-
Notifications
You must be signed in to change notification settings - Fork 415
/
pipeline_latent_consistency.py
230 lines (199 loc) · 10.9 KB
/
pipeline_latent_consistency.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
# Copyright 2023 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 logging
from typing import Callable, List, Optional, Union
import numpy as np
import torch
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
from .pipeline_stable_diffusion import StableDiffusionPipelineMixin
logger = logging.getLogger(__name__)
class LatentConsistencyPipelineMixin(StableDiffusionPipelineMixin):
# Adapted from https://github.com/huggingface/diffusers/blob/v0.22.0/src/diffusers/pipelines/latent_consistency/pipeline_latent_consistency.py#L264
def __call__(
self,
prompt: Optional[Union[str, List[str]]] = None,
height: Optional[int] = None,
width: Optional[int] = None,
num_inference_steps: int = 4,
original_inference_steps: int = None,
guidance_scale: float = 8.5,
num_images_per_prompt: int = 1,
generator: Optional[np.random.RandomState] = None,
latents: Optional[np.ndarray] = None,
prompt_embeds: Optional[np.ndarray] = None,
output_type: str = "pil",
return_dict: bool = True,
callback: Optional[Callable[[int, int, np.ndarray], None]] = None,
callback_steps: int = 1,
):
r"""
Function invoked when calling the pipeline for generation.
Args:
prompt (`Optional[Union[str, List[str]]]`, defaults to None):
The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
instead.
height (`Optional[int]`, defaults to None):
The height in pixels of the generated image.
width (`Optional[int]`, defaults to None):
The width in pixels of the generated image.
num_inference_steps (`int`, 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`, 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.
num_images_per_prompt (`int`, defaults to 1):
The number of images to generate per prompt.
generator (`Optional[np.random.RandomState]`, defaults to `None`)::
A np.random.RandomState to make generation deterministic.
latents (`Optional[np.ndarray]`, defaults to `None`):
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 (`Optional[np.ndarray]`, defaults to `None`):
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.
output_type (`str`, 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`, defaults to `True`):
Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
plain tuple.
callback (Optional[Callable], defaults to `None`):
A function that will be called every `callback_steps` steps during inference. The function will be
called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
callback_steps (`int`, defaults to 1):
The frequency at which the `callback` function will be called. If not specified, the callback will be
called at every step.
guidance_rescale (`float`, defaults to 0.0):
Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
[Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
Guidance rescale factor should fix overexposure when using zero terminal SNR.
Returns:
[`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
[`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
When returning a tuple, the first element is a list with the generated images, and the second element is a
list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
(nsfw) content, according to the `safety_checker`.
"""
height = height or self.unet.config["sample_size"] * self.vae_scale_factor
width = width or self.unet.config["sample_size"] * self.vae_scale_factor
# Don't need to get negative prompts due to LCM guided distillation
negative_prompt = None
negative_prompt_embeds = None
# check inputs. Raise error if not correct
self.check_inputs(
prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds
)
# define call parameters
if isinstance(prompt, str):
batch_size = 1
elif isinstance(prompt, list):
batch_size = len(prompt)
else:
batch_size = prompt_embeds.shape[0]
if generator is None:
generator = np.random
prompt_embeds = self._encode_prompt(
prompt,
num_images_per_prompt,
False,
negative_prompt,
prompt_embeds=prompt_embeds,
negative_prompt_embeds=negative_prompt_embeds,
)
# set timesteps
self.scheduler.set_timesteps(num_inference_steps, original_inference_steps=original_inference_steps)
timesteps = self.scheduler.timesteps
latents = self.prepare_latents(
batch_size * num_images_per_prompt,
self.unet.config["in_channels"],
height,
width,
prompt_embeds.dtype,
generator,
latents,
)
bs = batch_size * num_images_per_prompt
# get Guidance Scale Embedding
w = np.full(bs, guidance_scale - 1, dtype=prompt_embeds.dtype)
w_embedding = self.get_guidance_scale_embedding(
w, embedding_dim=self.unet.config["time_cond_proj_dim"], dtype=prompt_embeds.dtype
)
# Adapted from diffusers to extend it for other runtimes than ORT
timestep_dtype = self.unet.input_dtype.get("timestep", np.float32)
num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
for i, t in enumerate(self.progress_bar(timesteps)):
timestep = np.array([t], dtype=timestep_dtype)
noise_pred = self.unet(
sample=latents,
timestep=timestep,
encoder_hidden_states=prompt_embeds,
timestep_cond=w_embedding,
)[0]
# compute the previous noisy sample x_t -> x_t-1
latents, denoised = self.scheduler.step(
torch.from_numpy(noise_pred), t, torch.from_numpy(latents), return_dict=False
)
latents, denoised = latents.numpy(), denoised.numpy()
# call the callback, if provided
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
if callback is not None and i % callback_steps == 0:
callback(i, t, latents)
if output_type == "latent":
image = denoised
has_nsfw_concept = None
else:
denoised /= self.vae_decoder.config["scaling_factor"]
# it seems likes there is a strange result for using half-precision vae decoder if batchsize>1
image = np.concatenate(
[self.vae_decoder(latent_sample=denoised[i : i + 1])[0] for i in range(denoised.shape[0])]
)
image, has_nsfw_concept = self.run_safety_checker(image)
if has_nsfw_concept is None:
do_denormalize = [True] * image.shape[0]
else:
do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
if not return_dict:
return (image, has_nsfw_concept)
return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
# Adapted from https://github.com/huggingface/diffusers/blob/v0.22.0/src/diffusers/pipelines/latent_consistency/pipeline_latent_consistency.py#L264
def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=None):
"""
See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
Args:
timesteps (`torch.Tensor`):
generate embedding vectors at these timesteps
embedding_dim (`int`, *optional*, defaults to 512):
dimension of the embeddings to generate
dtype:
data type of the generated embeddings
Returns:
`torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
"""
w = w * 1000
half_dim = embedding_dim // 2
emb = np.log(10000.0) / (half_dim - 1)
emb = np.exp(np.arange(half_dim, dtype=dtype) * -emb)
emb = w[:, None] * emb[None, :]
emb = np.concatenate([np.sin(emb), np.cos(emb)], axis=1)
if embedding_dim % 2 == 1: # zero pad
emb = np.pad(emb, [(0, 0), (0, 1)])
assert emb.shape == (w.shape[0], embedding_dim)
return emb