ameerazam08
commited on
Commit
•
84bf924
1
Parent(s):
80e89cb
Upload 9 files
Browse files- .gitattributes +3 -0
- LICENSE +21 -0
- figures/.DS_Store +0 -0
- figures/main_figure.jpg +3 -0
- figures/sample_bunny_2K.png +3 -0
- figures/sample_icecream_4K.png +3 -0
- pipeline_diffusehigh_sdxl.py +798 -0
- requirements.txt +5 -0
- utils/utils.py +11 -0
.gitattributes
CHANGED
@@ -33,3 +33,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
36 |
+
figures/main_figure.jpg filter=lfs diff=lfs merge=lfs -text
|
37 |
+
figures/sample_bunny_2K.png filter=lfs diff=lfs merge=lfs -text
|
38 |
+
figures/sample_icecream_4K.png filter=lfs diff=lfs merge=lfs -text
|
LICENSE
ADDED
@@ -0,0 +1,21 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
MIT License
|
2 |
+
|
3 |
+
Copyright (c) 2024 yhyun225
|
4 |
+
|
5 |
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
6 |
+
of this software and associated documentation files (the "Software"), to deal
|
7 |
+
in the Software without restriction, including without limitation the rights
|
8 |
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
9 |
+
copies of the Software, and to permit persons to whom the Software is
|
10 |
+
furnished to do so, subject to the following conditions:
|
11 |
+
|
12 |
+
The above copyright notice and this permission notice shall be included in all
|
13 |
+
copies or substantial portions of the Software.
|
14 |
+
|
15 |
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
16 |
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
17 |
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
18 |
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
19 |
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
20 |
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
21 |
+
SOFTWARE.
|
figures/.DS_Store
ADDED
Binary file (6.15 kB). View file
|
|
figures/main_figure.jpg
ADDED
Git LFS Details
|
figures/sample_bunny_2K.png
ADDED
Git LFS Details
|
figures/sample_icecream_4K.png
ADDED
Git LFS Details
|
pipeline_diffusehigh_sdxl.py
ADDED
@@ -0,0 +1,798 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import numpy as np
|
2 |
+
import PIL
|
3 |
+
import inspect
|
4 |
+
import os
|
5 |
+
from tqdm import tqdm
|
6 |
+
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
|
7 |
+
|
8 |
+
import torch
|
9 |
+
import torch.nn.functional as F
|
10 |
+
|
11 |
+
from transformers import (
|
12 |
+
CLIPImageProcessor,
|
13 |
+
CLIPTextModel,
|
14 |
+
CLIPTextModelWithProjection,
|
15 |
+
CLIPTokenizer,
|
16 |
+
CLIPVisionModelWithProjection,
|
17 |
+
)
|
18 |
+
|
19 |
+
from diffusers.schedulers import KarrasDiffusionSchedulers
|
20 |
+
from diffusers.image_processor import PipelineImageInput
|
21 |
+
from diffusers import (
|
22 |
+
AutoencoderKL,
|
23 |
+
UNet2DConditionModel,
|
24 |
+
StableDiffusionXLPipeline,
|
25 |
+
DDIMScheduler,
|
26 |
+
EulerDiscreteScheduler,
|
27 |
+
)
|
28 |
+
from diffusers.utils import BaseOutput
|
29 |
+
from diffusers.utils.torch_utils import randn_tensor
|
30 |
+
from pytorch_wavelets import DWTForward, DWTInverse
|
31 |
+
from torchvision.transforms import GaussianBlur
|
32 |
+
|
33 |
+
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
|
34 |
+
def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
|
35 |
+
"""
|
36 |
+
Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
|
37 |
+
Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
|
38 |
+
"""
|
39 |
+
std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
|
40 |
+
std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
|
41 |
+
# rescale the results from guidance (fixes overexposure)
|
42 |
+
noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
|
43 |
+
# mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
|
44 |
+
noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
|
45 |
+
return noise_cfg
|
46 |
+
|
47 |
+
|
48 |
+
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
|
49 |
+
def retrieve_timesteps(
|
50 |
+
scheduler,
|
51 |
+
num_inference_steps: Optional[int] = None,
|
52 |
+
device: Optional[Union[str, torch.device]] = None,
|
53 |
+
timesteps: Optional[List[int]] = None,
|
54 |
+
**kwargs,
|
55 |
+
):
|
56 |
+
"""
|
57 |
+
Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
|
58 |
+
custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
|
59 |
+
|
60 |
+
Args:
|
61 |
+
scheduler (`SchedulerMixin`):
|
62 |
+
The scheduler to get timesteps from.
|
63 |
+
num_inference_steps (`int`):
|
64 |
+
The number of diffusion steps used when generating samples with a pre-trained model. If used,
|
65 |
+
`timesteps` must be `None`.
|
66 |
+
device (`str` or `torch.device`, *optional*):
|
67 |
+
The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
|
68 |
+
timesteps (`List[int]`, *optional*):
|
69 |
+
Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
|
70 |
+
timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
|
71 |
+
must be `None`.
|
72 |
+
|
73 |
+
Returns:
|
74 |
+
`Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
|
75 |
+
second element is the number of inference steps.
|
76 |
+
"""
|
77 |
+
if timesteps is not None:
|
78 |
+
accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
|
79 |
+
if not accepts_timesteps:
|
80 |
+
raise ValueError(
|
81 |
+
f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
|
82 |
+
f" timestep schedules. Please check whether you are using the correct scheduler."
|
83 |
+
)
|
84 |
+
scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
|
85 |
+
timesteps = scheduler.timesteps
|
86 |
+
num_inference_steps = len(timesteps)
|
87 |
+
else:
|
88 |
+
scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
|
89 |
+
timesteps = scheduler.timesteps
|
90 |
+
return timesteps, num_inference_steps
|
91 |
+
|
92 |
+
|
93 |
+
def gaussian_blur_image_sharpening(image, kernel_size=3, sigma=(0.1, 2.0), alpha=1):
|
94 |
+
gaussian_blur = GaussianBlur(kernel_size=kernel_size, sigma=sigma)
|
95 |
+
image_blurred = gaussian_blur(image)
|
96 |
+
image_sharpened = (alpha + 1) * image - alpha * image_blurred
|
97 |
+
|
98 |
+
return image_sharpened
|
99 |
+
|
100 |
+
|
101 |
+
class DiffuseHighSDXLPipelineOutput(BaseOutput):
|
102 |
+
"""
|
103 |
+
Output class for Stable Diffusion pipelines.
|
104 |
+
|
105 |
+
Args:
|
106 |
+
images (`List[PIL.Image.Image]` or `np.ndarray`)
|
107 |
+
List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width,
|
108 |
+
num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline.
|
109 |
+
"""
|
110 |
+
|
111 |
+
images: Union[List[PIL.Image.Image], np.ndarray]
|
112 |
+
guidance_images: Union[List[PIL.Image.Image], np.ndarray]
|
113 |
+
|
114 |
+
|
115 |
+
class DiffuseHighSDXLPipeline(StableDiffusionXLPipeline):
|
116 |
+
def __init__(
|
117 |
+
self,
|
118 |
+
vae: AutoencoderKL,
|
119 |
+
text_encoder: CLIPTextModel,
|
120 |
+
text_encoder_2: CLIPTextModelWithProjection,
|
121 |
+
tokenizer: CLIPTokenizer,
|
122 |
+
tokenizer_2: CLIPTokenizer,
|
123 |
+
unet: UNet2DConditionModel,
|
124 |
+
scheduler: KarrasDiffusionSchedulers,
|
125 |
+
image_encoder: CLIPVisionModelWithProjection = None,
|
126 |
+
feature_extractor: CLIPImageProcessor = None,
|
127 |
+
force_zeros_for_empty_prompt: bool = True,
|
128 |
+
add_watermarker: Optional[bool] = None,
|
129 |
+
):
|
130 |
+
super().__init__(
|
131 |
+
vae=vae,
|
132 |
+
text_encoder=text_encoder,
|
133 |
+
text_encoder_2=text_encoder_2,
|
134 |
+
tokenizer=tokenizer,
|
135 |
+
tokenizer_2=tokenizer_2,
|
136 |
+
unet=unet,
|
137 |
+
scheduler=scheduler,
|
138 |
+
image_encoder=image_encoder,
|
139 |
+
feature_extractor=feature_extractor,
|
140 |
+
force_zeros_for_empty_prompt=force_zeros_for_empty_prompt,
|
141 |
+
add_watermarker=add_watermarker
|
142 |
+
)
|
143 |
+
|
144 |
+
def _encode_vae_image(
|
145 |
+
self,
|
146 |
+
image: torch.Tensor,
|
147 |
+
normalize: bool = True,
|
148 |
+
):
|
149 |
+
if normalize:
|
150 |
+
image = image * 2 - 1
|
151 |
+
|
152 |
+
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
|
153 |
+
|
154 |
+
if needs_upcasting:
|
155 |
+
self.upcast_vae()
|
156 |
+
|
157 |
+
image = image.to(self.device)
|
158 |
+
latents = self.vae.encode(image).latent_dist.mode() * self.vae.config.scaling_factor
|
159 |
+
|
160 |
+
if needs_upcasting:
|
161 |
+
self.vae.to(dtype=torch.float16)
|
162 |
+
|
163 |
+
return latents.to(self.dtype)
|
164 |
+
|
165 |
+
def _decode_vae_latent(
|
166 |
+
self,
|
167 |
+
latents: torch.Tensor,
|
168 |
+
output_type: Optional[str] = 'pt',
|
169 |
+
):
|
170 |
+
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
|
171 |
+
|
172 |
+
if needs_upcasting:
|
173 |
+
self.upcast_vae()
|
174 |
+
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
|
175 |
+
|
176 |
+
latents = latents.to(self.device)
|
177 |
+
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
|
178 |
+
image = self.image_processor.postprocess(image, output_type=output_type)
|
179 |
+
|
180 |
+
if needs_upcasting:
|
181 |
+
self.vae.to(dtype=torch.float16)
|
182 |
+
|
183 |
+
return image
|
184 |
+
|
185 |
+
def edm_scheduler_step(
|
186 |
+
self,
|
187 |
+
model_output: torch.FloatTensor,
|
188 |
+
timestep: Union[float, torch.FloatTensor],
|
189 |
+
sample: torch.FloatTensor,
|
190 |
+
s_churn: float = 0.0,
|
191 |
+
s_tmin: float = 0.0,
|
192 |
+
s_tmax: float = 0.0,
|
193 |
+
s_noise: float = 1.0,
|
194 |
+
LL_guidance: Optional[torch.FloatTensor] = None,
|
195 |
+
generator: Optional[torch.Generator] = None,
|
196 |
+
return_pred_original_sample: bool = False,
|
197 |
+
):
|
198 |
+
assert isinstance(self.scheduler, EulerDiscreteScheduler)
|
199 |
+
config = self.scheduler.config
|
200 |
+
|
201 |
+
if self.scheduler.step_index is None:
|
202 |
+
self.scheduler._init_step_index(timestep)
|
203 |
+
|
204 |
+
step_index = self.scheduler.step_index
|
205 |
+
|
206 |
+
sigma = self.scheduler.sigmas[step_index]
|
207 |
+
|
208 |
+
gamma = min(s_churn / (len(self.scheduler.sigmas) - 1), 2**0.5 - 1) if s_tmin <= sigma <= s_tmax else 0.0
|
209 |
+
|
210 |
+
noise = randn_tensor(
|
211 |
+
model_output.shape, dtype=model_output.dtype, device=model_output.device, generator=generator
|
212 |
+
)
|
213 |
+
|
214 |
+
eps = noise * s_noise
|
215 |
+
sigma_hat = sigma * (gamma + 1)
|
216 |
+
|
217 |
+
if gamma > 0:
|
218 |
+
sample = sample + eps * (sigma_hat**2 - sigma**2) ** 0.5
|
219 |
+
|
220 |
+
# 1. compute predicted original sample (x_0) from sigma-scaled predicted noise
|
221 |
+
if config.prediction_type == "original_sample" or config.prediction_type == "sample":
|
222 |
+
pred_original_sample = model_output
|
223 |
+
elif config.prediction_type == "epsilon":
|
224 |
+
pred_original_sample = sample - sigma_hat * model_output
|
225 |
+
elif config.prediction_type == "v_prediction":
|
226 |
+
# denoised = model_output * c_out + input * c_skip
|
227 |
+
pred_original_sample = model_output * (-sigma / (sigma**2 + 1) ** 0.5) + (sample / (sigma**2 + 1))
|
228 |
+
else:
|
229 |
+
raise ValueError(
|
230 |
+
f"prediction_type given as {config.prediction_type} must be one of `epsilon`, or `v_prediction`"
|
231 |
+
)
|
232 |
+
|
233 |
+
# 2. If gudiance LL component is given, perform structural guidance
|
234 |
+
if LL_guidance is not None:
|
235 |
+
pred_original_image = self._decode_vae_latent(pred_original_sample, output_type='pt')
|
236 |
+
|
237 |
+
_, HH = self.DWT(pred_original_image)
|
238 |
+
coeffs = (LL_guidance, HH)
|
239 |
+
pred_original_image = self.iDWT(coeffs)
|
240 |
+
|
241 |
+
pred_original_sample = self._encode_vae_image(pred_original_image)
|
242 |
+
|
243 |
+
# 3. Convert to an ODE derivative
|
244 |
+
derivative = (sample - pred_original_sample) / sigma_hat
|
245 |
+
|
246 |
+
dt = self.scheduler.sigmas[self.scheduler.step_index + 1] - sigma_hat
|
247 |
+
|
248 |
+
prev_sample = sample + derivative * dt
|
249 |
+
|
250 |
+
self.scheduler._step_index += 1
|
251 |
+
|
252 |
+
if return_pred_original_sample:
|
253 |
+
return (prev_sample, pred_original_sample)
|
254 |
+
|
255 |
+
return (prev_sample, )
|
256 |
+
|
257 |
+
|
258 |
+
@torch.no_grad()
|
259 |
+
def __call__(
|
260 |
+
self,
|
261 |
+
prompt: Union[str, List[str]] = None,
|
262 |
+
prompt_2: Optional[Union[str, List[str]]] = None,
|
263 |
+
num_inference_steps: int = 50,
|
264 |
+
timesteps: List[int] = None,
|
265 |
+
denoising_end: Optional[float] = None,
|
266 |
+
guidance_scale: float = 5,
|
267 |
+
negative_prompt: Optional[Union[str, List[str]]] = None,
|
268 |
+
negative_prompt_2: Optional[Union[str, List[str]]] = None,
|
269 |
+
num_images_per_prompt: Optional[int] = 1,
|
270 |
+
eta: float = 0.0,
|
271 |
+
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
|
272 |
+
latents: Optional[torch.FloatTensor] = None,
|
273 |
+
prompt_embeds: Optional[torch.FloatTensor] = None,
|
274 |
+
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
|
275 |
+
pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
|
276 |
+
negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
|
277 |
+
ip_adapter_image: Optional[PipelineImageInput] = None,
|
278 |
+
output_type: Optional[str] = "pil",
|
279 |
+
return_dict: bool = True,
|
280 |
+
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
|
281 |
+
guidance_rescale: float = 0.0,
|
282 |
+
original_size: Optional[Tuple[int, int]] = None,
|
283 |
+
crops_coords_top_left: Tuple[int, int] = (0, 0),
|
284 |
+
target_size: Optional[Tuple[int, int]] = None,
|
285 |
+
negative_original_size: Optional[Tuple[int, int]] = None,
|
286 |
+
negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
|
287 |
+
negative_target_size: Optional[Tuple[int, int]] = None,
|
288 |
+
clip_skip: Optional[int] = None,
|
289 |
+
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
|
290 |
+
callback_steps: Optional[int] = 1,
|
291 |
+
callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
|
292 |
+
callback_on_step_end_tensor_inputs: List[str] = ["latents"],
|
293 |
+
### DiffuseHigh parameters ###
|
294 |
+
target_height: Union[int, List[int]] = [2048, 3072, 4096],
|
295 |
+
target_width: Union[int, List[int]] = [2048, 3072, 4096],
|
296 |
+
guidance_image: Optional[Union[torch.FloatTensor, PIL.Image.Image, np.ndarray]] = None,
|
297 |
+
noising_steps: int = 15,
|
298 |
+
diffusehigh_guidance_scale: float = 10.0,
|
299 |
+
# >>> DWT parameters
|
300 |
+
enable_dwt: bool = True,
|
301 |
+
dwt_level: Optional[int] = 1,
|
302 |
+
dwt_wave: Optional[str] = "db4",
|
303 |
+
dwt_mode: Optional[str] = "symmetric",
|
304 |
+
dwt_steps: Optional[int] = 5,
|
305 |
+
# >>> Sharpening parameters
|
306 |
+
enable_sharpening: bool = True,
|
307 |
+
sharpening_kernel_size: int = 3,
|
308 |
+
sharpening_sigma: Optional[Union[Tuple[float, float], float]] = (0.1, 2.0),
|
309 |
+
sharpening_alpha: float = 1.0,
|
310 |
+
**kwargs,
|
311 |
+
):
|
312 |
+
r"""
|
313 |
+
Function invoked when calling the pipeline for generation.
|
314 |
+
|
315 |
+
Args:
|
316 |
+
prompt (`str` or `List[str]`, *optional*):
|
317 |
+
The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
|
318 |
+
instead.
|
319 |
+
prompt_2 (`str` or `List[str]`, *optional*):
|
320 |
+
The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
|
321 |
+
used in both text-encoders
|
322 |
+
height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
|
323 |
+
The height in pixels of the generated image. This is set to 1024 by default for the best results.
|
324 |
+
Anything below 512 pixels won't work well for
|
325 |
+
[stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
|
326 |
+
and checkpoints that are not specifically fine-tuned on low resolutions.
|
327 |
+
width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
|
328 |
+
The width in pixels of the generated image. This is set to 1024 by default for the best results.
|
329 |
+
Anything below 512 pixels won't work well for
|
330 |
+
[stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
|
331 |
+
and checkpoints that are not specifically fine-tuned on low resolutions.
|
332 |
+
num_inference_steps (`int`, *optional*, defaults to 50):
|
333 |
+
The number of denoising steps. More denoising steps usually lead to a higher quality image at the
|
334 |
+
expense of slower inference.
|
335 |
+
timesteps (`List[int]`, *optional*):
|
336 |
+
Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
|
337 |
+
in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
|
338 |
+
passed will be used. Must be in descending order.
|
339 |
+
denoising_end (`float`, *optional*):
|
340 |
+
When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
|
341 |
+
completed before it is intentionally prematurely terminated. As a result, the returned sample will
|
342 |
+
still retain a substantial amount of noise as determined by the discrete timesteps selected by the
|
343 |
+
scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a
|
344 |
+
"Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
|
345 |
+
Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output)
|
346 |
+
guidance_scale (`float`, *optional*, defaults to 5.0):
|
347 |
+
Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
|
348 |
+
`guidance_scale` is defined as `w` of equation 2. of [Imagen
|
349 |
+
Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
|
350 |
+
1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
|
351 |
+
usually at the expense of lower image quality.
|
352 |
+
negative_prompt (`str` or `List[str]`, *optional*):
|
353 |
+
The prompt or prompts not to guide the image generation. If not defined, one has to pass
|
354 |
+
`negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
|
355 |
+
less than `1`).
|
356 |
+
negative_prompt_2 (`str` or `List[str]`, *optional*):
|
357 |
+
The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
|
358 |
+
`text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
|
359 |
+
num_images_per_prompt (`int`, *optional*, defaults to 1):
|
360 |
+
The number of images to generate per prompt.
|
361 |
+
eta (`float`, *optional*, defaults to 0.0):
|
362 |
+
Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
|
363 |
+
[`schedulers.DDIMScheduler`], will be ignored for others.
|
364 |
+
generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
|
365 |
+
One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
|
366 |
+
to make generation deterministic.
|
367 |
+
latents (`torch.FloatTensor`, *optional*):
|
368 |
+
Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
|
369 |
+
generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
|
370 |
+
tensor will ge generated by sampling using the supplied random `generator`.
|
371 |
+
prompt_embeds (`torch.FloatTensor`, *optional*):
|
372 |
+
Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
|
373 |
+
provided, text embeddings will be generated from `prompt` input argument.
|
374 |
+
negative_prompt_embeds (`torch.FloatTensor`, *optional*):
|
375 |
+
Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
|
376 |
+
weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
|
377 |
+
argument.
|
378 |
+
pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
|
379 |
+
Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
|
380 |
+
If not provided, pooled text embeddings will be generated from `prompt` input argument.
|
381 |
+
negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
|
382 |
+
Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
|
383 |
+
weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
|
384 |
+
input argument.
|
385 |
+
ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
|
386 |
+
output_type (`str`, *optional*, defaults to `"pil"`):
|
387 |
+
The output format of the generate image. Choose between
|
388 |
+
[PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
|
389 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
390 |
+
Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead
|
391 |
+
of a plain tuple.
|
392 |
+
cross_attention_kwargs (`dict`, *optional*):
|
393 |
+
A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
|
394 |
+
`self.processor` in
|
395 |
+
[diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
|
396 |
+
guidance_rescale (`float`, *optional*, defaults to 0.0):
|
397 |
+
Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
|
398 |
+
Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
|
399 |
+
[Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
|
400 |
+
Guidance rescale factor should fix overexposure when using zero terminal SNR.
|
401 |
+
original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
|
402 |
+
If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
|
403 |
+
`original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as
|
404 |
+
explained in section 2.2 of
|
405 |
+
[https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
|
406 |
+
crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
|
407 |
+
`crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
|
408 |
+
`crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
|
409 |
+
`crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
|
410 |
+
[https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
|
411 |
+
target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
|
412 |
+
For most cases, `target_size` should be set to the desired height and width of the generated image. If
|
413 |
+
not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in
|
414 |
+
section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
|
415 |
+
negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
|
416 |
+
To negatively condition the generation process based on a specific image resolution. Part of SDXL's
|
417 |
+
micro-conditioning as explained in section 2.2 of
|
418 |
+
[https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
|
419 |
+
information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
|
420 |
+
negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
|
421 |
+
To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's
|
422 |
+
micro-conditioning as explained in section 2.2 of
|
423 |
+
[https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
|
424 |
+
information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
|
425 |
+
negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
|
426 |
+
To negatively condition the generation process based on a target image resolution. It should be as same
|
427 |
+
as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of
|
428 |
+
[https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
|
429 |
+
information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
|
430 |
+
callback_on_step_end (`Callable`, *optional*):
|
431 |
+
A function that calls at the end of each denoising steps during the inference. The function is called
|
432 |
+
with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
|
433 |
+
callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
|
434 |
+
`callback_on_step_end_tensor_inputs`.
|
435 |
+
callback_on_step_end_tensor_inputs (`List`, *optional*):
|
436 |
+
The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
|
437 |
+
will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
|
438 |
+
`._callback_tensor_inputs` attribute of your pipeline class.
|
439 |
+
target_height ('List[int]' or int):
|
440 |
+
The height of the image being generated. If list is given, the pipeline generates corresponding intermediate
|
441 |
+
resolution images in a progressive manner.
|
442 |
+
target_width ('List[int]' or int):
|
443 |
+
The width of the image being generated. If list is given, the pipeline generates corresponding intermediate
|
444 |
+
resolution images in a progressive manner.
|
445 |
+
|
446 |
+
|
447 |
+
|
448 |
+
Examples:
|
449 |
+
|
450 |
+
Returns:
|
451 |
+
[`DiffuseHighSDXLPipelineOutput`] or `tuple`:
|
452 |
+
[`DiffuseHighSDXLPipelineOutput`] if `return_dict` is True, otherwise a
|
453 |
+
`tuple`. When returning a tuple, the first element is a list with the generated images.
|
454 |
+
"""
|
455 |
+
# 0. Default height and width to unet
|
456 |
+
height = self.default_sample_size * self.vae_scale_factor
|
457 |
+
width = self.default_sample_size * self.vae_scale_factor
|
458 |
+
|
459 |
+
original_size = original_size or (height, width)
|
460 |
+
target_size = target_size or (height, width)
|
461 |
+
|
462 |
+
# 1. Check inputs. Raise error if not correct
|
463 |
+
self.check_inputs(
|
464 |
+
prompt,
|
465 |
+
prompt_2,
|
466 |
+
height,
|
467 |
+
width,
|
468 |
+
callback_steps,
|
469 |
+
negative_prompt,
|
470 |
+
negative_prompt_2,
|
471 |
+
prompt_embeds,
|
472 |
+
negative_prompt_embeds,
|
473 |
+
pooled_prompt_embeds,
|
474 |
+
negative_pooled_prompt_embeds,
|
475 |
+
callback_on_step_end_tensor_inputs,
|
476 |
+
)
|
477 |
+
|
478 |
+
self._guidance_scale = guidance_scale
|
479 |
+
self._guidance_rescale = guidance_rescale
|
480 |
+
self._clip_skip = clip_skip
|
481 |
+
self._cross_attention_kwargs = cross_attention_kwargs
|
482 |
+
self._denoising_end = denoising_end
|
483 |
+
|
484 |
+
# 2. Define call parameters
|
485 |
+
if prompt is not None and isinstance(prompt, str):
|
486 |
+
batch_size = 1
|
487 |
+
elif prompt is not None and isinstance(prompt, list):
|
488 |
+
batch_size = len(prompt)
|
489 |
+
else:
|
490 |
+
batch_size = prompt_embeds.shape[0]
|
491 |
+
|
492 |
+
device = self._execution_device
|
493 |
+
|
494 |
+
# 3. Encode input prompt
|
495 |
+
lora_scale = (
|
496 |
+
self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
|
497 |
+
)
|
498 |
+
|
499 |
+
(
|
500 |
+
prompt_embeds,
|
501 |
+
negative_prompt_embeds,
|
502 |
+
pooled_prompt_embeds,
|
503 |
+
negative_pooled_prompt_embeds,
|
504 |
+
) = self.encode_prompt(
|
505 |
+
prompt=prompt,
|
506 |
+
prompt_2=prompt_2,
|
507 |
+
device=device,
|
508 |
+
num_images_per_prompt=num_images_per_prompt,
|
509 |
+
do_classifier_free_guidance=self.do_classifier_free_guidance,
|
510 |
+
negative_prompt=negative_prompt,
|
511 |
+
negative_prompt_2=negative_prompt_2,
|
512 |
+
prompt_embeds=prompt_embeds,
|
513 |
+
negative_prompt_embeds=negative_prompt_embeds,
|
514 |
+
pooled_prompt_embeds=pooled_prompt_embeds,
|
515 |
+
negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
|
516 |
+
lora_scale=lora_scale,
|
517 |
+
clip_skip=self.clip_skip,
|
518 |
+
)
|
519 |
+
|
520 |
+
# 4. Prepare timesteps
|
521 |
+
timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
|
522 |
+
|
523 |
+
# 5. Prepare latent variables
|
524 |
+
num_channels_latents = self.unet.config.in_channels
|
525 |
+
latents = self.prepare_latents(
|
526 |
+
batch_size * num_images_per_prompt,
|
527 |
+
num_channels_latents,
|
528 |
+
height,
|
529 |
+
width,
|
530 |
+
prompt_embeds.dtype,
|
531 |
+
device,
|
532 |
+
generator,
|
533 |
+
latents,
|
534 |
+
)
|
535 |
+
|
536 |
+
# 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
|
537 |
+
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
|
538 |
+
|
539 |
+
# 7. Prepare added time ids & embeddings
|
540 |
+
add_text_embeds = pooled_prompt_embeds
|
541 |
+
if self.text_encoder_2 is None:
|
542 |
+
text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
|
543 |
+
else:
|
544 |
+
text_encoder_projection_dim = self.text_encoder_2.config.projection_dim
|
545 |
+
|
546 |
+
add_time_ids = self._get_add_time_ids(
|
547 |
+
original_size,
|
548 |
+
crops_coords_top_left,
|
549 |
+
target_size,
|
550 |
+
dtype=prompt_embeds.dtype,
|
551 |
+
text_encoder_projection_dim=text_encoder_projection_dim,
|
552 |
+
)
|
553 |
+
if negative_original_size is not None and negative_target_size is not None:
|
554 |
+
negative_add_time_ids = self._get_add_time_ids(
|
555 |
+
negative_original_size,
|
556 |
+
negative_crops_coords_top_left,
|
557 |
+
negative_target_size,
|
558 |
+
dtype=prompt_embeds.dtype,
|
559 |
+
text_encoder_projection_dim=text_encoder_projection_dim,
|
560 |
+
)
|
561 |
+
else:
|
562 |
+
negative_add_time_ids = add_time_ids
|
563 |
+
|
564 |
+
if self.do_classifier_free_guidance:
|
565 |
+
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
|
566 |
+
add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
|
567 |
+
add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
|
568 |
+
|
569 |
+
prompt_embeds = prompt_embeds.to(device)
|
570 |
+
add_text_embeds = add_text_embeds.to(device)
|
571 |
+
add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
|
572 |
+
|
573 |
+
if ip_adapter_image is not None:
|
574 |
+
image_embeds, negative_image_embeds = self.encode_image(ip_adapter_image, device, num_images_per_prompt)
|
575 |
+
if self.do_classifier_free_guidance:
|
576 |
+
image_embeds = torch.cat([negative_image_embeds, image_embeds])
|
577 |
+
image_embeds = image_embeds.to(device)
|
578 |
+
|
579 |
+
# 8. Denoising loop
|
580 |
+
num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
|
581 |
+
|
582 |
+
# 8.1 Apply denoising_end
|
583 |
+
if (
|
584 |
+
self.denoising_end is not None
|
585 |
+
and isinstance(self.denoising_end, float)
|
586 |
+
and self.denoising_end > 0
|
587 |
+
and self.denoising_end < 1
|
588 |
+
):
|
589 |
+
discrete_timestep_cutoff = int(
|
590 |
+
round(
|
591 |
+
self.scheduler.config.num_train_timesteps
|
592 |
+
- (self.denoising_end * self.scheduler.config.num_train_timesteps)
|
593 |
+
)
|
594 |
+
)
|
595 |
+
num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
|
596 |
+
timesteps = timesteps[:num_inference_steps]
|
597 |
+
|
598 |
+
# 9. Optionally get Guidance Scale Embedding
|
599 |
+
timestep_cond = None
|
600 |
+
if self.unet.config.time_cond_proj_dim is not None:
|
601 |
+
guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
|
602 |
+
timestep_cond = self.get_guidance_scale_embedding(
|
603 |
+
guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
|
604 |
+
).to(device=device, dtype=latents.dtype)
|
605 |
+
|
606 |
+
# 10. Obtain clean image for structral guidance (can be given by user or generated)
|
607 |
+
if guidance_image is None:
|
608 |
+
self._num_timesteps = len(timesteps)
|
609 |
+
with self.progress_bar(total=num_inference_steps) as progress_bar:
|
610 |
+
for i, t in enumerate(timesteps):
|
611 |
+
# expand the latents if we are doing classifier free guidance
|
612 |
+
latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
|
613 |
+
|
614 |
+
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
|
615 |
+
|
616 |
+
# predict the noise residual
|
617 |
+
added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
|
618 |
+
if ip_adapter_image is not None:
|
619 |
+
added_cond_kwargs["image_embeds"] = image_embeds
|
620 |
+
noise_pred = self.unet(
|
621 |
+
latent_model_input,
|
622 |
+
t,
|
623 |
+
encoder_hidden_states=prompt_embeds,
|
624 |
+
timestep_cond=timestep_cond,
|
625 |
+
cross_attention_kwargs=self.cross_attention_kwargs,
|
626 |
+
added_cond_kwargs=added_cond_kwargs,
|
627 |
+
return_dict=False,
|
628 |
+
)[0]
|
629 |
+
|
630 |
+
# perform guidance
|
631 |
+
if self.do_classifier_free_guidance:
|
632 |
+
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
633 |
+
noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
|
634 |
+
|
635 |
+
if self.do_classifier_free_guidance and self.guidance_rescale > 0.0:
|
636 |
+
# Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
|
637 |
+
noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale)
|
638 |
+
|
639 |
+
# compute the previous noisy sample x_t -> x_t-1
|
640 |
+
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
|
641 |
+
|
642 |
+
if callback_on_step_end is not None:
|
643 |
+
callback_kwargs = {}
|
644 |
+
for k in callback_on_step_end_tensor_inputs:
|
645 |
+
callback_kwargs[k] = locals()[k]
|
646 |
+
callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
|
647 |
+
|
648 |
+
latents = callback_outputs.pop("latents", latents)
|
649 |
+
prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
|
650 |
+
negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
|
651 |
+
add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds)
|
652 |
+
negative_pooled_prompt_embeds = callback_outputs.pop(
|
653 |
+
"negative_pooled_prompt_embeds", negative_pooled_prompt_embeds
|
654 |
+
)
|
655 |
+
add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids)
|
656 |
+
negative_add_time_ids = callback_outputs.pop("negative_add_time_ids", negative_add_time_ids)
|
657 |
+
|
658 |
+
# call the callback, if provided
|
659 |
+
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
|
660 |
+
progress_bar.update()
|
661 |
+
if callback is not None and i % callback_steps == 0:
|
662 |
+
step_idx = i // getattr(self.scheduler, "order", 1)
|
663 |
+
callback(step_idx, t, latents)
|
664 |
+
|
665 |
+
image = self._decode_vae_latent(latents, output_type='pt')
|
666 |
+
else:
|
667 |
+
image = self.image_processor.preprocess(guidance_image, height, width)
|
668 |
+
if self.image_processor.config.do_normalize:
|
669 |
+
image = (image + 1.) * 0.5
|
670 |
+
|
671 |
+
image = image.to(self.device)
|
672 |
+
|
673 |
+
original_guidance_image = image
|
674 |
+
|
675 |
+
# |-------------------------------- DiffuseHigh process --------------------------------|
|
676 |
+
# DWT & inverse DWT works on torch.float32
|
677 |
+
if enable_dwt:
|
678 |
+
self.DWT = DWTForward(J=dwt_level, wave=dwt_wave, mode=dwt_mode).to(self.device)
|
679 |
+
self.iDWT = DWTInverse(wave=dwt_wave, mode=dwt_mode).to(self.device)
|
680 |
+
|
681 |
+
# 11. Prepare progressive DiffuseHigh pipeline
|
682 |
+
self.scheduler.set_timesteps(num_inference_steps)
|
683 |
+
diffusehigh_timesteps = self.scheduler.timesteps[-noising_steps:]
|
684 |
+
self.enable_vae_tiling() # Vae tiling mode in order to prevent OOM issues
|
685 |
+
|
686 |
+
if isinstance(target_width, int):
|
687 |
+
target_width = [target_width]
|
688 |
+
if isinstance(target_height, int):
|
689 |
+
target_height = [target_height]
|
690 |
+
|
691 |
+
assert len(target_width) == len(target_height)
|
692 |
+
|
693 |
+
#12. Progressive DiffuseHigh Pipeline
|
694 |
+
for h, w in zip(target_height, target_width):
|
695 |
+
# interpolate the image to the desired resolution
|
696 |
+
guidance_image = F.interpolate(image, (h, w), mode="bicubic", align_corners=False)
|
697 |
+
|
698 |
+
# apply sharpening operation to the image
|
699 |
+
if enable_sharpening:
|
700 |
+
guidance_image = gaussian_blur_image_sharpening(
|
701 |
+
guidance_image,
|
702 |
+
kernel_size=sharpening_kernel_size,
|
703 |
+
sigma=sharpening_sigma,
|
704 |
+
alpha=sharpening_alpha,
|
705 |
+
)
|
706 |
+
|
707 |
+
# extract low-frequency component (structural guidance) from the guidance image
|
708 |
+
if enable_dwt:
|
709 |
+
LL, _ = self.DWT(guidance_image)
|
710 |
+
|
711 |
+
# obtain latent of the interpolated image and noise it
|
712 |
+
latents = self._encode_vae_image(guidance_image)
|
713 |
+
noise = randn_tensor(latents.shape, generator, device=latents.device, dtype=latents.dtype)
|
714 |
+
latents = self.scheduler.add_noise(latents, noise, diffusehigh_timesteps[None, 0])
|
715 |
+
|
716 |
+
for i, t in tqdm(enumerate(diffusehigh_timesteps), total=diffusehigh_timesteps.shape[0]):
|
717 |
+
# expand the latents if we are doing classifier free guidance
|
718 |
+
latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
|
719 |
+
|
720 |
+
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
|
721 |
+
|
722 |
+
# predict the noise residual
|
723 |
+
added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
|
724 |
+
|
725 |
+
noise_pred = self.unet(
|
726 |
+
latent_model_input,
|
727 |
+
t,
|
728 |
+
encoder_hidden_states=prompt_embeds,
|
729 |
+
timestep_cond=timestep_cond,
|
730 |
+
cross_attention_kwargs=self.cross_attention_kwargs, # None
|
731 |
+
added_cond_kwargs=added_cond_kwargs, # None
|
732 |
+
return_dict=False,
|
733 |
+
)[0]
|
734 |
+
|
735 |
+
# perform guidance
|
736 |
+
if self.do_classifier_free_guidance:
|
737 |
+
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
738 |
+
noise_pred = noise_pred_uncond + diffusehigh_guidance_scale * (noise_pred_text - noise_pred_uncond)
|
739 |
+
|
740 |
+
# EDM sampler step
|
741 |
+
latents = self.edm_scheduler_step(
|
742 |
+
noise_pred,
|
743 |
+
t,
|
744 |
+
latents,
|
745 |
+
**extra_step_kwargs,
|
746 |
+
LL_guidance=LL if (enable_dwt and i < dwt_steps) else None,
|
747 |
+
)[0]
|
748 |
+
|
749 |
+
image = self._decode_vae_latent(latents)
|
750 |
+
|
751 |
+
if isinstance(self.scheduler, EulerDiscreteScheduler):
|
752 |
+
self.scheduler._step_index = None
|
753 |
+
|
754 |
+
# Offload all models
|
755 |
+
self.maybe_free_model_hooks()
|
756 |
+
|
757 |
+
if output_type != 'pt':
|
758 |
+
image = self.image_processor.postprocess(image * 2 - 1, output_type=output_type)
|
759 |
+
guidance_image = self.image_processor.postprocess(original_guidance_image * 2 -1 , output_type=output_type)
|
760 |
+
|
761 |
+
if not return_dict:
|
762 |
+
return (image, guidance_image)
|
763 |
+
|
764 |
+
return DiffuseHighSDXLPipelineOutput(images=image, guidance_image=guidance_image)
|
765 |
+
|
766 |
+
|
767 |
+
def set_seeds(seed):
|
768 |
+
os.environ["PYTHONHASHSEED"] = str(seed)
|
769 |
+
np.random.seed(seed)
|
770 |
+
torch.manual_seed(seed)
|
771 |
+
torch.cuda.manual_seed(seed)
|
772 |
+
torch.backends.cudnn.deterministic = True
|
773 |
+
torch.backends.cudnn.benchmark = True
|
774 |
+
|
775 |
+
# DEBUGGING
|
776 |
+
if __name__ == "__main__":
|
777 |
+
set_seeds(23)
|
778 |
+
|
779 |
+
model = DiffuseHighSDXLPipeline.from_pretrained(
|
780 |
+
"stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16, #scheduler=scheduler
|
781 |
+
).to("cuda")
|
782 |
+
|
783 |
+
prompt = "Cinematic photo of delicious chocolate icecream."
|
784 |
+
|
785 |
+
negative_prompt = "blurry, ugly, duplicate, poorly drawn, deformed, mosaic"
|
786 |
+
|
787 |
+
image = model(
|
788 |
+
prompt,
|
789 |
+
negative_prompt=negative_prompt,
|
790 |
+
target_height=[2048, 3072, 4096],
|
791 |
+
target_width=[2048, 3072, 4096],
|
792 |
+
enable_dwt=True,
|
793 |
+
dwt_steps=5,
|
794 |
+
enable_sharpening=True,
|
795 |
+
sharpness_factor=1.0,
|
796 |
+
).images[0]
|
797 |
+
|
798 |
+
image.save("sample.png")
|
requirements.txt
ADDED
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
diffusers==0.24.0
|
2 |
+
accelerate
|
3 |
+
transformers
|
4 |
+
pywavelets
|
5 |
+
pytorch-wavelets
|
utils/utils.py
ADDED
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import torch
|
3 |
+
import numpy as np
|
4 |
+
|
5 |
+
def set_seeds(seed):
|
6 |
+
os.environ["PYTHONHASHSEED"] = str(seed)
|
7 |
+
np.random.seed(seed)
|
8 |
+
torch.manual_seed(seed)
|
9 |
+
torch.cuda.manual_seed(seed)
|
10 |
+
torch.backends.cudnn.deterministic = True
|
11 |
+
torch.backends.cudnn.benchmark = True
|