + + elif ( + isinstance(self.controlnet, ControlNetUnionModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, ControlNetUnionModel) + ): + if not isinstance(controlnet_conditioning_scale, float): + raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.") + + else: + assert False + + if not isinstance(control_guidance_start, (tuple, list)): + control_guidance_start = [control_guidance_start] + + if not isinstance(control_guidance_end, (tuple, list)): + control_guidance_end = [control_guidance_end] + + if len(control_guidance_start) != len(control_guidance_end): + raise ValueError( + f"`control_guidance_start` has {len(control_guidance_start)} elements, but `control_guidance_end` has {len(control_guidance_end)} elements. Make sure to provide the same number of elements to each list." + ) + + for start, end in zip(control_guidance_start, control_guidance_end): + if start >= end: + raise ValueError( + f"control guidance start: {start} cannot be larger or equal to control guidance end: {end}." + ) + if start < 0.0: + raise ValueError(f"control guidance start: {start} can't be smaller than 0.") + if end > 1.0: + raise ValueError(f"control guidance end: {end} can't be larger than 1.0.") + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + # Copied from diffusers.pipelines.controlnet.pipeline_controlnet_inpaint_sd_xl.StableDiffusionXLControlNetInpaintPipeline.prepare_control_image + def prepare_control_image( + self, + image, + width, + height, + batch_size, + num_images_per_prompt, + device, + dtype, + crops_coords, + resize_mode, + do_classifier_free_guidance=False, + guess_mode=False, + ): + image = self.control_image_processor.preprocess( + image, height=height, width=width, crops_coords=crops_coords, resize_mode=resize_mode + ).to(dtype=torch.float32) + image_batch_size = image.shape[0] + + if image_batch_size == 1: + repeat_by = batch_size + else: + # image batch size is the same as prompt batch size + repeat_by = num_images_per_prompt + + image = image.repeat_interleave(repeat_by, dim=0) + + image = image.to(device=device, dtype=dtype) + + if do_classifier_free_guidance and not guess_mode: + image = torch.cat([image] * 2) + + return image + + # Copied from diffusers.pipelines.controlnet.pipeline_controlnet_inpaint_sd_xl.StableDiffusionXLControlNetInpaintPipeline.prepare_latents + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + image=None, + timestep=None, + is_strength_max=True, + add_noise=True, + return_noise=False, + return_image_latents=False, + ): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if (image is None or timestep is None) and not is_strength_max: + raise ValueError( + "Since strength < 1. initial latents are to be initialised as a combination of Image + Noise." + "However, either the image or the noise timestep has not been provided." + ) + + if return_image_latents or (latents is None and not is_strength_max): + image = image.to(device=device, dtype=dtype) + + if image.shape[1] == 4: + image_latents = image + else: + image_latents = self._encode_vae_image(image=image, generator=generator) + image_latents = image_latents.repeat(batch_size // image_latents.shape[0], 1, 1, 1) + + if latents is None and add_noise: + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # if strength is 1. then initialise the latents to noise, else initial to image + noise + latents = noise if is_strength_max else self.scheduler.add_noise(image_latents, noise, timestep) + # if pure noise then scale the initial latents by the Scheduler's init sigma + latents = latents * self.scheduler.init_noise_sigma if is_strength_max else latents + elif add_noise: + noise = latents.to(device) + latents = noise * self.scheduler.init_noise_sigma + else: + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = image_latents.to(device) + + outputs = (latents,) + + if return_noise: + outputs += (noise,) + + if return_image_latents: + outputs += (image_latents,) + + return outputs + + # Copied from diffusers.pipelines.controlnet.pipeline_controlnet_inpaint_sd_xl.StableDiffusionXLControlNetInpaintPipeline._encode_vae_image + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + dtype = image.dtype + if self.vae.config.force_upcast: + image = image.float() + self.vae.to(dtype=torch.float32) + + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + if self.vae.config.force_upcast: + self.vae.to(dtype) + + image_latents = image_latents.to(dtype) + image_latents = self.vae.config.scaling_factor * image_latents + + return image_latents + + # Copied from diffusers.pipelines.controlnet.pipeline_controlnet_inpaint_sd_xl.StableDiffusionXLControlNetInpaintPipeline.prepare_mask_latents + def prepare_mask_latents( + self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance + ): + # resize the mask to latents shape as we concatenate the mask to the latents + # we do that before converting to dtype to avoid breaking in case we're using cpu_offload + # and half precision + mask = torch.nn.functional.interpolate( + mask, size=(height // self.vae_scale_factor, width // self.vae_scale_factor) + ) + mask = mask.to(device=device, dtype=dtype) + + # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method + if mask.shape[0] < batch_size: + if not batch_size % mask.shape[0] == 0: + raise ValueError( + "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to" + f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number" + " of masks that you pass is divisible by the total requested batch size." + ) + mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1) + + mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask + + masked_image_latents = None + if masked_image is not None: + masked_image = masked_image.to(device=device, dtype=dtype) + masked_image_latents = self._encode_vae_image(masked_image, generator=generator) + if masked_image_latents.shape[0] < batch_size: + if not batch_size % masked_image_latents.shape[0] == 0: + raise ValueError( + "The passed images and the required batch size don't match. Images are supposed to be duplicated" + f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed." + " Make sure the number of images that you pass is divisible by the total requested batch size." + ) + masked_image_latents = masked_image_latents.repeat( + batch_size // masked_image_latents.shape[0], 1, 1, 1 + ) + + masked_image_latents = ( + torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents + ) + + # aligning device to prevent device errors when concating it with the latent model input + masked_image_latents = masked_image_latents.to(device=device, dtype=dtype) + + return mask, masked_image_latents + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device, denoising_start=None): + # get the original timestep using init_timestep + if denoising_start is None: + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + t_start = max(num_inference_steps - init_timestep, 0) + + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + else: + # Strength is irrelevant if we directly request a timestep to start at; + # that is, strength is determined by the denoising_start instead. + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (denoising_start * self.scheduler.config.num_train_timesteps) + ) + ) + + num_inference_steps = (self.scheduler.timesteps < discrete_timestep_cutoff).sum().item() + if self.scheduler.order == 2 and num_inference_steps % 2 == 0: + # if the scheduler is a 2nd order scheduler we might have to do +1 + # because `num_inference_steps` might be even given that every timestep + # (except the highest one) is duplicated. If `num_inference_steps` is even it would + # mean that we cut the timesteps in the middle of the denoising step + # (between 1st and 2nd derivative) which leads to incorrect results. By adding 1 + # we ensure that the denoising process always ends after the 2nd derivate step of the scheduler + num_inference_steps = num_inference_steps + 1 + + # because t_n+1 >= t_n, we slice the timesteps starting from the end + t_start = len(self.scheduler.timesteps) - num_inference_steps + timesteps = self.scheduler.timesteps[t_start:] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start) + return timesteps, num_inference_steps + + def _get_add_time_ids( + self, + original_size, + crops_coords_top_left, + target_size, + aesthetic_score, + negative_aesthetic_score, + dtype, + text_encoder_projection_dim=None, + ): + if self.config.requires_aesthetics_score: + add_time_ids = list(original_size + crops_coords_top_left + (aesthetic_score,)) + add_neg_time_ids = list(original_size + crops_coords_top_left + (negative_aesthetic_score,)) + else: + add_time_ids = list(original_size + crops_coords_top_left + target_size) + add_neg_time_ids = list(original_size + crops_coords_top_left + target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if ( + expected_add_embed_dim > passed_add_embed_dim + and (expected_add_embed_dim - passed_add_embed_dim) == self.unet.config.addition_time_embed_dim + ): + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to enable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=True)` to make sure `aesthetic_score` {aesthetic_score} and `negative_aesthetic_score` {negative_aesthetic_score} is correctly used by the model." + ) + elif ( + expected_add_embed_dim < passed_add_embed_dim + and (passed_add_embed_dim - expected_add_embed_dim) == self.unet.config.addition_time_embed_dim + ): + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to disable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=False)` to make sure `target_size` {target_size} is correctly used by the model." + ) + elif expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + add_neg_time_ids = torch.tensor([add_neg_time_ids], dtype=dtype) + + return add_time_ids, add_neg_time_ids + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae + def upcast_vae(self): + dtype = self.vae.dtype + self.vae.to(dtype=torch.float32) + use_torch_2_0_or_xformers = isinstance( + self.vae.decoder.mid_block.attentions[0].processor, + ( + AttnProcessor2_0, + XFormersAttnProcessor, + ), + ) + # if xformers or torch_2_0 is used attention block does not need + # to be in float32 which can save lots of memory + if use_torch_2_0_or_xformers: + self.vae.post_quant_conv.to(dtype) + self.vae.decoder.conv_in.to(dtype) + self.vae.decoder.mid_block.to(dtype) + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + prompt_2: Optional[Union[str, List[str]]] = None, + image: PipelineImageInput = None, + mask_image: PipelineImageInput = None, + control_image_list: Union[ControlNetUnionInput, ControlNetUnionInputProMax] = None, + height: Optional[int] = None, + width: Optional[int] = None, + padding_mask_crop: Optional[int] = None, + strength: float = 0.9999, + num_inference_steps: int = 50, + denoising_start: Optional[float] = None, + denoising_end: Optional[float] = None, + guidance_scale: float = 5.0, + negative_prompt: Optional[Union[str, List[str]]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + ip_adapter_image: Optional[PipelineImageInput] = None, + ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None, + pooled_prompt_embeds: Optional[torch.Tensor] = None, + negative_pooled_prompt_embeds: Optional[torch.Tensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + controlnet_conditioning_scale: Union[float, List[float]] = 1.0, + guess_mode: bool = False, + control_guidance_start: Union[float, List[float]] = 0.0, + control_guidance_end: Union[float, List[float]] = 1.0, + guidance_rescale: float = 0.0, + original_size: Tuple[int, int] = None, + crops_coords_top_left: Tuple[int, int] = (0, 0), + target_size: Tuple[int, int] = None, + aesthetic_score: float = 6.0, + negative_aesthetic_score: float = 2.5, + clip_skip: Optional[int] = None, + callback_on_step_end: Optional[ + Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks] + ] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + **kwargs, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + image (`PIL.Image.Image`): + `Image`, or tensor representing an image batch which will be inpainted, *i.e.* parts of the image will + be masked out with `mask_image` and repainted according to `prompt`. + mask_image (`PIL.Image.Image`): + `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be + repainted, while black pixels will be preserved. If `mask_image` is a PIL image, it will be converted + to a single channel (luminance) before use. If it's a tensor, it should contain one color channel (L) + instead of 3, so the expected shape would be `(B, H, W, 1)`. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. + padding_mask_crop (`int`, *optional*, defaults to `None`): + The size of margin in the crop to be applied to the image and masking. If `None`, no crop is applied to + image and mask_image. If `padding_mask_crop` is not `None`, it will first find a rectangular region + with the same aspect ration of the image and contains all masked area, and then expand that area based + on `padding_mask_crop`. The image and mask_image will then be cropped based on the expanded area before + resizing to the original image size for inpainting. This is useful when the masked area is small while + the image is large and contain information irrelevant for inpainting, such as background. + strength (`float`, *optional*, defaults to 0.9999): + Conceptually, indicates how much to transform the masked portion of the reference `image`. Must be + between 0 and 1. `image` will be used as a starting point, adding more noise to it the larger the + `strength`. The number of denoising steps depends on the amount of noise initially added. When + `strength` is 1, added noise will be maximum and the denoising process will run for the full number of + iterations specified in `num_inference_steps`. A value of 1, therefore, essentially ignores the masked + portion of the reference `image`. Note that in the case of `denoising_start` being declared as an + integer, the value of `strength` will be ignored. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + denoising_start (`float`, *optional*): + When specified, indicates the fraction (between 0.0 and 1.0) of the total denoising process to be + bypassed before it is initiated. Consequently, the initial part of the denoising process is skipped and + it is assumed that the passed `image` is a partly denoised image. Note that when this is specified, + strength will be ignored. The `denoising_start` parameter is particularly beneficial when this pipeline + is integrated into a "Mixture of Denoisers" multi-pipeline setup, as detailed in [**Refining the Image + Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output). + denoising_end (`float`, *optional*): + When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be + completed before it is intentionally prematurely terminated. As a result, the returned sample will + still retain a substantial amount of noise (ca. final 20% of timesteps still needed) and should be + denoised by a successor pipeline that has `denoising_start` set to 0.8 so that it only denoises the + final 20% of the scheduler. The denoising_end parameter should ideally be utilized when this pipeline + forms a part of a "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image + Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output). + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(width, height)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(width, height)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + aesthetic_score (`float`, *optional*, defaults to 6.0): + Used to simulate an aesthetic score of the generated image by influencing the positive text condition. + Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_aesthetic_score (`float`, *optional*, defaults to 2.5): + Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). Can be used to + simulate an aesthetic score of the generated image by influencing the negative text condition. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a + `tuple. `tuple. When returning a tuple, the first element is a list with the generated images. + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`", + ) + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet + + if not isinstance(control_image_list, (ControlNetUnionInput, ControlNetUnionInputProMax)): + raise ValueError( + "Expected type of `control_image_list` to be one of `ControlNetUnionInput` or `ControlNetUnionInputProMax`" + ) + if len(control_image_list) != controlnet.config.num_control_type: + if isinstance(control_image_list, ControlNetUnionInput): + raise ValueError( + f"Expected num_control_type {controlnet.config.num_control_type}, got {len(control_image_list)}. Try `ControlNetUnionInputProMax`." + ) + elif isinstance(control_image_list, ControlNetUnionInputProMax): + raise ValueError( + f"Expected num_control_type {controlnet.config.num_control_type}, got {len(control_image_list)}. Try `ControlNetUnionInput`." + ) + + # align format for control guidance + if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list): + control_guidance_start = len(control_guidance_end) * [control_guidance_start] + elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list): + control_guidance_end = len(control_guidance_start) * [control_guidance_end] + + # # 0.0 Default height and width to unet + # height = height or self.unet.config.sample_size * self.vae_scale_factor + # width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 0.1 align format for control guidance + if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list): + control_guidance_start = len(control_guidance_end) * [control_guidance_start] + elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list): + control_guidance_end = len(control_guidance_start) * [control_guidance_end] + + # 1. Check inputs + control_type = [] + for image_type in control_image_list: + if control_image_list[image_type]: + self.check_inputs( + prompt, + prompt_2, + control_image_list[image_type], + mask_image, + strength, + num_inference_steps, + callback_steps, + output_type, + negative_prompt, + negative_prompt_2, + prompt_embeds, + negative_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + controlnet_conditioning_scale, + control_guidance_start, + control_guidance_end, + callback_on_step_end_tensor_inputs, + padding_mask_crop, + ) + control_type.append(1) + else: + control_type.append(0) + + control_type = torch.Tensor(control_type) + + self._guidance_scale = guidance_scale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + text_encoder_lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=self.clip_skip, + ) + + # 3.1 Encode ip_adapter_image + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 4. set timesteps + def denoising_value_valid(dnv): + return isinstance(dnv, float) and 0 < dnv < 1 + + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps, num_inference_steps = self.get_timesteps( + num_inference_steps, + strength, + device, + denoising_start=denoising_start if denoising_value_valid(denoising_start) else None, + ) + # check that number of inference steps is not < 1 - as this doesn't make sense + if num_inference_steps < 1: + raise ValueError( + f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline" + f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." + ) + # at which timestep to set the initial noise (n.b. 50% if strength is 0.5) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + # create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise + is_strength_max = strength == 1.0 + self._num_timesteps = len(timesteps) + + # 5. Preprocess mask and image - resizes image and mask w.r.t height and width + # 5.1 Prepare init image + if padding_mask_crop is not None: + height, width = self.image_processor.get_default_height_width(image, height, width) + crops_coords = self.mask_processor.get_crop_region(mask_image, width, height, pad=padding_mask_crop) + resize_mode = "fill" + else: + crops_coords = None + resize_mode = "default" + + original_image = image + init_image = self.image_processor.preprocess( + image, height=height, width=width, crops_coords=crops_coords, resize_mode=resize_mode + ) + init_image = init_image.to(dtype=torch.float32) + + # 5.2 Prepare control images + for image_type in control_image_list: + if control_image_list[image_type]: + control_image = self.prepare_control_image( + image=control_image_list[image_type], + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=controlnet.dtype, + crops_coords=crops_coords, + resize_mode=resize_mode, + do_classifier_free_guidance=self.do_classifier_free_guidance, + guess_mode=guess_mode, + ) + height, width = control_image.shape[-2:] + control_image_list[image_type] = control_image + + # 5.3 Prepare mask + mask = self.mask_processor.preprocess( + mask_image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords + ) + + masked_image = init_image * (mask < 0.5) + _, _, height, width = init_image.shape + + # 6. Prepare latent variables + num_channels_latents = self.vae.config.latent_channels + num_channels_unet = self.unet.config.in_channels + return_image_latents = num_channels_unet == 4 + + add_noise = True if denoising_start is None else False + latents_outputs = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + image=init_image, + timestep=latent_timestep, + is_strength_max=is_strength_max, + add_noise=add_noise, + return_noise=True, + return_image_latents=return_image_latents, + ) + + if return_image_latents: + latents, noise, image_latents = latents_outputs + else: + latents, noise = latents_outputs + + # 7. Prepare mask latent variables + mask, _ = self.prepare_mask_latents( + mask, + masked_image, + batch_size * num_images_per_prompt, + height, + width, + prompt_embeds.dtype, + device, + generator, + self.do_classifier_free_guidance, + ) + + # 8. Check that sizes of mask, masked image and latents match + if num_channels_unet != 4: + raise ValueError( + f"The unet {self.unet.__class__} should have either 4 or 9 input channels, not {self.unet.config.in_channels}." + ) + # 8.1 Prepare extra step kwargs. + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 8.2 Create tensor stating which controlnets to keep + controlnet_keep = [] + for i in range(len(timesteps)): + controlnet_keep.append( + 1.0 + - float(i / len(timesteps) < control_guidance_start or (i + 1) / len(timesteps) > control_guidance_end) + ) + + # 9. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + height, width = latents.shape[-2:] + height = height * self.vae_scale_factor + width = width * self.vae_scale_factor + + original_size = original_size or (height, width) + target_size = target_size or (height, width) + + # 10. Prepare added time ids & embeddings + add_text_embeds = pooled_prompt_embeds + if self.text_encoder_2 is None: + text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) + else: + text_encoder_projection_dim = self.text_encoder_2.config.projection_dim + + add_time_ids, add_neg_time_ids = self._get_add_time_ids( + original_size, + crops_coords_top_left, + target_size, + aesthetic_score, + negative_aesthetic_score, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + add_time_ids = add_time_ids.repeat(batch_size * num_images_per_prompt, 1) + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) + add_neg_time_ids = add_neg_time_ids.repeat(batch_size * num_images_per_prompt, 1) + add_time_ids = torch.cat([add_neg_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device) + + # 11. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + + if ( + denoising_end is not None + and denoising_start is not None + and denoising_value_valid(denoising_end) + and denoising_value_valid(denoising_start) + and denoising_start >= denoising_end + ): + raise ValueError( + f"`denoising_start`: {denoising_start} cannot be larger than or equal to `denoising_end`: " + + f" {denoising_end} when using type float." + ) + elif denoising_end is not None and denoising_value_valid(denoising_end): + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (denoising_end * self.scheduler.config.num_train_timesteps) + ) + ) + num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) + timesteps = timesteps[:num_inference_steps] + + control_type = ( + control_type.reshape(1, -1) + .to(device, dtype=prompt_embeds.dtype) + .repeat(batch_size * num_images_per_prompt * 2, 1) + ) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + added_cond_kwargs = { + "text_embeds": add_text_embeds, + "time_ids": add_time_ids, + } + + # controlnet(s) inference + if guess_mode and self.do_classifier_free_guidance: + # Infer ControlNet only for the conditional batch. + control_model_input = latents + control_model_input = self.scheduler.scale_model_input(control_model_input, t) + controlnet_prompt_embeds = prompt_embeds.chunk(2)[1] + controlnet_added_cond_kwargs = { + "text_embeds": add_text_embeds.chunk(2)[1], + "time_ids": add_time_ids.chunk(2)[1], + } + else: + control_model_input = latent_model_input + controlnet_prompt_embeds = prompt_embeds + controlnet_added_cond_kwargs = added_cond_kwargs + + if isinstance(controlnet_keep[i], list): + cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])] + else: + controlnet_cond_scale = controlnet_conditioning_scale + if isinstance(controlnet_cond_scale, list): + controlnet_cond_scale = controlnet_cond_scale[0] + cond_scale = controlnet_cond_scale * controlnet_keep[i] + + # # Resize control_image to match the size of the input to the controlnet + # if control_image.shape[-2:] != control_model_input.shape[-2:]: + # control_image = F.interpolate(control_image, size=control_model_input.shape[-2:], mode="bilinear", align_corners=False) + + down_block_res_samples, mid_block_res_sample = self.controlnet( + control_model_input, + t, + encoder_hidden_states=controlnet_prompt_embeds, + controlnet_cond=control_image_list, + control_type=control_type, + conditioning_scale=cond_scale, + guess_mode=guess_mode, + added_cond_kwargs=controlnet_added_cond_kwargs, + return_dict=False, + ) + + if guess_mode and self.do_classifier_free_guidance: + # Inferred ControlNet only for the conditional batch. + # To apply the output of ControlNet to both the unconditional and conditional batches, + # add 0 to the unconditional batch to keep it unchanged. + down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples] + mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample]) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + added_cond_kwargs["image_embeds"] = image_embeds + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=self.cross_attention_kwargs, + down_block_additional_residuals=down_block_res_samples, + mid_block_additional_residual=mid_block_res_sample, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and guidance_rescale > 0.0: + # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + init_latents_proper = image_latents + if self.do_classifier_free_guidance: + init_mask, _ = mask.chunk(2) + else: + init_mask = mask + + if i < len(timesteps) - 1: + noise_timestep = timesteps[i + 1] + init_latents_proper = self.scheduler.add_noise( + init_latents_proper, noise, torch.tensor([noise_timestep]) + ) + + latents = (1 - init_mask) * init_latents_proper + init_mask * latents + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + # make sure the VAE is in float32 mode, as it overflows in float16 + if self.vae.dtype == torch.float16 and self.vae.config.force_upcast: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + + # If we do sequential model offloading, let's offload unet and controlnet + # manually for max memory savings + if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: + self.unet.to("cpu") + self.controlnet.to("cpu") + torch.cuda.empty_cache() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + else: + return StableDiffusionXLPipelineOutput(images=latents) + + # apply watermark if available + if self.watermark is not None: + image = self.watermark.apply_watermark(image) + + image = self.image_processor.postprocess(image, output_type=output_type) + + if padding_mask_crop is not None: + image = [self.image_processor.apply_overlay(mask_image, original_image, i, crops_coords) for i in image] + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return StableDiffusionXLPipelineOutput(images=image) diff --git a/src/diffusers/pipelines/controlnet/pipeline_controlnet_union_sd_xl.py b/src/diffusers/pipelines/controlnet/pipeline_controlnet_union_sd_xl.py new file mode 100644 index 000000000000..58a8ba62e24e --- /dev/null +++ b/src/diffusers/pipelines/controlnet/pipeline_controlnet_union_sd_xl.py @@ -0,0 +1,1531 @@ +# Copyright 2024 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import inspect +from typing import Any, Callable, Dict, List, Optional, Tuple, Union + +import numpy as np +import PIL.Image +import torch +import torch.nn.functional as F +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionModelWithProjection, +) + +from diffusers.utils.import_utils import is_invisible_watermark_available + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import ( + FromSingleFileMixin, + IPAdapterMixin, + StableDiffusionXLLoraLoaderMixin, + TextualInversionLoaderMixin, +) +from ...models import AutoencoderKL, ControlNetModel, ControlNetUnionModel, ImageProjection, UNet2DConditionModel +from ...models.attention_processor import ( + AttnProcessor2_0, + XFormersAttnProcessor, +) +from ...models.controlnets import ControlNetUnionInput, ControlNetUnionInputProMax +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import is_compiled_module, is_torch_version, randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput + + +if is_invisible_watermark_available(): + from ..stable_diffusion_xl.watermark import StableDiffusionXLWatermarker + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> # !pip install controlnet_aux + >>> from controlnet_aux import LineartAnimeDetector + >>> from diffusers import StableDiffusionXLControlNetUnionPipeline, ControlNetUnionModel, AutoencoderKL + >>> from diffusers.models.controlnets import ControlNetUnionInput + >>> from diffusers.utils import load_image + >>> import torch + + >>> prompt = "A cat" + >>> # download an image + >>> image = load_image( + ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/kandinsky/cat.png" + ... ).resize((1024, 1024)) + >>> # initialize the models and pipeline + >>> controlnet = ControlNetUnionModel.from_pretrained( + ... "xinsir/controlnet-union-sdxl-1.0", torch_dtype=torch.float16 + ... ) + >>> vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16) + >>> pipe = StableDiffusionXLControlNetUnionPipeline.from_pretrained( + ... "stabilityai/stable-diffusion-xl-base-1.0", + ... controlnet=controlnet, + ... vae=vae, + ... torch_dtype=torch.float16, + ... ) + >>> pipe.enable_model_cpu_offload() + >>> # prepare image + >>> processor = LineartAnimeDetector.from_pretrained("lllyasviel/Annotators") + >>> controlnet_img = processor(image, output_type="pil") + >>> # set ControlNetUnion input + >>> union_input = ControlNetUnionInput( + ... canny=controlnet_img, + ... ) + >>> # generate image + >>> image = pipe(prompt, image=union_input).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + sigmas: Optional[List[float]] = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`List[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionXLControlNetUnionPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionXLLoraLoaderMixin, + IPAdapterMixin, + FromSingleFileMixin, +): + r""" + Pipeline for text-to-image generation using Stable Diffusion XL with ControlNet guidance. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + text_encoder_2 ([`~transformers.CLIPTextModelWithProjection`]): + Second frozen text-encoder + ([laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + tokenizer_2 ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + controlnet ([`ControlNetUnionModel`]`): + Provides additional conditioning to the `unet` during the denoising process. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`): + Whether the negative prompt embeddings should always be set to 0. Also see the config of + `stabilityai/stable-diffusion-xl-base-1-0`. + add_watermarker (`bool`, *optional*): + Whether to use the [invisible_watermark](https://github.com/ShieldMnt/invisible-watermark/) library to + watermark output images. If not defined, it defaults to `True` if the package is installed; otherwise no + watermarker is used. + """ + + # leave controlnet out on purpose because it iterates with unet + model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae" + _optional_components = [ + "tokenizer", + "tokenizer_2", + "text_encoder", + "text_encoder_2", + "feature_extractor", + "image_encoder", + ] + _callback_tensor_inputs = [ + "latents", + "prompt_embeds", + "negative_prompt_embeds", + "add_text_embeds", + "add_time_ids", + "negative_pooled_prompt_embeds", + "negative_add_time_ids", + "image", + ] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + tokenizer_2: CLIPTokenizer, + unet: UNet2DConditionModel, + controlnet: ControlNetUnionModel, + scheduler: KarrasDiffusionSchedulers, + force_zeros_for_empty_prompt: bool = True, + add_watermarker: Optional[bool] = None, + feature_extractor: CLIPImageProcessor = None, + image_encoder: CLIPVisionModelWithProjection = None, + ): + super().__init__() + + if not isinstance(controlnet, ControlNetUnionModel): + raise ValueError("Expected `controlnet` to be of type `ControlNetUnionModel`.") + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + unet=unet, + controlnet=controlnet, + scheduler=scheduler, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True) + self.control_image_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True, do_normalize=False + ) + add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available() + + if add_watermarker: + self.watermark = StableDiffusionXLWatermarker() + else: + self.watermark = None + + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt + def encode_prompt( + self, + prompt: str, + prompt_2: Optional[str] = None, + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: Optional[str] = None, + negative_prompt_2: Optional[str] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + pooled_prompt_embeds: Optional[torch.Tensor] = None, + negative_pooled_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale) + else: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] + text_encoders = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # textual inversion: process multi-vector tokens if necessary + prompt_embeds_list = [] + prompts = [prompt, prompt_2] + for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, tokenizer) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + + # We are only ALWAYS interested in the pooled output of the final text encoder + pooled_prompt_embeds = prompt_embeds[0] + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + # "2" because SDXL always indexes from the penultimate layer. + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + + uncond_tokens: List[str] + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = [negative_prompt, negative_prompt_2] + + negative_prompt_embeds_list = [] + for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = tokenizer( + negative_prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + # We are only ALWAYS interested in the pooled output of the final text encoder + negative_pooled_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) + + if self.text_encoder_2 is not None: + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + if self.text_encoder_2 is not None: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.controlnet.pipeline_controlnet_sd_xl.StableDiffusionXLControlNetPipeline.check_image + def check_image(self, image, prompt, prompt_embeds): + image_is_pil = isinstance(image, PIL.Image.Image) + image_is_tensor = isinstance(image, torch.Tensor) + image_is_np = isinstance(image, np.ndarray) + image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image) + image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor) + image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray) + + if ( + not image_is_pil + and not image_is_tensor + and not image_is_np + and not image_is_pil_list + and not image_is_tensor_list + and not image_is_np_list + ): + raise TypeError( + f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}" + ) + + if image_is_pil: + image_batch_size = 1 + else: + image_batch_size = len(image) + + if prompt is not None and isinstance(prompt, str): + prompt_batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + prompt_batch_size = len(prompt) + elif prompt_embeds is not None: + prompt_batch_size = prompt_embeds.shape[0] + + if image_batch_size != 1 and image_batch_size != prompt_batch_size: + raise ValueError( + f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}" + ) + + def check_inputs( + self, + prompt, + prompt_2, + image: PipelineImageInput, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + negative_pooled_prompt_embeds=None, + controlnet_conditioning_scale=1.0, + control_guidance_start=0.0, + control_guidance_end=1.0, + callback_on_step_end_tensor_inputs=None, + ): + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + # Check `image` + is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance( + self.controlnet, torch._dynamo.eval_frame.OptimizedModule + ) + if ( + isinstance(self.controlnet, ControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, ControlNetModel) + ): + self.check_image(image, prompt, prompt_embeds) + elif ( + isinstance(self.controlnet, ControlNetUnionModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, ControlNetUnionModel) + ): + self.check_image(image, prompt, prompt_embeds) + + else: + assert False + + # Check `controlnet_conditioning_scale` + if ( + isinstance(self.controlnet, ControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, ControlNetModel) + ): + if not isinstance(controlnet_conditioning_scale, float): + raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.") + + elif ( + isinstance(self.controlnet, ControlNetUnionModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, ControlNetUnionModel) + ): + if not isinstance(controlnet_conditioning_scale, float): + raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.") + + else: + assert False + + if not isinstance(control_guidance_start, (tuple, list)): + control_guidance_start = [control_guidance_start] + + if not isinstance(control_guidance_end, (tuple, list)): + control_guidance_end = [control_guidance_end] + + if len(control_guidance_start) != len(control_guidance_end): + raise ValueError( + f"`control_guidance_start` has {len(control_guidance_start)} elements, but `control_guidance_end` has {len(control_guidance_end)} elements. Make sure to provide the same number of elements to each list." + ) + + for start, end in zip(control_guidance_start, control_guidance_end): + if start >= end: + raise ValueError( + f"control guidance start: {start} cannot be larger or equal to control guidance end: {end}." + ) + if start < 0.0: + raise ValueError(f"control guidance start: {start} can't be smaller than 0.") + if end > 1.0: + raise ValueError(f"control guidance end: {end} can't be larger than 1.0.") + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + def check_input( + self, + image: Union[ControlNetUnionInput, ControlNetUnionInputProMax], + ): + controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet + + if not isinstance(image, (ControlNetUnionInput, ControlNetUnionInputProMax)): + raise ValueError( + "Expected type of `image` to be one of `ControlNetUnionInput` or `ControlNetUnionInputProMax`" + ) + if len(image) != controlnet.config.num_control_type: + if isinstance(image, ControlNetUnionInput): + raise ValueError( + f"Expected num_control_type {controlnet.config.num_control_type}, got {len(image)}. Try `ControlNetUnionInputProMax`." + ) + elif isinstance(image, ControlNetUnionInputProMax): + raise ValueError( + f"Expected num_control_type {controlnet.config.num_control_type}, got {len(image)}. Try `ControlNetUnionInput`." + ) + + # Copied from diffusers.pipelines.controlnet.pipeline_controlnet.StableDiffusionControlNetPipeline.prepare_image + def prepare_image( + self, + image, + width, + height, + batch_size, + num_images_per_prompt, + device, + dtype, + do_classifier_free_guidance=False, + guess_mode=False, + ): + image = self.control_image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32) + image_batch_size = image.shape[0] + + if image_batch_size == 1: + repeat_by = batch_size + else: + # image batch size is the same as prompt batch size + repeat_by = num_images_per_prompt + + image = image.repeat_interleave(repeat_by, dim=0) + + image = image.to(device=device, dtype=dtype) + + if do_classifier_free_guidance and not guess_mode: + image = torch.cat([image] * 2) + + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline._get_add_time_ids + def _get_add_time_ids( + self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None + ): + add_time_ids = list(original_size + crops_coords_top_left + target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + return add_time_ids + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae + def upcast_vae(self): + dtype = self.vae.dtype + self.vae.to(dtype=torch.float32) + use_torch_2_0_or_xformers = isinstance( + self.vae.decoder.mid_block.attentions[0].processor, + ( + AttnProcessor2_0, + XFormersAttnProcessor, + ), + ) + # if xformers or torch_2_0 is used attention block does not need + # to be in float32 which can save lots of memory + if use_torch_2_0_or_xformers: + self.vae.post_quant_conv.to(dtype) + self.vae.decoder.conv_in.to(dtype) + self.vae.decoder.mid_block.to(dtype) + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def denoising_end(self): + return self._denoising_end + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + prompt_2: Optional[Union[str, List[str]]] = None, + image: Union[ControlNetUnionInput, ControlNetUnionInputProMax] = None, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 50, + timesteps: List[int] = None, + sigmas: List[float] = None, + denoising_end: Optional[float] = None, + guidance_scale: float = 5.0, + negative_prompt: Optional[Union[str, List[str]]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + pooled_prompt_embeds: Optional[torch.Tensor] = None, + negative_pooled_prompt_embeds: Optional[torch.Tensor] = None, + ip_adapter_image: Optional[PipelineImageInput] = None, + ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + controlnet_conditioning_scale: Union[float, List[float]] = 1.0, + guess_mode: bool = False, + control_guidance_start: Union[float, List[float]] = 0.0, + control_guidance_end: Union[float, List[float]] = 1.0, + original_size: Tuple[int, int] = None, + crops_coords_top_left: Tuple[int, int] = (0, 0), + target_size: Tuple[int, int] = None, + negative_original_size: Optional[Tuple[int, int]] = None, + negative_crops_coords_top_left: Tuple[int, int] = (0, 0), + negative_target_size: Optional[Tuple[int, int]] = None, + clip_skip: Optional[int] = None, + callback_on_step_end: Optional[ + Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks] + ] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders. + image (`Union[ControlNetUnionInput, ControlNetUnionInputProMax]`): + In turn this supports (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, + `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[List[torch.FloatTensor]]`, + `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`): + The ControlNet input condition to provide guidance to the `unet` for generation. If the type is + specified as `torch.Tensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be accepted + as an image. The dimensions of the output image defaults to `image`'s dimensions. If height and/or + width are passed, `image` is resized accordingly. If multiple ControlNets are specified in `init`, + images must be passed as a list such that each element of the list can be correctly batched for input + to a single ControlNet. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`List[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + denoising_end (`float`, *optional*): + When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be + completed before it is intentionally prematurely terminated. As a result, the returned sample will + still retain a substantial amount of noise as determined by the discrete timesteps selected by the + scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a + "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image + Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output) + guidance_scale (`float`, *optional*, defaults to 5.0): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. This is sent to `tokenizer_2` + and `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies + to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, pooled text embeddings are generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs (prompt + weighting). If not provided, pooled `negative_prompt_embeds` are generated from `negative_prompt` input + argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0): + The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added + to the residual in the original `unet`. If multiple ControlNets are specified in `init`, you can set + the corresponding scale as a list. + guess_mode (`bool`, *optional*, defaults to `False`): + The ControlNet encoder tries to recognize the content of the input image even if you remove all + prompts. A `guidance_scale` value between 3.0 and 5.0 is recommended. + control_guidance_start (`float` or `List[float]`, *optional*, defaults to 0.0): + The percentage of total steps at which the ControlNet starts applying. + control_guidance_end (`float` or `List[float]`, *optional*, defaults to 1.0): + The percentage of total steps at which the ControlNet stops applying. + original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a specific image resolution. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)): + To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a target image resolution. It should be as same + as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned containing the output images. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet + + self.check_input(image) + + # align format for control guidance + if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list): + control_guidance_start = len(control_guidance_end) * [control_guidance_start] + elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list): + control_guidance_end = len(control_guidance_start) * [control_guidance_end] + + # 1. Check inputs. Raise error if not correct + control_type = [] + for image_type in image: + if image[image_type]: + self.check_inputs( + prompt, + prompt_2, + image[image_type], + negative_prompt, + negative_prompt_2, + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + negative_pooled_prompt_embeds, + controlnet_conditioning_scale, + control_guidance_start, + control_guidance_end, + callback_on_step_end_tensor_inputs, + ) + control_type.append(1) + else: + control_type.append(0) + + control_type = torch.Tensor(control_type) + + self._guidance_scale = guidance_scale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._denoising_end = denoising_end + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + global_pool_conditions = controlnet.config.global_pool_conditions + guess_mode = guess_mode or global_pool_conditions + + # 3.1 Encode input prompt + text_encoder_lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt, + prompt_2, + device, + num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=self.clip_skip, + ) + + # 3.2 Encode ip_adapter_image + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 4. Prepare image + for image_type in image: + if image[image_type]: + image[image_type] = self.prepare_image( + image=image[image_type], + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=controlnet.dtype, + do_classifier_free_guidance=self.do_classifier_free_guidance, + guess_mode=guess_mode, + ) + height, width = image[image_type].shape[-2:] + + # 5. Prepare timesteps + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, device, timesteps, sigmas + ) + self._num_timesteps = len(timesteps) + + # 6. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6.5 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7.1 Create tensor stating which controlnets to keep + controlnet_keep = [] + for i in range(len(timesteps)): + controlnet_keep.append( + 1.0 + - float(i / len(timesteps) < control_guidance_start or (i + 1) / len(timesteps) > control_guidance_end) + ) + + # 7.2 Prepare added time ids & embeddings + for image_type in image: + if isinstance(image[image_type], torch.Tensor): + original_size = original_size or image[image_type].shape[-2:] + + target_size = target_size or (height, width) + add_text_embeds = pooled_prompt_embeds + if self.text_encoder_2 is None: + text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) + else: + text_encoder_projection_dim = self.text_encoder_2.config.projection_dim + + add_time_ids = self._get_add_time_ids( + original_size, + crops_coords_top_left, + target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + + if negative_original_size is not None and negative_target_size is not None: + negative_add_time_ids = self._get_add_time_ids( + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + else: + negative_add_time_ids = add_time_ids + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) + add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1) + + # 8. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + + # 8.1 Apply denoising_end + if ( + self.denoising_end is not None + and isinstance(self.denoising_end, float) + and self.denoising_end > 0 + and self.denoising_end < 1 + ): + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (self.denoising_end * self.scheduler.config.num_train_timesteps) + ) + ) + num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) + timesteps = timesteps[:num_inference_steps] + + is_unet_compiled = is_compiled_module(self.unet) + is_controlnet_compiled = is_compiled_module(self.controlnet) + is_torch_higher_equal_2_1 = is_torch_version(">=", "2.1") + + control_type = ( + control_type.reshape(1, -1) + .to(device, dtype=prompt_embeds.dtype) + .repeat(batch_size * num_images_per_prompt * 2, 1) + ) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # Relevant thread: + # https://dev-discuss.pytorch.org/t/cudagraphs-in-pytorch-2-0/1428 + if (is_unet_compiled and is_controlnet_compiled) and is_torch_higher_equal_2_1: + torch._inductor.cudagraph_mark_step_begin() + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + added_cond_kwargs = { + "text_embeds": add_text_embeds, + "time_ids": add_time_ids, + } + + # controlnet(s) inference + if guess_mode and self.do_classifier_free_guidance: + # Infer ControlNet only for the conditional batch. + control_model_input = latents + control_model_input = self.scheduler.scale_model_input(control_model_input, t) + controlnet_prompt_embeds = prompt_embeds.chunk(2)[1] + controlnet_added_cond_kwargs = { + "text_embeds": add_text_embeds.chunk(2)[1], + "time_ids": add_time_ids.chunk(2)[1], + } + else: + control_model_input = latent_model_input + controlnet_prompt_embeds = prompt_embeds + controlnet_added_cond_kwargs = added_cond_kwargs + + if isinstance(controlnet_keep[i], list): + cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])] + else: + controlnet_cond_scale = controlnet_conditioning_scale + if isinstance(controlnet_cond_scale, list): + controlnet_cond_scale = controlnet_cond_scale[0] + cond_scale = controlnet_cond_scale * controlnet_keep[i] + + down_block_res_samples, mid_block_res_sample = self.controlnet( + control_model_input, + t, + encoder_hidden_states=controlnet_prompt_embeds, + controlnet_cond=image, + control_type=control_type, + conditioning_scale=cond_scale, + guess_mode=guess_mode, + added_cond_kwargs=controlnet_added_cond_kwargs, + return_dict=False, + ) + + if guess_mode and self.do_classifier_free_guidance: + # Inferred ControlNet only for the conditional batch. + # To apply the output of ControlNet to both the unconditional and conditional batches, + # add 0 to the unconditional batch to keep it unchanged. + down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples] + mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample]) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + added_cond_kwargs["image_embeds"] = image_embeds + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + down_block_additional_residuals=down_block_res_samples, + mid_block_additional_residual=mid_block_res_sample, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds) + negative_pooled_prompt_embeds = callback_outputs.pop( + "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds + ) + add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids) + negative_add_time_ids = callback_outputs.pop("negative_add_time_ids", negative_add_time_ids) + image = callback_outputs.pop("image", image) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if not output_type == "latent": + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + + # unscale/denormalize the latents + # denormalize with the mean and std if available and not None + has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None + has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None + if has_latents_mean and has_latents_std: + latents_mean = ( + torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents_std = ( + torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean + else: + latents = latents / self.vae.config.scaling_factor + + image = self.vae.decode(latents, return_dict=False)[0] + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + else: + image = latents + + if not output_type == "latent": + # apply watermark if available + if self.watermark is not None: + image = self.watermark.apply_watermark(image) + + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return StableDiffusionXLPipelineOutput(images=image) diff --git a/src/diffusers/pipelines/controlnet/pipeline_controlnet_union_sd_xl_img2img.py b/src/diffusers/pipelines/controlnet/pipeline_controlnet_union_sd_xl_img2img.py new file mode 100644 index 000000000000..a3002eb565ff --- /dev/null +++ b/src/diffusers/pipelines/controlnet/pipeline_controlnet_union_sd_xl_img2img.py @@ -0,0 +1,1646 @@ +# Copyright 2024 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import inspect +from typing import Any, Callable, Dict, List, Optional, Tuple, Union + +import numpy as np +import PIL.Image +import torch +import torch.nn.functional as F +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionModelWithProjection, +) + +from diffusers.utils.import_utils import is_invisible_watermark_available + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import ( + FromSingleFileMixin, + IPAdapterMixin, + StableDiffusionXLLoraLoaderMixin, + TextualInversionLoaderMixin, +) +from ...models import AutoencoderKL, ControlNetModel, ControlNetUnionModel, ImageProjection, UNet2DConditionModel +from ...models.attention_processor import ( + AttnProcessor2_0, + XFormersAttnProcessor, +) +from ...models.controlnets import ControlNetUnionInput, ControlNetUnionInputProMax +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import is_compiled_module, randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput + + +if is_invisible_watermark_available(): + from ..stable_diffusion_xl.watermark import StableDiffusionXLWatermarker + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + # !pip install controlnet_aux + from diffusers import ( + StableDiffusionXLControlNetUnionImg2ImgPipeline, + ControlNetUnionModel, + AutoencoderKL, + ) + from diffusers.models.controlnets import ControlNetUnionInputProMax + from diffusers.utils import load_image + import torch + from PIL import Image + import numpy as np + + prompt = "A cat" + # download an image + image = load_image( + "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/kandinsky/cat.png" + ) + # initialize the models and pipeline + controlnet = ControlNetUnionModel.from_pretrained( + "brad-twinkl/controlnet-union-sdxl-1.0-promax", torch_dtype=torch.float16 + ) + vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16) + pipe = StableDiffusionXLControlNetUnionImg2ImgPipeline.from_pretrained( + "stabilityai/stable-diffusion-xl-base-1.0", + controlnet=controlnet, + vae=vae, + torch_dtype=torch.float16, + ).to("cuda") + # `enable_model_cpu_offload` is not recommended due to multiple generations + height = image.height + width = image.width + ratio = np.sqrt(1024.0 * 1024.0 / (width * height)) + # 3 * 3 upscale correspond to 16 * 3 multiply, 2 * 2 correspond to 16 * 2 multiply and so on. + scale_image_factor = 3 + base_factor = 16 + factor = scale_image_factor * base_factor + W, H = int(width * ratio) // factor * factor, int(height * ratio) // factor * factor + image = image.resize((W, H)) + target_width = W // scale_image_factor + target_height = H // scale_image_factor + images = [] + crops_coords_list = [ + (0, 0), + (0, width // 2), + (height // 2, 0), + (width // 2, height // 2), + 0, + 0, + 0, + 0, + 0, + ] + for i in range(scale_image_factor): + for j in range(scale_image_factor): + left = j * target_width + top = i * target_height + right = left + target_width + bottom = top + target_height + cropped_image = image.crop((left, top, right, bottom)) + cropped_image = cropped_image.resize((W, H)) + images.append(cropped_image) + # set ControlNetUnion input + result_images = [] + for sub_img, crops_coords in zip(images, crops_coords_list): + union_input = ControlNetUnionInputProMax( + tile=sub_img, + ) + new_width, new_height = W, H + out = pipe( + prompt=[prompt] * 1, + image=sub_img, + control_image_list=union_input, + width=new_width, + height=new_height, + num_inference_steps=30, + crops_coords_top_left=(W, H), + target_size=(W, H), + original_size=(W * 2, H * 2), + ) + result_images.append(out.images[0]) + new_im = Image.new("RGB", (new_width * scale_image_factor, new_height * scale_image_factor)) + new_im.paste(result_images[0], (0, 0)) + new_im.paste(result_images[1], (new_width, 0)) + new_im.paste(result_images[2], (new_width * 2, 0)) + new_im.paste(result_images[3], (0, new_height)) + new_im.paste(result_images[4], (new_width, new_height)) + new_im.paste(result_images[5], (new_width * 2, new_height)) + new_im.paste(result_images[6], (0, new_height * 2)) + new_im.paste(result_images[7], (new_width, new_height * 2)) + new_im.paste(result_images[8], (new_width * 2, new_height * 2)) + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class StableDiffusionXLControlNetUnionImg2ImgPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionXLLoraLoaderMixin, + FromSingleFileMixin, + IPAdapterMixin, +): + r""" + Pipeline for image-to-image generation using Stable Diffusion XL with ControlNet guidance. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([` CLIPTextModelWithProjection`]): + Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + controlnet ([`ControlNetUnionModel`]): + Provides additional conditioning to the unet during the denoising process. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + requires_aesthetics_score (`bool`, *optional*, defaults to `"False"`): + Whether the `unet` requires an `aesthetic_score` condition to be passed during inference. Also see the + config of `stabilityai/stable-diffusion-xl-refiner-1-0`. + force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`): + Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of + `stabilityai/stable-diffusion-xl-base-1-0`. + add_watermarker (`bool`, *optional*): + Whether to use the [invisible_watermark library](https://github.com/ShieldMnt/invisible-watermark/) to + watermark output images. If not defined, it will default to True if the package is installed, otherwise no + watermarker will be used. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae" + _optional_components = [ + "tokenizer", + "tokenizer_2", + "text_encoder", + "text_encoder_2", + "feature_extractor", + "image_encoder", + ] + _callback_tensor_inputs = [ + "latents", + "prompt_embeds", + "negative_prompt_embeds", + "add_text_embeds", + "add_time_ids", + "negative_pooled_prompt_embeds", + "add_neg_time_ids", + ] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + tokenizer_2: CLIPTokenizer, + unet: UNet2DConditionModel, + controlnet: ControlNetUnionModel, + scheduler: KarrasDiffusionSchedulers, + requires_aesthetics_score: bool = False, + force_zeros_for_empty_prompt: bool = True, + add_watermarker: Optional[bool] = None, + feature_extractor: CLIPImageProcessor = None, + image_encoder: CLIPVisionModelWithProjection = None, + ): + super().__init__() + + if not isinstance(controlnet, ControlNetUnionModel): + raise ValueError("Expected `controlnet` to be of type `ControlNetUnionModel`.") + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + unet=unet, + controlnet=controlnet, + scheduler=scheduler, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True) + self.control_image_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True, do_normalize=False + ) + add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available() + + if add_watermarker: + self.watermark = StableDiffusionXLWatermarker() + else: + self.watermark = None + + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.register_to_config(requires_aesthetics_score=requires_aesthetics_score) + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt + def encode_prompt( + self, + prompt: str, + prompt_2: Optional[str] = None, + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: Optional[str] = None, + negative_prompt_2: Optional[str] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + pooled_prompt_embeds: Optional[torch.Tensor] = None, + negative_pooled_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale) + else: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] + text_encoders = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # textual inversion: process multi-vector tokens if necessary + prompt_embeds_list = [] + prompts = [prompt, prompt_2] + for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, tokenizer) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + + # We are only ALWAYS interested in the pooled output of the final text encoder + pooled_prompt_embeds = prompt_embeds[0] + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + # "2" because SDXL always indexes from the penultimate layer. + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + + uncond_tokens: List[str] + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = [negative_prompt, negative_prompt_2] + + negative_prompt_embeds_list = [] + for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = tokenizer( + negative_prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + # We are only ALWAYS interested in the pooled output of the final text encoder + negative_pooled_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) + + if self.text_encoder_2 is not None: + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + if self.text_encoder_2 is not None: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + prompt_2, + image, + strength, + num_inference_steps, + callback_steps, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + controlnet_conditioning_scale=1.0, + control_guidance_start=0.0, + control_guidance_end=1.0, + callback_on_step_end_tensor_inputs=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + if num_inference_steps is None: + raise ValueError("`num_inference_steps` cannot be None.") + elif not isinstance(num_inference_steps, int) or num_inference_steps <= 0: + raise ValueError( + f"`num_inference_steps` has to be a positive integer but is {num_inference_steps} of type" + f" {type(num_inference_steps)}." + ) + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + # Check `image` + is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance( + self.controlnet, torch._dynamo.eval_frame.OptimizedModule + ) + if ( + isinstance(self.controlnet, ControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, ControlNetModel) + ): + self.check_image(image, prompt, prompt_embeds) + elif ( + isinstance(self.controlnet, ControlNetUnionModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, ControlNetUnionModel) + ): + self.check_image(image, prompt, prompt_embeds) + else: + assert False + + # Check `controlnet_conditioning_scale` + if ( + isinstance(self.controlnet, ControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, ControlNetModel) + ): + if not isinstance(controlnet_conditioning_scale, float): + raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.") + + elif ( + isinstance(self.controlnet, ControlNetUnionModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, ControlNetUnionModel) + ): + if not isinstance(controlnet_conditioning_scale, float): + raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.") + + else: + assert False + + if not isinstance(control_guidance_start, (tuple, list)): + control_guidance_start = [control_guidance_start] + + if not isinstance(control_guidance_end, (tuple, list)): + control_guidance_end = [control_guidance_end] + + if len(control_guidance_start) != len(control_guidance_end): + raise ValueError( + f"`control_guidance_start` has {len(control_guidance_start)} elements, but `control_guidance_end` has {len(control_guidance_end)} elements. Make sure to provide the same number of elements to each list." + ) + + for start, end in zip(control_guidance_start, control_guidance_end): + if start >= end: + raise ValueError( + f"control guidance start: {start} cannot be larger or equal to control guidance end: {end}." + ) + if start < 0.0: + raise ValueError(f"control guidance start: {start} can't be smaller than 0.") + if end > 1.0: + raise ValueError(f"control guidance end: {end} can't be larger than 1.0.") + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + # Copied from diffusers.pipelines.controlnet.pipeline_controlnet_sd_xl.StableDiffusionXLControlNetPipeline.check_image + def check_image(self, image, prompt, prompt_embeds): + image_is_pil = isinstance(image, PIL.Image.Image) + image_is_tensor = isinstance(image, torch.Tensor) + image_is_np = isinstance(image, np.ndarray) + image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image) + image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor) + image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray) + + if ( + not image_is_pil + and not image_is_tensor + and not image_is_np + and not image_is_pil_list + and not image_is_tensor_list + and not image_is_np_list + ): + raise TypeError( + f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}" + ) + + if image_is_pil: + image_batch_size = 1 + else: + image_batch_size = len(image) + + if prompt is not None and isinstance(prompt, str): + prompt_batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + prompt_batch_size = len(prompt) + elif prompt_embeds is not None: + prompt_batch_size = prompt_embeds.shape[0] + + if image_batch_size != 1 and image_batch_size != prompt_batch_size: + raise ValueError( + f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}" + ) + + # Copied from diffusers.pipelines.controlnet.pipeline_controlnet_sd_xl.StableDiffusionXLControlNetPipeline.prepare_image + def prepare_control_image( + self, + image, + width, + height, + batch_size, + num_images_per_prompt, + device, + dtype, + do_classifier_free_guidance=False, + guess_mode=False, + ): + image = self.control_image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32) + image_batch_size = image.shape[0] + + if image_batch_size == 1: + repeat_by = batch_size + else: + # image batch size is the same as prompt batch size + repeat_by = num_images_per_prompt + + image = image.repeat_interleave(repeat_by, dim=0) + + image = image.to(device=device, dtype=dtype) + + if do_classifier_free_guidance and not guess_mode: + image = torch.cat([image] * 2) + + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline.prepare_latents + def prepare_latents( + self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None, add_noise=True + ): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + latents_mean = latents_std = None + if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None: + latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1) + if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None: + latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1) + + # Offload text encoder if `enable_model_cpu_offload` was enabled + if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: + self.text_encoder_2.to("cpu") + torch.cuda.empty_cache() + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + if image.shape[1] == 4: + init_latents = image + + else: + # make sure the VAE is in float32 mode, as it overflows in float16 + if self.vae.config.force_upcast: + image = image.float() + self.vae.to(dtype=torch.float32) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + elif isinstance(generator, list): + if image.shape[0] < batch_size and batch_size % image.shape[0] == 0: + image = torch.cat([image] * (batch_size // image.shape[0]), dim=0) + elif image.shape[0] < batch_size and batch_size % image.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image.shape[0]} to effective batch_size {batch_size} " + ) + + init_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(batch_size) + ] + init_latents = torch.cat(init_latents, dim=0) + else: + init_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + if self.vae.config.force_upcast: + self.vae.to(dtype) + + init_latents = init_latents.to(dtype) + if latents_mean is not None and latents_std is not None: + latents_mean = latents_mean.to(device=device, dtype=dtype) + latents_std = latents_std.to(device=device, dtype=dtype) + init_latents = (init_latents - latents_mean) * self.vae.config.scaling_factor / latents_std + else: + init_latents = self.vae.config.scaling_factor * init_latents + + if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // init_latents.shape[0] + init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0) + elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts." + ) + else: + init_latents = torch.cat([init_latents], dim=0) + + if add_noise: + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # get latents + init_latents = self.scheduler.add_noise(init_latents, noise, timestep) + + latents = init_latents + + return latents + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline._get_add_time_ids + def _get_add_time_ids( + self, + original_size, + crops_coords_top_left, + target_size, + aesthetic_score, + negative_aesthetic_score, + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype, + text_encoder_projection_dim=None, + ): + if self.config.requires_aesthetics_score: + add_time_ids = list(original_size + crops_coords_top_left + (aesthetic_score,)) + add_neg_time_ids = list( + negative_original_size + negative_crops_coords_top_left + (negative_aesthetic_score,) + ) + else: + add_time_ids = list(original_size + crops_coords_top_left + target_size) + add_neg_time_ids = list(negative_original_size + crops_coords_top_left + negative_target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if ( + expected_add_embed_dim > passed_add_embed_dim + and (expected_add_embed_dim - passed_add_embed_dim) == self.unet.config.addition_time_embed_dim + ): + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to enable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=True)` to make sure `aesthetic_score` {aesthetic_score} and `negative_aesthetic_score` {negative_aesthetic_score} is correctly used by the model." + ) + elif ( + expected_add_embed_dim < passed_add_embed_dim + and (passed_add_embed_dim - expected_add_embed_dim) == self.unet.config.addition_time_embed_dim + ): + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to disable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=False)` to make sure `target_size` {target_size} is correctly used by the model." + ) + elif expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + add_neg_time_ids = torch.tensor([add_neg_time_ids], dtype=dtype) + + return add_time_ids, add_neg_time_ids + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae + def upcast_vae(self): + dtype = self.vae.dtype + self.vae.to(dtype=torch.float32) + use_torch_2_0_or_xformers = isinstance( + self.vae.decoder.mid_block.attentions[0].processor, + ( + AttnProcessor2_0, + XFormersAttnProcessor, + ), + ) + # if xformers or torch_2_0 is used attention block does not need + # to be in float32 which can save lots of memory + if use_torch_2_0_or_xformers: + self.vae.post_quant_conv.to(dtype) + self.vae.decoder.conv_in.to(dtype) + self.vae.decoder.mid_block.to(dtype) + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + prompt_2: Optional[Union[str, List[str]]] = None, + image: PipelineImageInput = None, + control_image_list: Union[ControlNetUnionInput, ControlNetUnionInputProMax] = None, + height: Optional[int] = None, + width: Optional[int] = None, + strength: float = 0.8, + num_inference_steps: int = 50, + guidance_scale: float = 5.0, + negative_prompt: Optional[Union[str, List[str]]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + pooled_prompt_embeds: Optional[torch.Tensor] = None, + negative_pooled_prompt_embeds: Optional[torch.Tensor] = None, + ip_adapter_image: Optional[PipelineImageInput] = None, + ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + controlnet_conditioning_scale: Union[float, List[float]] = 0.8, + guess_mode: bool = False, + control_guidance_start: Union[float, List[float]] = 0.0, + control_guidance_end: Union[float, List[float]] = 1.0, + original_size: Tuple[int, int] = None, + crops_coords_top_left: Tuple[int, int] = (0, 0), + target_size: Tuple[int, int] = None, + negative_original_size: Optional[Tuple[int, int]] = None, + negative_crops_coords_top_left: Tuple[int, int] = (0, 0), + negative_target_size: Optional[Tuple[int, int]] = None, + aesthetic_score: float = 6.0, + negative_aesthetic_score: float = 2.5, + clip_skip: Optional[int] = None, + callback_on_step_end: Optional[ + Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks] + ] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + **kwargs, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,: + `List[List[torch.Tensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`): + The initial image will be used as the starting point for the image generation process. Can also accept + image latents as `image`, if passing latents directly, it will not be encoded again. + control_image_list (`Union[ControlNetUnionInput, ControlNetUnionInputProMax]`): + In turn this supports (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, + `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[List[torch.FloatTensor]]`, + `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`):: + The ControlNet input condition. ControlNet uses this input condition to generate guidance to Unet. If + the type is specified as `torch.Tensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also + be accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If height + and/or width are passed, `image` is resized according to them. If multiple ControlNets are specified in + init, images must be passed as a list such that each element of the list can be correctly batched for + input to a single controlnet. + height (`int`, *optional*, defaults to the size of control_image): + The height in pixels of the generated image. Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + width (`int`, *optional*, defaults to the size of control_image): + The width in pixels of the generated image. Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + strength (`float`, *optional*, defaults to 0.8): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0): + The outputs of the controlnet are multiplied by `controlnet_conditioning_scale` before they are added + to the residual in the original unet. If multiple ControlNets are specified in init, you can set the + corresponding scale as a list. + guess_mode (`bool`, *optional*, defaults to `False`): + In this mode, the ControlNet encoder will try best to recognize the content of the input image even if + you remove all prompts. The `guidance_scale` between 3.0 and 5.0 is recommended. + control_guidance_start (`float` or `List[float]`, *optional*, defaults to 0.0): + The percentage of total steps at which the controlnet starts applying. + control_guidance_end (`float` or `List[float]`, *optional*, defaults to 1.0): + The percentage of total steps at which the controlnet stops applying. + original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a specific image resolution. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)): + To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a target image resolution. It should be as same + as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + aesthetic_score (`float`, *optional*, defaults to 6.0): + Used to simulate an aesthetic score of the generated image by influencing the positive text condition. + Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_aesthetic_score (`float`, *optional*, defaults to 2.5): + Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). Can be used to + simulate an aesthetic score of the generated image by influencing the negative text condition. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple` + containing the output images. + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`", + ) + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet + + if not isinstance(control_image_list, (ControlNetUnionInput, ControlNetUnionInputProMax)): + raise ValueError( + "Expected type of `control_image_list` to be one of `ControlNetUnionInput` or `ControlNetUnionInputProMax`" + ) + if len(control_image_list) != controlnet.config.num_control_type: + if isinstance(control_image_list, ControlNetUnionInput): + raise ValueError( + f"Expected num_control_type {controlnet.config.num_control_type}, got {len(control_image_list)}. Try `ControlNetUnionInputProMax`." + ) + elif isinstance(control_image_list, ControlNetUnionInputProMax): + raise ValueError( + f"Expected num_control_type {controlnet.config.num_control_type}, got {len(control_image_list)}. Try `ControlNetUnionInput`." + ) + + # align format for control guidance + if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list): + control_guidance_start = len(control_guidance_end) * [control_guidance_start] + elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list): + control_guidance_end = len(control_guidance_start) * [control_guidance_end] + + # 1. Check inputs. Raise error if not correct + control_type = [] + for image_type in control_image_list: + if control_image_list[image_type]: + self.check_inputs( + prompt, + prompt_2, + control_image_list[image_type], + strength, + num_inference_steps, + callback_steps, + negative_prompt, + negative_prompt_2, + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + controlnet_conditioning_scale, + control_guidance_start, + control_guidance_end, + callback_on_step_end_tensor_inputs, + ) + control_type.append(1) + else: + control_type.append(0) + + control_type = torch.Tensor(control_type) + + self._guidance_scale = guidance_scale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + global_pool_conditions = controlnet.config.global_pool_conditions + guess_mode = guess_mode or global_pool_conditions + + # 3.1. Encode input prompt + text_encoder_lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt, + prompt_2, + device, + num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=self.clip_skip, + ) + + # 3.2 Encode ip_adapter_image + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 4. Prepare image and controlnet_conditioning_image + image = self.image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32) + + for image_type in control_image_list: + if control_image_list[image_type]: + control_image = self.prepare_control_image( + image=control_image_list[image_type], + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=controlnet.dtype, + do_classifier_free_guidance=self.do_classifier_free_guidance, + guess_mode=guess_mode, + ) + height, width = control_image.shape[-2:] + control_image_list[image_type] = control_image + + # 5. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + self._num_timesteps = len(timesteps) + + # 6. Prepare latent variables + if latents is None: + latents = self.prepare_latents( + image, + latent_timestep, + batch_size, + num_images_per_prompt, + prompt_embeds.dtype, + device, + generator, + True, + ) + + # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7.1 Create tensor stating which controlnets to keep + controlnet_keep = [] + for i in range(len(timesteps)): + controlnet_keep.append( + 1.0 + - float(i / len(timesteps) < control_guidance_start or (i + 1) / len(timesteps) > control_guidance_end) + ) + + # 7.2 Prepare added time ids & embeddings + for image_type in control_image_list: + if isinstance(control_image_list[image_type], torch.Tensor): + original_size = original_size or control_image_list[image_type].shape[-2:] + target_size = target_size or (height, width) + + if negative_original_size is None: + negative_original_size = original_size + if negative_target_size is None: + negative_target_size = target_size + add_text_embeds = pooled_prompt_embeds + + if self.text_encoder_2 is None: + text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) + else: + text_encoder_projection_dim = self.text_encoder_2.config.projection_dim + + add_time_ids, add_neg_time_ids = self._get_add_time_ids( + original_size, + crops_coords_top_left, + target_size, + aesthetic_score, + negative_aesthetic_score, + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + add_time_ids = add_time_ids.repeat(batch_size * num_images_per_prompt, 1) + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) + add_neg_time_ids = add_neg_time_ids.repeat(batch_size * num_images_per_prompt, 1) + add_time_ids = torch.cat([add_neg_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device) + control_type = ( + control_type.reshape(1, -1) + .to(device, dtype=prompt_embeds.dtype) + .repeat(batch_size * num_images_per_prompt * 2, 1) + ) + + # 8. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + added_cond_kwargs = { + "text_embeds": add_text_embeds, + "time_ids": add_time_ids, + } + + # controlnet(s) inference + if guess_mode and self.do_classifier_free_guidance: + # Infer ControlNet only for the conditional batch. + control_model_input = latents + control_model_input = self.scheduler.scale_model_input(control_model_input, t) + controlnet_prompt_embeds = prompt_embeds.chunk(2)[1] + controlnet_added_cond_kwargs = { + "text_embeds": add_text_embeds.chunk(2)[1], + "time_ids": add_time_ids.chunk(2)[1], + } + else: + control_model_input = latent_model_input + controlnet_prompt_embeds = prompt_embeds + controlnet_added_cond_kwargs = added_cond_kwargs + + if isinstance(controlnet_keep[i], list): + cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])] + else: + controlnet_cond_scale = controlnet_conditioning_scale + if isinstance(controlnet_cond_scale, list): + controlnet_cond_scale = controlnet_cond_scale[0] + cond_scale = controlnet_cond_scale * controlnet_keep[i] + down_block_res_samples, mid_block_res_sample = self.controlnet( + control_model_input, + t, + encoder_hidden_states=controlnet_prompt_embeds, + controlnet_cond=control_image_list, + control_type=control_type, + conditioning_scale=cond_scale, + guess_mode=guess_mode, + added_cond_kwargs=controlnet_added_cond_kwargs, + return_dict=False, + ) + + if guess_mode and self.do_classifier_free_guidance: + # Inferred ControlNet only for the conditional batch. + # To apply the output of ControlNet to both the unconditional and conditional batches, + # add 0 to the unconditional batch to keep it unchanged. + down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples] + mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample]) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + added_cond_kwargs["image_embeds"] = image_embeds + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=self.cross_attention_kwargs, + down_block_additional_residuals=down_block_res_samples, + mid_block_additional_residual=mid_block_res_sample, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds) + negative_pooled_prompt_embeds = callback_outputs.pop( + "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds + ) + add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids) + add_neg_time_ids = callback_outputs.pop("add_neg_time_ids", add_neg_time_ids) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + # If we do sequential model offloading, let's offload unet and controlnet + # manually for max memory savings + if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: + self.unet.to("cpu") + self.controlnet.to("cpu") + torch.cuda.empty_cache() + + if not output_type == "latent": + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + + # unscale/denormalize the latents + # denormalize with the mean and std if available and not None + has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None + has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None + if has_latents_mean and has_latents_std: + latents_mean = ( + torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents_std = ( + torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean + else: + latents = latents / self.vae.config.scaling_factor + + image = self.vae.decode(latents, return_dict=False)[0] + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + else: + image = latents + return StableDiffusionXLPipelineOutput(images=image) + + # apply watermark if available + if self.watermark is not None: + image = self.watermark.apply_watermark(image) + + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return StableDiffusionXLPipelineOutput(images=image) diff --git a/src/diffusers/utils/dummy_pt_objects.py b/src/diffusers/utils/dummy_pt_objects.py index 5091ff318f1b..8ebda50adbac 100644 --- a/src/diffusers/utils/dummy_pt_objects.py +++ b/src/diffusers/utils/dummy_pt_objects.py @@ -212,6 +212,21 @@ def from_pretrained(cls, *args, **kwargs): requires_backends(cls, ["torch"]) +class ControlNetUnionModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + class ControlNetXSAdapter(metaclass=DummyObject): _backends = ["torch"] diff --git a/src/diffusers/utils/dummy_torch_and_transformers_objects.py b/src/diffusers/utils/dummy_torch_and_transformers_objects.py index 4fc7cd6aefff..4ddd1fbff4dd 100644 --- a/src/diffusers/utils/dummy_torch_and_transformers_objects.py +++ b/src/diffusers/utils/dummy_torch_and_transformers_objects.py @@ -1967,6 +1967,51 @@ def from_pretrained(cls, *args, **kwargs): requires_backends(cls, ["torch", "transformers"]) +class StableDiffusionXLControlNetUnionImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLControlNetUnionInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLControlNetUnionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + class StableDiffusionXLControlNetXSPipeline(metaclass=DummyObject): _backends = ["torch", "transformers"]