Spaces:
Paused
Paused
Create pipeline_calls.py
Browse files- pipeline_calls.py +552 -0
pipeline_calls.py
ADDED
@@ -0,0 +1,552 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright 2023 Google LLC
|
2 |
+
#
|
3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
4 |
+
# you may not use this file except in compliance with the License.
|
5 |
+
# You may obtain a copy of the License at
|
6 |
+
#
|
7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
8 |
+
#
|
9 |
+
# Unless required by applicable law or agreed to in writing, software
|
10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12 |
+
# See the License for the specific language governing permissions and
|
13 |
+
# limitations under the License.
|
14 |
+
|
15 |
+
|
16 |
+
from __future__ import annotations
|
17 |
+
from typing import Any
|
18 |
+
import torch
|
19 |
+
import numpy as np
|
20 |
+
from diffusers import ControlNetModel, StableDiffusionXLControlNetPipeline
|
21 |
+
from diffusers.image_processor import PipelineImageInput
|
22 |
+
from diffusers.utils.torch_utils import is_compiled_module, is_torch_version
|
23 |
+
from transformers import DPTImageProcessor, DPTForDepthEstimation
|
24 |
+
from diffusers import StableDiffusionPanoramaPipeline
|
25 |
+
from PIL import Image
|
26 |
+
import copy
|
27 |
+
|
28 |
+
T = torch.Tensor
|
29 |
+
TN = T | None
|
30 |
+
|
31 |
+
|
32 |
+
def get_depth_map(image: Image, feature_processor: DPTImageProcessor, depth_estimator: DPTForDepthEstimation) -> Image:
|
33 |
+
image = feature_processor(images=image, return_tensors="pt").pixel_values.to("cuda")
|
34 |
+
with torch.no_grad(), torch.autocast("cuda"):
|
35 |
+
depth_map = depth_estimator(image).predicted_depth
|
36 |
+
|
37 |
+
depth_map = torch.nn.functional.interpolate(
|
38 |
+
depth_map.unsqueeze(1),
|
39 |
+
size=(1024, 1024),
|
40 |
+
mode="bicubic",
|
41 |
+
align_corners=False,
|
42 |
+
)
|
43 |
+
depth_min = torch.amin(depth_map, dim=[1, 2, 3], keepdim=True)
|
44 |
+
depth_max = torch.amax(depth_map, dim=[1, 2, 3], keepdim=True)
|
45 |
+
depth_map = (depth_map - depth_min) / (depth_max - depth_min)
|
46 |
+
image = torch.cat([depth_map] * 3, dim=1)
|
47 |
+
|
48 |
+
image = image.permute(0, 2, 3, 1).cpu().numpy()[0]
|
49 |
+
image = Image.fromarray((image * 255.0).clip(0, 255).astype(np.uint8))
|
50 |
+
return image
|
51 |
+
|
52 |
+
|
53 |
+
def concat_zero_control(control_reisduel: T) -> T:
|
54 |
+
b = control_reisduel.shape[0] // 2
|
55 |
+
zerso_reisduel = torch.zeros_like(control_reisduel[0:1])
|
56 |
+
return torch.cat((zerso_reisduel, control_reisduel[:b], zerso_reisduel, control_reisduel[b::]))
|
57 |
+
|
58 |
+
|
59 |
+
@torch.no_grad()
|
60 |
+
def controlnet_call(
|
61 |
+
pipeline: StableDiffusionXLControlNetPipeline,
|
62 |
+
prompt: str | list[str] = None,
|
63 |
+
prompt_2: str | list[str] | None = None,
|
64 |
+
image: PipelineImageInput = None,
|
65 |
+
height: int | None = None,
|
66 |
+
width: int | None = None,
|
67 |
+
num_inference_steps: int = 50,
|
68 |
+
guidance_scale: float = 5.0,
|
69 |
+
negative_prompt: str | list[str] | None = None,
|
70 |
+
negative_prompt_2: str | list[str] | None = None,
|
71 |
+
num_images_per_prompt: int = 1,
|
72 |
+
eta: float = 0.0,
|
73 |
+
generator: torch.Generator | None = None,
|
74 |
+
latents: TN = None,
|
75 |
+
prompt_embeds: TN = None,
|
76 |
+
negative_prompt_embeds: TN = None,
|
77 |
+
pooled_prompt_embeds: TN = None,
|
78 |
+
negative_pooled_prompt_embeds: TN = None,
|
79 |
+
cross_attention_kwargs: dict[str, Any] | None = None,
|
80 |
+
controlnet_conditioning_scale: float | list[float] = 1.0,
|
81 |
+
control_guidance_start: float | list[float] = 0.0,
|
82 |
+
control_guidance_end: float | list[float] = 1.0,
|
83 |
+
original_size: tuple[int, int] = None,
|
84 |
+
crops_coords_top_left: tuple[int, int] = (0, 0),
|
85 |
+
target_size: tuple[int, int] | None = None,
|
86 |
+
negative_original_size: tuple[int, int] | None = None,
|
87 |
+
negative_crops_coords_top_left: tuple[int, int] = (0, 0),
|
88 |
+
negative_target_size:tuple[int, int] | None = None,
|
89 |
+
clip_skip: int | None = None,
|
90 |
+
) -> list[Image]:
|
91 |
+
controlnet = pipeline.controlnet._orig_mod if is_compiled_module(pipeline.controlnet) else pipeline.controlnet
|
92 |
+
|
93 |
+
# align format for control guidance
|
94 |
+
if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list):
|
95 |
+
control_guidance_start = len(control_guidance_end) * [control_guidance_start]
|
96 |
+
elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list):
|
97 |
+
control_guidance_end = len(control_guidance_start) * [control_guidance_end]
|
98 |
+
elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list):
|
99 |
+
mult = 1
|
100 |
+
control_guidance_start, control_guidance_end = (
|
101 |
+
mult * [control_guidance_start],
|
102 |
+
mult * [control_guidance_end],
|
103 |
+
)
|
104 |
+
|
105 |
+
# 1. Check inputs. Raise error if not correct
|
106 |
+
pipeline.check_inputs(
|
107 |
+
prompt,
|
108 |
+
prompt_2,
|
109 |
+
image,
|
110 |
+
1,
|
111 |
+
negative_prompt,
|
112 |
+
negative_prompt_2,
|
113 |
+
prompt_embeds,
|
114 |
+
negative_prompt_embeds,
|
115 |
+
pooled_prompt_embeds,
|
116 |
+
negative_pooled_prompt_embeds,
|
117 |
+
controlnet_conditioning_scale,
|
118 |
+
control_guidance_start,
|
119 |
+
control_guidance_end,
|
120 |
+
)
|
121 |
+
|
122 |
+
pipeline._guidance_scale = guidance_scale
|
123 |
+
|
124 |
+
# 2. Define call parameters
|
125 |
+
if prompt is not None and isinstance(prompt, str):
|
126 |
+
batch_size = 1
|
127 |
+
elif prompt is not None and isinstance(prompt, list):
|
128 |
+
batch_size = len(prompt)
|
129 |
+
else:
|
130 |
+
batch_size = prompt_embeds.shape[0]
|
131 |
+
|
132 |
+
device = pipeline._execution_device
|
133 |
+
|
134 |
+
# 3. Encode input prompt
|
135 |
+
text_encoder_lora_scale = (
|
136 |
+
cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
|
137 |
+
)
|
138 |
+
(
|
139 |
+
prompt_embeds,
|
140 |
+
negative_prompt_embeds,
|
141 |
+
pooled_prompt_embeds,
|
142 |
+
negative_pooled_prompt_embeds,
|
143 |
+
) = pipeline.encode_prompt(
|
144 |
+
prompt,
|
145 |
+
prompt_2,
|
146 |
+
device,
|
147 |
+
1,
|
148 |
+
True,
|
149 |
+
negative_prompt,
|
150 |
+
negative_prompt_2,
|
151 |
+
prompt_embeds=prompt_embeds,
|
152 |
+
negative_prompt_embeds=negative_prompt_embeds,
|
153 |
+
pooled_prompt_embeds=pooled_prompt_embeds,
|
154 |
+
negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
|
155 |
+
lora_scale=text_encoder_lora_scale,
|
156 |
+
clip_skip=clip_skip,
|
157 |
+
)
|
158 |
+
|
159 |
+
# 4. Prepare image
|
160 |
+
if isinstance(controlnet, ControlNetModel):
|
161 |
+
image = pipeline.prepare_image(
|
162 |
+
image=image,
|
163 |
+
width=width,
|
164 |
+
height=height,
|
165 |
+
batch_size=1,
|
166 |
+
num_images_per_prompt=1,
|
167 |
+
device=device,
|
168 |
+
dtype=controlnet.dtype,
|
169 |
+
do_classifier_free_guidance=True,
|
170 |
+
guess_mode=False,
|
171 |
+
)
|
172 |
+
height, width = image.shape[-2:]
|
173 |
+
image = torch.stack([image[0]] * num_images_per_prompt + [image[1]] * num_images_per_prompt)
|
174 |
+
else:
|
175 |
+
assert False
|
176 |
+
# 5. Prepare timesteps
|
177 |
+
pipeline.scheduler.set_timesteps(num_inference_steps, device=device)
|
178 |
+
timesteps = pipeline.scheduler.timesteps
|
179 |
+
|
180 |
+
# 6. Prepare latent variables
|
181 |
+
num_channels_latents = pipeline.unet.config.in_channels
|
182 |
+
latents = pipeline.prepare_latents(
|
183 |
+
1 + num_images_per_prompt,
|
184 |
+
num_channels_latents,
|
185 |
+
height,
|
186 |
+
width,
|
187 |
+
prompt_embeds.dtype,
|
188 |
+
device,
|
189 |
+
generator,
|
190 |
+
latents,
|
191 |
+
)
|
192 |
+
|
193 |
+
# 6.5 Optionally get Guidance Scale Embedding
|
194 |
+
timestep_cond = None
|
195 |
+
|
196 |
+
# 7. Prepare extra step kwargs.
|
197 |
+
extra_step_kwargs = pipeline.prepare_extra_step_kwargs(generator, eta)
|
198 |
+
|
199 |
+
# 7.1 Create tensor stating which controlnets to keep
|
200 |
+
controlnet_keep = []
|
201 |
+
for i in range(len(timesteps)):
|
202 |
+
keeps = [
|
203 |
+
1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e)
|
204 |
+
for s, e in zip(control_guidance_start, control_guidance_end)
|
205 |
+
]
|
206 |
+
controlnet_keep.append(keeps[0] if isinstance(controlnet, ControlNetModel) else keeps)
|
207 |
+
|
208 |
+
# 7.2 Prepare added time ids & embeddings
|
209 |
+
if isinstance(image, list):
|
210 |
+
original_size = original_size or image[0].shape[-2:]
|
211 |
+
else:
|
212 |
+
original_size = original_size or image.shape[-2:]
|
213 |
+
target_size = target_size or (height, width)
|
214 |
+
|
215 |
+
add_text_embeds = pooled_prompt_embeds
|
216 |
+
if pipeline.text_encoder_2 is None:
|
217 |
+
text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
|
218 |
+
else:
|
219 |
+
text_encoder_projection_dim = pipeline.text_encoder_2.config.projection_dim
|
220 |
+
|
221 |
+
add_time_ids = pipeline._get_add_time_ids(
|
222 |
+
original_size,
|
223 |
+
crops_coords_top_left,
|
224 |
+
target_size,
|
225 |
+
dtype=prompt_embeds.dtype,
|
226 |
+
text_encoder_projection_dim=text_encoder_projection_dim,
|
227 |
+
)
|
228 |
+
|
229 |
+
if negative_original_size is not None and negative_target_size is not None:
|
230 |
+
negative_add_time_ids = pipeline._get_add_time_ids(
|
231 |
+
negative_original_size,
|
232 |
+
negative_crops_coords_top_left,
|
233 |
+
negative_target_size,
|
234 |
+
dtype=prompt_embeds.dtype,
|
235 |
+
text_encoder_projection_dim=text_encoder_projection_dim,
|
236 |
+
)
|
237 |
+
else:
|
238 |
+
negative_add_time_ids = add_time_ids
|
239 |
+
|
240 |
+
prompt_embeds = torch.stack([prompt_embeds[0]] + [prompt_embeds[1]] * num_images_per_prompt)
|
241 |
+
negative_prompt_embeds = torch.stack([negative_prompt_embeds[0]] + [negative_prompt_embeds[1]] * num_images_per_prompt)
|
242 |
+
negative_pooled_prompt_embeds = torch.stack([negative_pooled_prompt_embeds[0]] + [negative_pooled_prompt_embeds[1]] * num_images_per_prompt)
|
243 |
+
add_text_embeds = torch.stack([add_text_embeds[0]] + [add_text_embeds[1]] * num_images_per_prompt)
|
244 |
+
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
|
245 |
+
add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
|
246 |
+
add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
|
247 |
+
|
248 |
+
prompt_embeds = prompt_embeds.to(device)
|
249 |
+
add_text_embeds = add_text_embeds.to(device)
|
250 |
+
add_time_ids = add_time_ids.to(device).repeat(1 + num_images_per_prompt, 1)
|
251 |
+
batch_size = num_images_per_prompt + 1
|
252 |
+
# 8. Denoising loop
|
253 |
+
num_warmup_steps = len(timesteps) - num_inference_steps * pipeline.scheduler.order
|
254 |
+
is_unet_compiled = is_compiled_module(pipeline.unet)
|
255 |
+
is_controlnet_compiled = is_compiled_module(pipeline.controlnet)
|
256 |
+
is_torch_higher_equal_2_1 = is_torch_version(">=", "2.1")
|
257 |
+
added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
|
258 |
+
controlnet_prompt_embeds = torch.cat((prompt_embeds[1:batch_size], prompt_embeds[1:batch_size]))
|
259 |
+
controlnet_added_cond_kwargs = {key: torch.cat((item[1:batch_size,], item[1:batch_size])) for key, item in added_cond_kwargs.items()}
|
260 |
+
with pipeline.progress_bar(total=num_inference_steps) as progress_bar:
|
261 |
+
for i, t in enumerate(timesteps):
|
262 |
+
# Relevant thread:
|
263 |
+
# https://dev-discuss.pytorch.org/t/cudagraphs-in-pytorch-2-0/1428
|
264 |
+
if (is_unet_compiled and is_controlnet_compiled) and is_torch_higher_equal_2_1:
|
265 |
+
torch._inductor.cudagraph_mark_step_begin()
|
266 |
+
# expand the latents if we are doing classifier free guidance
|
267 |
+
latent_model_input = torch.cat([latents] * 2)
|
268 |
+
latent_model_input = pipeline.scheduler.scale_model_input(latent_model_input, t)
|
269 |
+
|
270 |
+
# controlnet(s) inference
|
271 |
+
control_model_input = torch.cat((latent_model_input[1:batch_size,], latent_model_input[batch_size+1:]))
|
272 |
+
|
273 |
+
if isinstance(controlnet_keep[i], list):
|
274 |
+
cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])]
|
275 |
+
else:
|
276 |
+
controlnet_cond_scale = controlnet_conditioning_scale
|
277 |
+
if isinstance(controlnet_cond_scale, list):
|
278 |
+
controlnet_cond_scale = controlnet_cond_scale[0]
|
279 |
+
cond_scale = controlnet_cond_scale * controlnet_keep[i]
|
280 |
+
if cond_scale > 0:
|
281 |
+
down_block_res_samples, mid_block_res_sample = pipeline.controlnet(
|
282 |
+
control_model_input,
|
283 |
+
t,
|
284 |
+
encoder_hidden_states=controlnet_prompt_embeds,
|
285 |
+
controlnet_cond=image,
|
286 |
+
conditioning_scale=cond_scale,
|
287 |
+
guess_mode=False,
|
288 |
+
added_cond_kwargs=controlnet_added_cond_kwargs,
|
289 |
+
return_dict=False,
|
290 |
+
)
|
291 |
+
|
292 |
+
mid_block_res_sample = concat_zero_control(mid_block_res_sample)
|
293 |
+
down_block_res_samples = [concat_zero_control(down_block_res_sample) for down_block_res_sample in down_block_res_samples]
|
294 |
+
else:
|
295 |
+
mid_block_res_sample = down_block_res_samples = None
|
296 |
+
# predict the noise residual
|
297 |
+
noise_pred = pipeline.unet(
|
298 |
+
latent_model_input,
|
299 |
+
t,
|
300 |
+
encoder_hidden_states=prompt_embeds,
|
301 |
+
timestep_cond=timestep_cond,
|
302 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
303 |
+
down_block_additional_residuals=down_block_res_samples,
|
304 |
+
mid_block_additional_residual=mid_block_res_sample,
|
305 |
+
added_cond_kwargs=added_cond_kwargs,
|
306 |
+
return_dict=False,
|
307 |
+
)[0]
|
308 |
+
|
309 |
+
# perform guidance
|
310 |
+
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
311 |
+
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
|
312 |
+
|
313 |
+
# compute the previous noisy sample x_t -> x_t-1
|
314 |
+
latents = pipeline.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
|
315 |
+
|
316 |
+
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % pipeline.scheduler.order == 0):
|
317 |
+
progress_bar.update()
|
318 |
+
|
319 |
+
# manually for max memory savings
|
320 |
+
if pipeline.vae.dtype == torch.float16 and pipeline.vae.config.force_upcast:
|
321 |
+
pipeline.upcast_vae()
|
322 |
+
latents = latents.to(next(iter(pipeline.vae.post_quant_conv.parameters())).dtype)
|
323 |
+
|
324 |
+
# make sure the VAE is in float32 mode, as it overflows in float16
|
325 |
+
needs_upcasting = pipeline.vae.dtype == torch.float16 and pipeline.vae.config.force_upcast
|
326 |
+
|
327 |
+
if needs_upcasting:
|
328 |
+
pipeline.upcast_vae()
|
329 |
+
latents = latents.to(next(iter(pipeline.vae.post_quant_conv.parameters())).dtype)
|
330 |
+
|
331 |
+
image = pipeline.vae.decode(latents / pipeline.vae.config.scaling_factor, return_dict=False)[0]
|
332 |
+
|
333 |
+
# cast back to fp16 if needed
|
334 |
+
if needs_upcasting:
|
335 |
+
pipeline.vae.to(dtype=torch.float16)
|
336 |
+
|
337 |
+
if pipeline.watermark is not None:
|
338 |
+
image = pipeline.watermark.apply_watermark(image)
|
339 |
+
|
340 |
+
image = pipeline.image_processor.postprocess(image, output_type='pil')
|
341 |
+
|
342 |
+
# Offload all models
|
343 |
+
pipeline.maybe_free_model_hooks()
|
344 |
+
return image
|
345 |
+
|
346 |
+
|
347 |
+
@torch.no_grad()
|
348 |
+
def panorama_call(
|
349 |
+
pipeline: StableDiffusionPanoramaPipeline,
|
350 |
+
prompt: list[str],
|
351 |
+
height: int | None = 512,
|
352 |
+
width: int | None = 2048,
|
353 |
+
num_inference_steps: int = 50,
|
354 |
+
guidance_scale: float = 7.5,
|
355 |
+
view_batch_size: int = 1,
|
356 |
+
negative_prompt: str | list[str] | None = None,
|
357 |
+
num_images_per_prompt: int | None = 1,
|
358 |
+
eta: float = 0.0,
|
359 |
+
generator: torch.Generator | None = None,
|
360 |
+
reference_latent: TN = None,
|
361 |
+
latents: TN = None,
|
362 |
+
prompt_embeds: TN = None,
|
363 |
+
negative_prompt_embeds: TN = None,
|
364 |
+
cross_attention_kwargs: dict[str, Any] | None = None,
|
365 |
+
circular_padding: bool = False,
|
366 |
+
clip_skip: int | None = None,
|
367 |
+
stride=8
|
368 |
+
) -> list[Image]:
|
369 |
+
# 0. Default height and width to unet
|
370 |
+
height = height or pipeline.unet.config.sample_size * pipeline.vae_scale_factor
|
371 |
+
width = width or pipeline.unet.config.sample_size * pipeline.vae_scale_factor
|
372 |
+
|
373 |
+
# 1. Check inputs. Raise error if not correct
|
374 |
+
pipeline.check_inputs(
|
375 |
+
prompt, height, width, 1, negative_prompt, prompt_embeds, negative_prompt_embeds
|
376 |
+
)
|
377 |
+
|
378 |
+
device = pipeline._execution_device
|
379 |
+
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
|
380 |
+
# of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
|
381 |
+
# corresponds to doing no classifier free guidance.
|
382 |
+
do_classifier_free_guidance = guidance_scale > 1.0
|
383 |
+
|
384 |
+
# 3. Encode input prompt
|
385 |
+
text_encoder_lora_scale = (
|
386 |
+
cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
|
387 |
+
)
|
388 |
+
prompt_embeds, negative_prompt_embeds = pipeline.encode_prompt(
|
389 |
+
prompt,
|
390 |
+
device,
|
391 |
+
num_images_per_prompt,
|
392 |
+
do_classifier_free_guidance,
|
393 |
+
negative_prompt,
|
394 |
+
prompt_embeds=prompt_embeds,
|
395 |
+
negative_prompt_embeds=negative_prompt_embeds,
|
396 |
+
lora_scale=text_encoder_lora_scale,
|
397 |
+
clip_skip=clip_skip,
|
398 |
+
)
|
399 |
+
# For classifier free guidance, we need to do two forward passes.
|
400 |
+
# Here we concatenate the unconditional and text embeddings into a single batch
|
401 |
+
# to avoid doing two forward passes
|
402 |
+
|
403 |
+
# 4. Prepare timesteps
|
404 |
+
pipeline.scheduler.set_timesteps(num_inference_steps, device=device)
|
405 |
+
timesteps = pipeline.scheduler.timesteps
|
406 |
+
|
407 |
+
# 5. Prepare latent variables
|
408 |
+
num_channels_latents = pipeline.unet.config.in_channels
|
409 |
+
latents = pipeline.prepare_latents(
|
410 |
+
1,
|
411 |
+
num_channels_latents,
|
412 |
+
height,
|
413 |
+
width,
|
414 |
+
prompt_embeds.dtype,
|
415 |
+
device,
|
416 |
+
generator,
|
417 |
+
latents,
|
418 |
+
)
|
419 |
+
if reference_latent is None:
|
420 |
+
reference_latent = torch.randn(1, 4, pipeline.unet.config.sample_size, pipeline.unet.config.sample_size,
|
421 |
+
generator=generator)
|
422 |
+
reference_latent = reference_latent.to(device=device, dtype=pipeline.unet.dtype)
|
423 |
+
# 6. Define panorama grid and initialize views for synthesis.
|
424 |
+
# prepare batch grid
|
425 |
+
views = pipeline.get_views(height, width, circular_padding=circular_padding, stride=stride)
|
426 |
+
views_batch = [views[i: i + view_batch_size] for i in range(0, len(views), view_batch_size)]
|
427 |
+
views_scheduler_status = [copy.deepcopy(pipeline.scheduler.__dict__)] * len(views_batch)
|
428 |
+
count = torch.zeros_like(latents)
|
429 |
+
value = torch.zeros_like(latents)
|
430 |
+
# 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
|
431 |
+
extra_step_kwargs = pipeline.prepare_extra_step_kwargs(generator, eta)
|
432 |
+
|
433 |
+
# 8. Denoising loop
|
434 |
+
# Each denoising step also includes refinement of the latents with respect to the
|
435 |
+
# views.
|
436 |
+
num_warmup_steps = len(timesteps) - num_inference_steps * pipeline.scheduler.order
|
437 |
+
|
438 |
+
negative_prompt_embeds = torch.cat([negative_prompt_embeds[:1],
|
439 |
+
*[negative_prompt_embeds[1:]] * view_batch_size]
|
440 |
+
)
|
441 |
+
prompt_embeds = torch.cat([prompt_embeds[:1],
|
442 |
+
*[prompt_embeds[1:]] * view_batch_size]
|
443 |
+
)
|
444 |
+
|
445 |
+
with pipeline.progress_bar(total=num_inference_steps) as progress_bar:
|
446 |
+
for i, t in enumerate(timesteps):
|
447 |
+
count.zero_()
|
448 |
+
value.zero_()
|
449 |
+
|
450 |
+
# generate views
|
451 |
+
# Here, we iterate through different spatial crops of the latents and denoise them. These
|
452 |
+
# denoised (latent) crops are then averaged to produce the final latent
|
453 |
+
# for the current timestep via MultiDiffusion. Please see Sec. 4.1 in the
|
454 |
+
# MultiDiffusion paper for more details: https://arxiv.org/abs/2302.08113
|
455 |
+
# Batch views denoise
|
456 |
+
for j, batch_view in enumerate(views_batch):
|
457 |
+
vb_size = len(batch_view)
|
458 |
+
# get the latents corresponding to the current view coordinates
|
459 |
+
if circular_padding:
|
460 |
+
latents_for_view = []
|
461 |
+
for h_start, h_end, w_start, w_end in batch_view:
|
462 |
+
if w_end > latents.shape[3]:
|
463 |
+
# Add circular horizontal padding
|
464 |
+
latent_view = torch.cat(
|
465 |
+
(
|
466 |
+
latents[:, :, h_start:h_end, w_start:],
|
467 |
+
latents[:, :, h_start:h_end, : w_end - latents.shape[3]],
|
468 |
+
),
|
469 |
+
dim=-1,
|
470 |
+
)
|
471 |
+
else:
|
472 |
+
latent_view = latents[:, :, h_start:h_end, w_start:w_end]
|
473 |
+
latents_for_view.append(latent_view)
|
474 |
+
latents_for_view = torch.cat(latents_for_view)
|
475 |
+
else:
|
476 |
+
latents_for_view = torch.cat(
|
477 |
+
[
|
478 |
+
latents[:, :, h_start:h_end, w_start:w_end]
|
479 |
+
for h_start, h_end, w_start, w_end in batch_view
|
480 |
+
]
|
481 |
+
)
|
482 |
+
# rematch block's scheduler status
|
483 |
+
pipeline.scheduler.__dict__.update(views_scheduler_status[j])
|
484 |
+
|
485 |
+
# expand the latents if we are doing classifier free guidance
|
486 |
+
latent_reference_plus_view = torch.cat((reference_latent, latents_for_view))
|
487 |
+
latent_model_input = latent_reference_plus_view.repeat(2, 1, 1, 1)
|
488 |
+
prompt_embeds_input = torch.cat([negative_prompt_embeds[: 1 + vb_size],
|
489 |
+
prompt_embeds[: 1 + vb_size]]
|
490 |
+
)
|
491 |
+
latent_model_input = pipeline.scheduler.scale_model_input(latent_model_input, t)
|
492 |
+
# predict the noise residual
|
493 |
+
# return
|
494 |
+
noise_pred = pipeline.unet(
|
495 |
+
latent_model_input,
|
496 |
+
t,
|
497 |
+
encoder_hidden_states=prompt_embeds_input,
|
498 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
499 |
+
).sample
|
500 |
+
|
501 |
+
# perform guidance
|
502 |
+
|
503 |
+
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
504 |
+
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
|
505 |
+
# compute the previous noisy sample x_t -> x_t-1
|
506 |
+
latent_reference_plus_view = pipeline.scheduler.step(
|
507 |
+
noise_pred, t, latent_reference_plus_view, **extra_step_kwargs
|
508 |
+
).prev_sample
|
509 |
+
if j == len(views_batch) - 1:
|
510 |
+
reference_latent = latent_reference_plus_view[:1]
|
511 |
+
latents_denoised_batch = latent_reference_plus_view[1:]
|
512 |
+
# save views scheduler status after sample
|
513 |
+
views_scheduler_status[j] = copy.deepcopy(pipeline.scheduler.__dict__)
|
514 |
+
|
515 |
+
# extract value from batch
|
516 |
+
for latents_view_denoised, (h_start, h_end, w_start, w_end) in zip(
|
517 |
+
latents_denoised_batch.chunk(vb_size), batch_view
|
518 |
+
):
|
519 |
+
if circular_padding and w_end > latents.shape[3]:
|
520 |
+
# Case for circular padding
|
521 |
+
value[:, :, h_start:h_end, w_start:] += latents_view_denoised[
|
522 |
+
:, :, h_start:h_end, : latents.shape[3] - w_start
|
523 |
+
]
|
524 |
+
value[:, :, h_start:h_end, : w_end - latents.shape[3]] += latents_view_denoised[
|
525 |
+
:, :, h_start:h_end,
|
526 |
+
latents.shape[3] - w_start:
|
527 |
+
]
|
528 |
+
count[:, :, h_start:h_end, w_start:] += 1
|
529 |
+
count[:, :, h_start:h_end, : w_end - latents.shape[3]] += 1
|
530 |
+
else:
|
531 |
+
value[:, :, h_start:h_end, w_start:w_end] += latents_view_denoised
|
532 |
+
count[:, :, h_start:h_end, w_start:w_end] += 1
|
533 |
+
|
534 |
+
# take the MultiDiffusion step. Eq. 5 in MultiDiffusion paper: https://arxiv.org/abs/2302.08113
|
535 |
+
latents = torch.where(count > 0, value / count, value)
|
536 |
+
|
537 |
+
# call the callback, if provided
|
538 |
+
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % pipeline.scheduler.order == 0):
|
539 |
+
progress_bar.update()
|
540 |
+
|
541 |
+
if circular_padding:
|
542 |
+
image = pipeline.decode_latents_with_padding(latents)
|
543 |
+
else:
|
544 |
+
image = pipeline.vae.decode(latents / pipeline.vae.config.scaling_factor, return_dict=False)[0]
|
545 |
+
reference_image = pipeline.vae.decode(reference_latent / pipeline.vae.config.scaling_factor, return_dict=False)[0]
|
546 |
+
# image, has_nsfw_concept = pipeline.run_safety_checker(image, device, prompt_embeds.dtype)
|
547 |
+
# reference_image, _ = pipeline.run_safety_checker(reference_image, device, prompt_embeds.dtype)
|
548 |
+
|
549 |
+
image = pipeline.image_processor.postprocess(image, output_type='pil', do_denormalize=[True])
|
550 |
+
reference_image = pipeline.image_processor.postprocess(reference_image, output_type='pil', do_denormalize=[True])
|
551 |
+
pipeline.maybe_free_model_hooks()
|
552 |
+
return reference_image + image
|