diff --git a/src/diffusers/__init__.py b/src/diffusers/__init__.py index 85f3b7a127f7..6317433c2e08 100644 --- a/src/diffusers/__init__.py +++ b/src/diffusers/__init__.py @@ -288,6 +288,7 @@ "StableCascadePriorPipeline", "StableDiffusion3ControlNetPipeline", "StableDiffusion3Img2ImgPipeline", + "StableDiffusion3InpaintPipeline", "StableDiffusion3Pipeline", "StableDiffusionAdapterPipeline", "StableDiffusionAttendAndExcitePipeline", @@ -690,6 +691,7 @@ StableCascadePriorPipeline, StableDiffusion3ControlNetPipeline, StableDiffusion3Img2ImgPipeline, + StableDiffusion3InpaintPipeline, StableDiffusion3Pipeline, StableDiffusionAdapterPipeline, StableDiffusionAttendAndExcitePipeline, diff --git a/src/diffusers/pipelines/__init__.py b/src/diffusers/pipelines/__init__.py index aee2c609281f..b966124b46c2 100644 --- a/src/diffusers/pipelines/__init__.py +++ b/src/diffusers/pipelines/__init__.py @@ -243,7 +243,11 @@ "StableDiffusionLDM3DPipeline", ] ) - _import_structure["stable_diffusion_3"] = ["StableDiffusion3Pipeline", "StableDiffusion3Img2ImgPipeline"] + _import_structure["stable_diffusion_3"] = [ + "StableDiffusion3Pipeline", + "StableDiffusion3Img2ImgPipeline", + "StableDiffusion3InpaintPipeline", + ] _import_structure["stable_diffusion_attend_and_excite"] = ["StableDiffusionAttendAndExcitePipeline"] _import_structure["stable_diffusion_safe"] = ["StableDiffusionPipelineSafe"] _import_structure["stable_diffusion_sag"] = ["StableDiffusionSAGPipeline"] @@ -523,7 +527,11 @@ StableUnCLIPImg2ImgPipeline, StableUnCLIPPipeline, ) - from .stable_diffusion_3 import StableDiffusion3Img2ImgPipeline, StableDiffusion3Pipeline + from .stable_diffusion_3 import ( + StableDiffusion3Img2ImgPipeline, + StableDiffusion3InpaintPipeline, + StableDiffusion3Pipeline, + ) from .stable_diffusion_attend_and_excite import StableDiffusionAttendAndExcitePipeline from .stable_diffusion_diffedit import StableDiffusionDiffEditPipeline from .stable_diffusion_gligen import StableDiffusionGLIGENPipeline, StableDiffusionGLIGENTextImagePipeline diff --git a/src/diffusers/pipelines/stable_diffusion_3/__init__.py b/src/diffusers/pipelines/stable_diffusion_3/__init__.py index 0de2bbdb5434..b0604589a208 100644 --- a/src/diffusers/pipelines/stable_diffusion_3/__init__.py +++ b/src/diffusers/pipelines/stable_diffusion_3/__init__.py @@ -25,6 +25,7 @@ else: _import_structure["pipeline_stable_diffusion_3"] = ["StableDiffusion3Pipeline"] _import_structure["pipeline_stable_diffusion_3_img2img"] = ["StableDiffusion3Img2ImgPipeline"] + _import_structure["pipeline_stable_diffusion_3_inpaint"] = ["StableDiffusion3InpaintPipeline"] if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: try: @@ -35,6 +36,7 @@ else: from .pipeline_stable_diffusion_3 import StableDiffusion3Pipeline from .pipeline_stable_diffusion_3_img2img import StableDiffusion3Img2ImgPipeline + from .pipeline_stable_diffusion_3_inpaint import StableDiffusion3InpaintPipeline else: import sys diff --git a/src/diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3_inpaint.py b/src/diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3_inpaint.py new file mode 100644 index 000000000000..f340223cb38b --- /dev/null +++ b/src/diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3_inpaint.py @@ -0,0 +1,1173 @@ +# Copyright 2024 Stability AI 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 Callable, Dict, List, Optional, Union + +import torch +from transformers import ( + CLIPTextModelWithProjection, + CLIPTokenizer, + T5EncoderModel, + T5TokenizerFast, +) + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...models.autoencoders import AutoencoderKL +from ...models.transformers import SD3Transformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import ( + is_torch_xla_available, + logging, + replace_example_docstring, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import StableDiffusion3PipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusion3InpaintPipeline + >>> from diffusers.utils import load_image + + >>> pipe = StableDiffusion3InpaintPipeline.from_pretrained( + ... "stabilityai/stable-diffusion-3-medium-diffusers", torch_dtype=torch.float16 + ... ) + >>> pipe.to("cuda") + >>> prompt = "Face of a yellow cat, high resolution, sitting on a park bench" + >>> img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png" + >>> mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png" + >>> source = load_image(img_url) + >>> mask = load_image(mask_url) + >>> image = pipe(prompt=prompt, image=source, mask_image=mask).images[0] + >>> image.save("sd3_inpainting.png") + ``` +""" + + +# 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") + + +# 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, +): + """ + 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 StableDiffusion3InpaintPipeline(DiffusionPipeline): + r""" + Args: + transformer ([`SD3Transformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModelWithProjection`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant, + with an additional added projection layer that is initialized with a diagonal matrix with the `hidden_size` + as its dimension. + text_encoder_2 ([`CLIPTextModelWithProjection`]): + [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. + text_encoder_3 ([`T5EncoderModel`]): + Frozen text-encoder. Stable Diffusion 3 uses + [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel), specifically the + [t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) 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). + tokenizer_3 (`T5TokenizerFast`): + Tokenizer of class + [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer). + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->text_encoder_3->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds", "negative_pooled_prompt_embeds"] + + def __init__( + self, + transformer: SD3Transformer2DModel, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer_2: CLIPTokenizer, + text_encoder_3: T5EncoderModel, + tokenizer_3: T5TokenizerFast, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + text_encoder_3=text_encoder_3, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + tokenizer_3=tokenizer_3, + transformer=transformer, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) + self.image_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor, vae_latent_channels=self.vae.config.latent_channels + ) + self.mask_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor, + vae_latent_channels=self.vae.config.latent_channels, + do_normalize=False, + do_binarize=True, + do_convert_grayscale=True, + ) + self.tokenizer_max_length = self.tokenizer.model_max_length + self.default_sample_size = self.transformer.config.sample_size + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: Union[str, List[str]] = None, + num_images_per_prompt: int = 1, + max_sequence_length: int = 256, + device: Optional[torch.device] = None, + dtype: Optional[torch.dtype] = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if self.text_encoder_3 is None: + return torch.zeros( + ( + batch_size * num_images_per_prompt, + self.tokenizer_max_length, + self.transformer.config.joint_attention_dim, + ), + device=device, + dtype=dtype, + ) + + text_inputs = self.tokenizer_3( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_3(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 = self.tokenizer_3.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_3(text_input_ids.to(device))[0] + + dtype = self.text_encoder_3.dtype + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + + # duplicate text embeddings and attention mask 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(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline._get_clip_prompt_embeds + def _get_clip_prompt_embeds( + self, + prompt: Union[str, List[str]], + num_images_per_prompt: int = 1, + device: Optional[torch.device] = None, + clip_skip: Optional[int] = None, + clip_model_index: int = 0, + ): + device = device or self._execution_device + + clip_tokenizers = [self.tokenizer, self.tokenizer_2] + clip_text_encoders = [self.text_encoder, self.text_encoder_2] + + tokenizer = clip_tokenizers[clip_model_index] + text_encoder = clip_text_encoders[clip_model_index] + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_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[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_max_length} tokens: {removed_text}" + ) + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + pooled_prompt_embeds = prompt_embeds[0] + + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + _, 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(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt, 1) + pooled_prompt_embeds = pooled_prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds, pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.encode_prompt + def encode_prompt( + self, + prompt: Union[str, List[str]], + prompt_2: Union[str, List[str]], + prompt_3: Union[str, List[str]], + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: Optional[Union[str, List[str]]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + negative_prompt_3: Optional[Union[str, List[str]]] = None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + clip_skip: Optional[int] = None, + max_sequence_length: int = 256, + ): + r""" + + 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 all text-encoders + prompt_3 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is + used in all 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 all the text-encoders. + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and + `text_encoder_3`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.FloatTensor`, *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.FloatTensor`, *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.FloatTensor`, *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.FloatTensor`, *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. + 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 + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + prompt_3 = prompt_3 or prompt + prompt_3 = [prompt_3] if isinstance(prompt_3, str) else prompt_3 + + prompt_embed, pooled_prompt_embed = self._get_clip_prompt_embeds( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=clip_skip, + clip_model_index=0, + ) + prompt_2_embed, pooled_prompt_2_embed = self._get_clip_prompt_embeds( + prompt=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=clip_skip, + clip_model_index=1, + ) + clip_prompt_embeds = torch.cat([prompt_embed, prompt_2_embed], dim=-1) + + t5_prompt_embed = self._get_t5_prompt_embeds( + prompt=prompt_3, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + clip_prompt_embeds = torch.nn.functional.pad( + clip_prompt_embeds, (0, t5_prompt_embed.shape[-1] - clip_prompt_embeds.shape[-1]) + ) + + prompt_embeds = torch.cat([clip_prompt_embeds, t5_prompt_embed], dim=-2) + pooled_prompt_embeds = torch.cat([pooled_prompt_embed, pooled_prompt_2_embed], dim=-1) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + negative_prompt_3 = negative_prompt_3 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 + ) + negative_prompt_3 = ( + batch_size * [negative_prompt_3] if isinstance(negative_prompt_3, str) else negative_prompt_3 + ) + + 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`." + ) + + negative_prompt_embed, negative_pooled_prompt_embed = self._get_clip_prompt_embeds( + negative_prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=None, + clip_model_index=0, + ) + negative_prompt_2_embed, negative_pooled_prompt_2_embed = self._get_clip_prompt_embeds( + negative_prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=None, + clip_model_index=1, + ) + negative_clip_prompt_embeds = torch.cat([negative_prompt_embed, negative_prompt_2_embed], dim=-1) + + t5_negative_prompt_embed = self._get_t5_prompt_embeds( + prompt=negative_prompt_3, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + negative_clip_prompt_embeds = torch.nn.functional.pad( + negative_clip_prompt_embeds, + (0, t5_negative_prompt_embed.shape[-1] - negative_clip_prompt_embeds.shape[-1]), + ) + + negative_prompt_embeds = torch.cat([negative_clip_prompt_embeds, t5_negative_prompt_embed], dim=-2) + negative_pooled_prompt_embeds = torch.cat( + [negative_pooled_prompt_embed, negative_pooled_prompt_2_embed], dim=-1 + ) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.check_inputs + def check_inputs( + self, + prompt, + prompt_2, + prompt_3, + strength, + negative_prompt=None, + negative_prompt_2=None, + negative_prompt_3=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + 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_3 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_3`: {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)}") + elif prompt_3 is not None and (not isinstance(prompt_3, str) and not isinstance(prompt_3, list)): + raise ValueError(f"`prompt_3` has to be of type `str` or `list` but is {type(prompt_3)}") + + 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." + ) + elif negative_prompt_3 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_3`: {negative_prompt_3} 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`." + ) + + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(num_inference_steps * strength, num_inference_steps) + + t_start = int(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 + + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + image=None, + timestep=None, + is_strength_max=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] == 16: + 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: + 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.scale_noise(image_latents, timestep, noise) + else: + noise = latents.to(device) + latents = noise + + outputs = (latents,) + + if return_noise: + outputs += (noise,) + + if return_image_latents: + outputs += (image_latents,) + + return outputs + + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + 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) + + image_latents = (image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + return image_latents + + def prepare_mask_latents( + self, + mask, + masked_image, + batch_size, + num_images_per_prompt, + 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) + + batch_size = batch_size * num_images_per_prompt + + masked_image = masked_image.to(device=device, dtype=dtype) + + if masked_image.shape[1] == 16: + masked_image_latents = masked_image + else: + masked_image_latents = retrieve_latents(self.vae.encode(masked_image), generator=generator) + + masked_image_latents = (masked_image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + # 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) + 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) + + mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask + 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 + + @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 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, + prompt_3: Optional[Union[str, List[str]]] = None, + image: PipelineImageInput = None, + mask_image: PipelineImageInput = None, + masked_image_latents: PipelineImageInput = None, + height: int = None, + width: int = None, + padding_mask_crop: Optional[int] = None, + strength: float = 0.6, + num_inference_steps: int = 50, + timesteps: List[int] = None, + guidance_scale: float = 7.0, + negative_prompt: Optional[Union[str, List[str]]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + negative_prompt_3: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + 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, + output_type: Optional[str] = "pil", + return_dict: bool = True, + clip_skip: Optional[int] = None, + callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + max_sequence_length: int = 256, + ): + 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 `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead + prompt_3 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is + will be used instead + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + mask_image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to mask `image`. White pixels in the mask + are repainted while black pixels are preserved. If `mask_image` is a PIL image, it is converted to a + single channel (luminance) before use. If it's a numpy array or pytorch tensor, it should contain one + color channel (L) instead of 3, so the expected shape for pytorch tensor would be `(B, 1, H, W)`, `(B, + H, W)`, `(1, H, W)`, `(H, W)`. And for numpy array would be for `(B, H, W, 1)`, `(B, H, W)`, `(H, W, + 1)`, or `(H, W)`. + mask_image_latent (`torch.Tensor`, `List[torch.Tensor]`): + `Tensor` representing an image batch to mask `image` generated by VAE. If not provided, the mask + latents tensor will ge generated by `mask_image`. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + 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 1.0): + 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. + 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. + guidance_scale (`float`, *optional*, defaults to 5.0): + 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 instead + negative_prompt_3 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and + `text_encoder_3`. If not defined, `negative_prompt` is used instead + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + 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.FloatTensor`, *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.FloatTensor`, *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.FloatTensor`, *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.FloatTensor`, *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.FloatTensor`, *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. + 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_xl.StableDiffusionXLPipelineOutput`] instead + of a plain tuple. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + 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. + max_sequence_length (`int` defaults to 256): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] if `return_dict` is True, otherwise a + `tuple`. When returning a tuple, the first element is a list with the generated images. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + height = height or self.transformer.config.sample_size * self.vae_scale_factor + width = width or self.transformer.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + prompt_3, + strength, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + negative_prompt_3=negative_prompt_3, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._clip_skip = clip_skip + 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 + + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_3=prompt_3, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + negative_prompt_3=negative_prompt_3, + do_classifier_free_guidance=self.do_classifier_free_guidance, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + device=device, + clip_skip=self.clip_skip, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0) + + # 3. Prepare timesteps + timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + # 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." + ) + 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 + + # 4. Preprocess mask and image + if padding_mask_crop is not None: + 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. Prepare latent variables + num_channels_latents = self.vae.config.latent_channels + num_channels_transformer = self.transformer.config.in_channels + return_image_latents = num_channels_transformer == 16 + + 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, + return_noise=True, + return_image_latents=return_image_latents, + ) + + if return_image_latents: + latents, noise, image_latents = latents_outputs + else: + latents, noise = latents_outputs + + # 6. Prepare mask latent variables + mask_condition = self.mask_processor.preprocess( + mask_image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords + ) + + if masked_image_latents is None: + masked_image = init_image * (mask_condition < 0.5) + else: + masked_image = masked_image_latents + + mask, masked_image_latents = self.prepare_mask_latents( + mask_condition, + masked_image, + batch_size, + num_images_per_prompt, + height, + width, + prompt_embeds.dtype, + device, + generator, + self.do_classifier_free_guidance, + ) + + # match the inpainting pipeline and will be updated with input + mask inpainting model later + if num_channels_transformer == 33: + # default case for runwayml/stable-diffusion-inpainting + num_channels_mask = mask.shape[1] + num_channels_masked_image = masked_image_latents.shape[1] + if ( + num_channels_latents + num_channels_mask + num_channels_masked_image + != self.transformer.config.in_channels + ): + raise ValueError( + f"Incorrect configuration settings! The config of `pipeline.transformer`: {self.transformer.config} expects" + f" {self.transformer.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +" + f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}" + f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of" + " `pipeline.transformer` or your `mask_image` or `image` input." + ) + elif num_channels_transformer != 16: + raise ValueError( + f"The transformer {self.transformer.__class__} should have 16 input channels or 33 input channels, not {self.transformer.config.in_channels}." + ) + + # 7. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + 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 + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]) + + if num_channels_transformer == 33: + latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1) + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + pooled_projections=pooled_prompt_embeds, + 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 + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + if num_channels_transformer == 16: + 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.scale_noise( + init_latents_proper, torch.tensor([noise_timestep]), noise + ) + + latents = (1 - init_mask) * init_latents_proper + init_mask * latents + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + 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) + negative_pooled_prompt_embeds = callback_outputs.pop( + "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds + ) + mask = callback_outputs.pop("mask", mask) + masked_image_latents = callback_outputs.pop("masked_image_latents", masked_image_latents) + + # 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 XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[ + 0 + ] + else: + image = latents + + do_denormalize = [True] * image.shape[0] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + 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 StableDiffusion3PipelineOutput(images=image) diff --git a/src/diffusers/utils/dummy_torch_and_transformers_objects.py b/src/diffusers/utils/dummy_torch_and_transformers_objects.py index 482ac39de919..8a01882e2acc 100644 --- a/src/diffusers/utils/dummy_torch_and_transformers_objects.py +++ b/src/diffusers/utils/dummy_torch_and_transformers_objects.py @@ -962,6 +962,21 @@ def from_pretrained(cls, *args, **kwargs): requires_backends(cls, ["torch", "transformers"]) +class StableDiffusion3InpaintPipeline(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 StableDiffusion3Pipeline(metaclass=DummyObject): _backends = ["torch", "transformers"] diff --git a/tests/pipelines/stable_diffusion_3/test_pipeline_stable_diffusion_3_inpaint.py b/tests/pipelines/stable_diffusion_3/test_pipeline_stable_diffusion_3_inpaint.py new file mode 100644 index 000000000000..464ef6d017df --- /dev/null +++ b/tests/pipelines/stable_diffusion_3/test_pipeline_stable_diffusion_3_inpaint.py @@ -0,0 +1,199 @@ +import random +import unittest + +import numpy as np +import torch +from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModelWithProjection, CLIPTokenizer, T5EncoderModel + +from diffusers import ( + AutoencoderKL, + FlowMatchEulerDiscreteScheduler, + SD3Transformer2DModel, + StableDiffusion3InpaintPipeline, +) +from diffusers.utils.testing_utils import ( + enable_full_determinism, + floats_tensor, + torch_device, +) + +from ..pipeline_params import ( + TEXT_GUIDED_IMAGE_INPAINTING_BATCH_PARAMS, + TEXT_GUIDED_IMAGE_INPAINTING_PARAMS, + TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS, +) +from ..test_pipelines_common import PipelineLatentTesterMixin, PipelineTesterMixin + + +enable_full_determinism() + + +class StableDiffusion3InpaintPipelineFastTests(PipelineLatentTesterMixin, unittest.TestCase, PipelineTesterMixin): + pipeline_class = StableDiffusion3InpaintPipeline + params = TEXT_GUIDED_IMAGE_INPAINTING_PARAMS + required_optional_params = PipelineTesterMixin.required_optional_params + batch_params = TEXT_GUIDED_IMAGE_INPAINTING_BATCH_PARAMS + image_params = frozenset( + [] + ) # TO-DO: update image_params once pipeline is refactored with VaeImageProcessor.preprocess + image_latents_params = frozenset([]) + callback_cfg_params = TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS.union({"mask", "masked_image_latents"}) + + def get_dummy_components(self): + torch.manual_seed(0) + transformer = SD3Transformer2DModel( + sample_size=32, + patch_size=1, + in_channels=16, + num_layers=1, + attention_head_dim=8, + num_attention_heads=4, + joint_attention_dim=32, + caption_projection_dim=32, + pooled_projection_dim=64, + out_channels=16, + ) + clip_text_encoder_config = CLIPTextConfig( + bos_token_id=0, + eos_token_id=2, + hidden_size=32, + intermediate_size=37, + layer_norm_eps=1e-05, + num_attention_heads=4, + num_hidden_layers=5, + pad_token_id=1, + vocab_size=1000, + hidden_act="gelu", + projection_dim=32, + ) + + torch.manual_seed(0) + text_encoder = CLIPTextModelWithProjection(clip_text_encoder_config) + + torch.manual_seed(0) + text_encoder_2 = CLIPTextModelWithProjection(clip_text_encoder_config) + + text_encoder_3 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5") + + tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip") + tokenizer_2 = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip") + tokenizer_3 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5") + + torch.manual_seed(0) + vae = AutoencoderKL( + sample_size=32, + in_channels=3, + out_channels=3, + block_out_channels=(4,), + layers_per_block=1, + latent_channels=16, + norm_num_groups=1, + use_quant_conv=False, + use_post_quant_conv=False, + shift_factor=0.0609, + scaling_factor=1.5035, + ) + + scheduler = FlowMatchEulerDiscreteScheduler() + + return { + "scheduler": scheduler, + "text_encoder": text_encoder, + "text_encoder_2": text_encoder_2, + "text_encoder_3": text_encoder_3, + "tokenizer": tokenizer, + "tokenizer_2": tokenizer_2, + "tokenizer_3": tokenizer_3, + "transformer": transformer, + "vae": vae, + } + + def get_dummy_inputs(self, device, seed=0): + image = floats_tensor((1, 3, 32, 32), rng=random.Random(seed)).to(device) + mask_image = torch.ones((1, 1, 32, 32)).to(device) + if str(device).startswith("mps"): + generator = torch.manual_seed(seed) + else: + generator = torch.Generator(device="cpu").manual_seed(seed) + + inputs = { + "prompt": "A painting of a squirrel eating a burger", + "image": image, + "mask_image": mask_image, + "height": 32, + "width": 32, + "generator": generator, + "num_inference_steps": 2, + "guidance_scale": 5.0, + "output_type": "np", + "strength": 0.8, + } + return inputs + + def test_stable_diffusion_3_inpaint_different_prompts(self): + pipe = self.pipeline_class(**self.get_dummy_components()).to(torch_device) + + inputs = self.get_dummy_inputs(torch_device) + output_same_prompt = pipe(**inputs).images[0] + + inputs = self.get_dummy_inputs(torch_device) + inputs["prompt_2"] = "a different prompt" + inputs["prompt_3"] = "another different prompt" + output_different_prompts = pipe(**inputs).images[0] + + max_diff = np.abs(output_same_prompt - output_different_prompts).max() + + # Outputs should be different here + assert max_diff > 1e-2 + + def test_stable_diffusion_3_inpaint_different_negative_prompts(self): + pipe = self.pipeline_class(**self.get_dummy_components()).to(torch_device) + + inputs = self.get_dummy_inputs(torch_device) + output_same_prompt = pipe(**inputs).images[0] + + inputs = self.get_dummy_inputs(torch_device) + inputs["negative_prompt_2"] = "deformed" + inputs["negative_prompt_3"] = "blurry" + output_different_prompts = pipe(**inputs).images[0] + + max_diff = np.abs(output_same_prompt - output_different_prompts).max() + + # Outputs should be different here + assert max_diff > 1e-2 + + def test_stable_diffusion_3_inpaint_prompt_embeds(self): + pipe = self.pipeline_class(**self.get_dummy_components()).to(torch_device) + inputs = self.get_dummy_inputs(torch_device) + + output_with_prompt = pipe(**inputs).images[0] + + inputs = self.get_dummy_inputs(torch_device) + prompt = inputs.pop("prompt") + + do_classifier_free_guidance = inputs["guidance_scale"] > 1 + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = pipe.encode_prompt( + prompt, + prompt_2=None, + prompt_3=None, + do_classifier_free_guidance=do_classifier_free_guidance, + device=torch_device, + ) + output_with_embeds = pipe( + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + **inputs, + ).images[0] + + max_diff = np.abs(output_with_prompt - output_with_embeds).max() + assert max_diff < 1e-4 + + def test_multi_vae(self): + pass