Spaces:
Build error
Build error
updated
Browse files- app.py +1 -6
- text_to_animation/model.py +75 -321
- text_to_animation/models/cross_frame_attention_flax.py +0 -1
- text_to_animation/models/unet_3d_blocks_flax.py +717 -0
- text_to_animation/models/unet_3d_condition_flax.py +611 -0
- text_to_animation/pipelines/text_to_video_pipeline_flax.py +887 -142
- webui/app_control_animation.py +87 -120
app.py
CHANGED
@@ -39,7 +39,7 @@ Our code uses <a href="https://www.humphreyshi.com/home">Text2Video-Zero</a> and
|
|
39 |
notice = """
|
40 |
<p>For faster inference without waiting in queue, you may duplicate the space and upgrade to GPU in settings.
|
41 |
<br/>
|
42 |
-
<a href="https://
|
43 |
<img style="margin-top: 0em; margin-bottom: 0em" src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a>
|
44 |
</p>
|
45 |
"""
|
@@ -51,13 +51,8 @@ with gr.Blocks(css="style.css") as demo:
|
|
51 |
if on_huggingspace:
|
52 |
gr.HTML(notice)
|
53 |
|
54 |
-
# NOTE: In our final demo we should consider removing zero-shot t2v and pose conditional
|
55 |
with gr.Tab("Control Animation"):
|
56 |
create_demo_animation(model)
|
57 |
-
# with gr.Tab("Zero-Shot Text2Video"):
|
58 |
-
# create_demo_text_to_video(model)
|
59 |
-
# with gr.Tab("Pose Conditional"):
|
60 |
-
# create_demo_pose(model)
|
61 |
|
62 |
if on_huggingspace:
|
63 |
demo.queue(max_size=20)
|
|
|
39 |
notice = """
|
40 |
<p>For faster inference without waiting in queue, you may duplicate the space and upgrade to GPU in settings.
|
41 |
<br/>
|
42 |
+
<a href="https://huggingface.co/spaces/Pie31415/control-animation?duplicate=true">
|
43 |
<img style="margin-top: 0em; margin-bottom: 0em" src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a>
|
44 |
</p>
|
45 |
"""
|
|
|
51 |
if on_huggingspace:
|
52 |
gr.HTML(notice)
|
53 |
|
|
|
54 |
with gr.Tab("Control Animation"):
|
55 |
create_demo_animation(model)
|
|
|
|
|
|
|
|
|
56 |
|
57 |
if on_huggingspace:
|
58 |
demo.queue(max_size=20)
|
text_to_animation/model.py
CHANGED
@@ -3,7 +3,6 @@ from enum import Enum
|
|
3 |
import gc
|
4 |
import numpy as np
|
5 |
import jax.numpy as jnp
|
6 |
-
import tomesd
|
7 |
import jax
|
8 |
|
9 |
from PIL import Image
|
@@ -20,9 +19,12 @@ from diffusers import (
|
|
20 |
FlaxAutoencoderKL,
|
21 |
FlaxStableDiffusionControlNetPipeline,
|
22 |
StableDiffusionPipeline,
|
|
|
23 |
)
|
24 |
-
from text_to_animation.models.unet_2d_condition_flax import
|
25 |
-
|
|
|
|
|
26 |
|
27 |
from text_to_animation.pipelines.text_to_video_pipeline_flax import (
|
28 |
FlaxTextToVideoPipeline,
|
@@ -48,37 +50,31 @@ def replicate_devices(array):
|
|
48 |
|
49 |
|
50 |
class ControlAnimationModel:
|
51 |
-
def __init__(self,
|
52 |
-
self.device = device
|
53 |
self.dtype = dtype
|
54 |
self.rng = jax.random.PRNGKey(0)
|
55 |
-
self.pipe_dict = {
|
56 |
-
ModelType.Text2Video: FlaxTextToVideoPipeline, # TODO: Replace with our TextToVideo JAX Pipeline
|
57 |
-
ModelType.ControlNetPose: FlaxStableDiffusionControlNetPipeline,
|
58 |
-
}
|
59 |
self.pipe = None
|
60 |
self.model_type = None
|
61 |
|
62 |
self.states = {}
|
63 |
self.model_name = ""
|
64 |
|
65 |
-
self.from_local = True # if the attn model is available in local (after adaptation by adapt_attn.py)
|
66 |
-
|
67 |
def set_model(
|
68 |
self,
|
69 |
-
model_type: ModelType,
|
70 |
model_id: str,
|
71 |
-
controlnet,
|
72 |
-
controlnet_params,
|
73 |
-
tokenizer,
|
74 |
-
scheduler,
|
75 |
-
scheduler_state,
|
76 |
**kwargs,
|
77 |
):
|
78 |
if hasattr(self, "pipe") and self.pipe is not None:
|
79 |
del self.pipe
|
80 |
self.pipe = None
|
81 |
gc.collect()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
82 |
scheduler, scheduler_state = FlaxDDIMScheduler.from_pretrained(
|
83 |
model_id, subfolder="scheduler", from_pt=True
|
84 |
)
|
@@ -86,17 +82,12 @@ class ControlAnimationModel:
|
|
86 |
feature_extractor = CLIPFeatureExtractor.from_pretrained(
|
87 |
model_id, subfolder="feature_extractor"
|
88 |
)
|
89 |
-
|
90 |
-
unet,
|
91 |
-
|
92 |
-
|
93 |
-
|
94 |
-
|
95 |
-
)
|
96 |
-
else:
|
97 |
-
unet, unet_params = FlaxUNet2DConditionModel.from_pretrained(
|
98 |
-
model_id, subfolder="unet", from_pt=True, dtype=self.dtype
|
99 |
-
)
|
100 |
vae, vae_params = FlaxAutoencoderKL.from_pretrained(
|
101 |
model_id, subfolder="vae", from_pt=True, dtype=self.dtype
|
102 |
)
|
@@ -108,6 +99,7 @@ class ControlAnimationModel:
|
|
108 |
text_encoder=text_encoder,
|
109 |
tokenizer=tokenizer,
|
110 |
unet=unet,
|
|
|
111 |
controlnet=controlnet,
|
112 |
scheduler=scheduler,
|
113 |
safety_checker=None,
|
@@ -121,313 +113,52 @@ class ControlAnimationModel:
|
|
121 |
"text_encoder": text_encoder.params,
|
122 |
}
|
123 |
self.p_params = jax_utils.replicate(self.params)
|
124 |
-
|
125 |
-
self.model_type = model_type
|
126 |
self.model_name = model_id
|
127 |
|
128 |
-
|
129 |
-
|
130 |
-
# prompt_ids = self.pipe.prepare_text_inputs(prompt)
|
131 |
-
# n_prompt_ids = self.pipe.prepare_text_inputs(negative_prompt)
|
132 |
-
# latents = kwargs.pop('latents')
|
133 |
-
# # rng = jax.random.split(self.rng, jax.device_count())
|
134 |
-
# prng, self.rng = jax.random.split(self.rng)
|
135 |
-
# #prng = jax.numpy.stack([prng] * jax.device_count())#same prng seed on every device
|
136 |
-
# prng_seed = jax.random.split(prng, jax.device_count())
|
137 |
-
# image = replicate_devices(image[frame_ids])
|
138 |
-
# latents = replicate_devices(latents)
|
139 |
-
# prompt_ids = replicate_devices(prompt_ids)
|
140 |
-
# n_prompt_ids = replicate_devices(n_prompt_ids)
|
141 |
-
# return (self.pipe(image=image,
|
142 |
-
# latents=latents,
|
143 |
-
# prompt_ids=prompt_ids,
|
144 |
-
# neg_prompt_ids=n_prompt_ids,
|
145 |
-
# params=self.p_params,
|
146 |
-
# prng_seed=prng_seed, jit = True,
|
147 |
-
# ).images)[0]
|
148 |
-
|
149 |
-
def inference(self, image, split_to_chunks=False, chunk_size=8, **kwargs):
|
150 |
-
if not hasattr(self, "pipe") or self.pipe is None:
|
151 |
-
return
|
152 |
-
|
153 |
-
if "merging_ratio" in kwargs:
|
154 |
-
merging_ratio = kwargs.pop("merging_ratio")
|
155 |
-
|
156 |
-
# if merging_ratio > 0:
|
157 |
-
tomesd.apply_patch(self.pipe, ratio=merging_ratio)
|
158 |
-
|
159 |
-
# f = image.shape[0]
|
160 |
-
|
161 |
-
assert "prompt" in kwargs
|
162 |
-
prompt = [kwargs.pop("prompt")]
|
163 |
-
negative_prompt = [kwargs.pop("negative_prompt", "")]
|
164 |
-
|
165 |
-
frames_counter = 0
|
166 |
-
|
167 |
-
# Processing chunk-by-chunk
|
168 |
-
if split_to_chunks:
|
169 |
-
pass
|
170 |
-
# # not tested
|
171 |
-
# f = image.shape[0]
|
172 |
-
# chunk_ids = np.arange(0, f, chunk_size - 1)
|
173 |
-
# result = []
|
174 |
-
# for i in range(len(chunk_ids)):
|
175 |
-
# ch_start = chunk_ids[i]
|
176 |
-
# ch_end = f if i == len(chunk_ids) - 1 else chunk_ids[i + 1]
|
177 |
-
# frame_ids = [0] + list(range(ch_start, ch_end))
|
178 |
-
# print(f'Processing chunk {i + 1} / {len(chunk_ids)}')
|
179 |
-
# result.append(self.inference_chunk(image=image,
|
180 |
-
# frame_ids=frame_ids,
|
181 |
-
# prompt=prompt,
|
182 |
-
# negative_prompt=negative_prompt,
|
183 |
-
# **kwargs).images[1:])
|
184 |
-
# frames_counter += len(chunk_ids)-1
|
185 |
-
# if on_huggingspace and frames_counter >= 80:
|
186 |
-
# break
|
187 |
-
# result = np.concatenate(result)
|
188 |
-
# return result
|
189 |
-
else:
|
190 |
-
if "jit" in kwargs and kwargs.pop("jit"):
|
191 |
-
prompt_ids = self.pipe.prepare_text_inputs(prompt)
|
192 |
-
n_prompt_ids = self.pipe.prepare_text_inputs(negative_prompt)
|
193 |
-
latents = kwargs.pop("latents")
|
194 |
-
prng, self.rng = jax.random.split(self.rng)
|
195 |
-
prng_seed = jax.random.split(prng, jax.device_count())
|
196 |
-
image = replicate_devices(image)
|
197 |
-
latents = replicate_devices(latents)
|
198 |
-
prompt_ids = replicate_devices(prompt_ids)
|
199 |
-
n_prompt_ids = replicate_devices(n_prompt_ids)
|
200 |
-
return (
|
201 |
-
self.pipe(
|
202 |
-
image=image,
|
203 |
-
latents=latents,
|
204 |
-
prompt_ids=prompt_ids,
|
205 |
-
neg_prompt_ids=n_prompt_ids,
|
206 |
-
params=self.p_params,
|
207 |
-
prng_seed=prng_seed,
|
208 |
-
jit=True,
|
209 |
-
).images
|
210 |
-
)[0]
|
211 |
-
else:
|
212 |
-
prompt_ids = self.pipe.prepare_text_inputs(prompt)
|
213 |
-
n_prompt_ids = self.pipe.prepare_text_inputs(negative_prompt)
|
214 |
-
latents = kwargs.pop("latents")
|
215 |
-
prng_seed, self.rng = jax.random.split(self.rng)
|
216 |
-
return self.pipe(
|
217 |
-
image=image,
|
218 |
-
latents=latents,
|
219 |
-
prompt_ids=prompt_ids,
|
220 |
-
neg_prompt_ids=n_prompt_ids,
|
221 |
-
params=self.params,
|
222 |
-
prng_seed=prng_seed,
|
223 |
-
jit=False,
|
224 |
-
).images
|
225 |
-
|
226 |
-
def process_controlnet_pose(
|
227 |
self,
|
228 |
-
|
229 |
-
|
230 |
-
|
231 |
-
|
232 |
-
|
233 |
-
|
234 |
-
|
235 |
-
|
236 |
-
|
237 |
-
eta=0.0,
|
238 |
-
resolution=512,
|
239 |
-
use_cf_attn=True,
|
240 |
-
save_path=None,
|
241 |
-
):
|
242 |
-
print("Module Pose")
|
243 |
video_path = gradio_utils.motion_to_video_path(video_path)
|
244 |
-
if self.model_type != ModelType.ControlNetPose:
|
245 |
-
controlnet = FlaxControlNetModel.from_pretrained(
|
246 |
-
"fusing/stable-diffusion-v1-5-controlnet-openpose"
|
247 |
-
)
|
248 |
-
self.set_model(
|
249 |
-
ModelType.ControlNetPose,
|
250 |
-
model_id="runwayml/stable-diffusion-v1-5",
|
251 |
-
controlnet=controlnet,
|
252 |
-
)
|
253 |
-
self.pipe.scheduler = FlaxDDIMScheduler.from_config(
|
254 |
-
self.pipe.scheduler.config
|
255 |
-
)
|
256 |
-
if use_cf_attn:
|
257 |
-
self.pipe.unet.set_attn_processor(processor=self.controlnet_attn_proc)
|
258 |
-
self.pipe.controlnet.set_attn_processor(
|
259 |
-
processor=self.controlnet_attn_proc
|
260 |
-
)
|
261 |
|
262 |
-
|
263 |
-
|
264 |
-
if "Motion" in video_path
|
265 |
-
else video_path
|
266 |
-
)
|
267 |
|
268 |
-
|
269 |
-
negative_prompts =
|
270 |
|
271 |
video, fps = utils.prepare_video(
|
272 |
-
video_path, resolution,
|
273 |
-
)
|
274 |
-
control = (
|
275 |
-
utils.pre_process_pose(video, apply_pose_detect=False)
|
276 |
-
.to(self.device)
|
277 |
-
.to(self.dtype)
|
278 |
)
|
279 |
-
|
280 |
-
|
281 |
-
|
282 |
-
|
283 |
-
|
284 |
-
|
285 |
-
|
286 |
-
|
287 |
-
|
288 |
-
|
289 |
-
image=control,
|
290 |
-
prompt=prompt + ", " + added_prompt,
|
291 |
-
height=h,
|
292 |
-
width=w,
|
293 |
-
negative_prompt=negative_prompts,
|
294 |
-
num_inference_steps=num_inference_steps,
|
295 |
-
guidance_scale=guidance_scale,
|
296 |
-
controlnet_conditioning_scale=controlnet_conditioning_scale,
|
297 |
-
eta=eta,
|
298 |
-
latents=latents,
|
299 |
-
seed=seed,
|
300 |
-
output_type="numpy",
|
301 |
-
split_to_chunks=True,
|
302 |
-
chunk_size=chunk_size,
|
303 |
-
merging_ratio=merging_ratio,
|
304 |
)
|
305 |
-
return utils.create_gif(
|
306 |
-
result,
|
307 |
-
fps,
|
308 |
-
path=save_path,
|
309 |
-
watermark=gradio_utils.logo_name_to_path(watermark),
|
310 |
-
)
|
311 |
-
|
312 |
-
def process_text2video(
|
313 |
-
self,
|
314 |
-
prompt,
|
315 |
-
model_name="dreamlike-art/dreamlike-photoreal-2.0",
|
316 |
-
motion_field_strength_x=12,
|
317 |
-
motion_field_strength_y=12,
|
318 |
-
t0=44,
|
319 |
-
t1=47,
|
320 |
-
n_prompt="",
|
321 |
-
chunk_size=8,
|
322 |
-
video_length=8,
|
323 |
-
watermark="Picsart AI Research",
|
324 |
-
merging_ratio=0.0,
|
325 |
-
seed=0,
|
326 |
-
resolution=512,
|
327 |
-
fps=2,
|
328 |
-
use_cf_attn=True,
|
329 |
-
use_motion_field=True,
|
330 |
-
smooth_bg=False,
|
331 |
-
smooth_bg_strength=0.4,
|
332 |
-
path=None,
|
333 |
-
):
|
334 |
-
print("Module Text2Video")
|
335 |
-
if self.model_type != ModelType.Text2Video or model_name != self.model_name:
|
336 |
-
print("Model update")
|
337 |
-
unet = FlaxUNet2DConditionModel.from_pretrained(
|
338 |
-
model_name, subfolder="unet"
|
339 |
-
)
|
340 |
-
self.set_model(ModelType.Text2Video, model_id=model_name, unet=unet)
|
341 |
-
self.pipe.scheduler = FlaxDDIMScheduler.from_config(
|
342 |
-
self.pipe.scheduler.config
|
343 |
-
)
|
344 |
-
if use_cf_attn:
|
345 |
-
self.pipe.unet.set_attn_processor(processor=self.text2video_attn_proc)
|
346 |
-
self.generator.manual_seed(seed)
|
347 |
|
348 |
-
|
349 |
-
negative_prompts = "longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer difits, cropped, worst quality, low quality, deformed body, bloated, ugly, unrealistic"
|
350 |
-
|
351 |
-
prompt = prompt.rstrip()
|
352 |
-
if len(prompt) > 0 and (prompt[-1] == "," or prompt[-1] == "."):
|
353 |
-
prompt = prompt.rstrip()[:-1]
|
354 |
-
prompt = prompt.rstrip()
|
355 |
-
prompt = prompt + ", " + added_prompt
|
356 |
-
if len(n_prompt) > 0:
|
357 |
-
negative_prompt = n_prompt
|
358 |
-
else:
|
359 |
-
negative_prompt = None
|
360 |
-
|
361 |
-
result = self.inference(
|
362 |
-
prompt=prompt,
|
363 |
-
video_length=video_length,
|
364 |
-
height=resolution,
|
365 |
-
width=resolution,
|
366 |
-
num_inference_steps=50,
|
367 |
-
guidance_scale=7.5,
|
368 |
-
guidance_stop_step=1.0,
|
369 |
-
t0=t0,
|
370 |
-
t1=t1,
|
371 |
-
motion_field_strength_x=motion_field_strength_x,
|
372 |
-
motion_field_strength_y=motion_field_strength_y,
|
373 |
-
use_motion_field=use_motion_field,
|
374 |
-
smooth_bg=smooth_bg,
|
375 |
-
smooth_bg_strength=smooth_bg_strength,
|
376 |
-
seed=seed,
|
377 |
-
output_type="numpy",
|
378 |
-
negative_prompt=negative_prompt,
|
379 |
-
merging_ratio=merging_ratio,
|
380 |
-
split_to_chunks=True,
|
381 |
-
chunk_size=chunk_size,
|
382 |
-
)
|
383 |
-
return utils.create_video(
|
384 |
-
result, fps, path=path, watermark=gradio_utils.logo_name_to_path(watermark)
|
385 |
-
)
|
386 |
|
387 |
-
|
388 |
-
def to_pil_images(images: torch.Tensor) -> List[Image.Image]:
|
389 |
-
images = (images / 2 + 0.5).clamp(0, 1)
|
390 |
-
images = images.cpu().permute(0, 2, 3, 1).float().numpy()
|
391 |
-
images = np.round(images * 255).astype(np.uint8)
|
392 |
-
return [Image.fromarray(image) for image in images]
|
393 |
-
|
394 |
-
def generate_initial_frames(
|
395 |
-
self,
|
396 |
-
prompt: str,
|
397 |
-
model_link: str = "dreamlike-art/dreamlike-photoreal-2.0",
|
398 |
-
is_safetensor: bool = False,
|
399 |
-
n_prompt: str = "",
|
400 |
-
width: int = 512,
|
401 |
-
height: int = 512,
|
402 |
-
# batch_count: int = 4,
|
403 |
-
# batch_size: int = 1,
|
404 |
-
cfg_scale: float = 7.0,
|
405 |
-
seed: int = 0,
|
406 |
-
) -> List[Image.Image]:
|
407 |
-
generator = torch.Generator(device=self.device).manual_seed(seed)
|
408 |
-
pipe = StableDiffusionPipeline.from_pretrained(model_link)
|
409 |
-
|
410 |
-
batch_size = 4
|
411 |
-
prompt = [prompt] * batch_size
|
412 |
-
negative_prompt = [n_prompt] * batch_size
|
413 |
-
|
414 |
-
images = pipe(
|
415 |
-
prompt,
|
416 |
-
negative_prompt=negative_prompt,
|
417 |
-
width=width,
|
418 |
-
height=height,
|
419 |
-
guidance_scale=cfg_scale,
|
420 |
-
generator=generator,
|
421 |
-
).images
|
422 |
-
pil_images = self.to_pil_images(images)
|
423 |
-
|
424 |
-
return pil_images
|
425 |
|
426 |
def generate_animation(
|
427 |
self,
|
428 |
prompt: str,
|
|
|
|
|
429 |
model_link: str = "dreamlike-art/dreamlike-photoreal-2.0",
|
430 |
-
is_safetensor: bool = False,
|
431 |
motion_field_strength_x: int = 12,
|
432 |
motion_field_strength_y: int = 12,
|
433 |
t0: int = 44,
|
@@ -445,6 +176,29 @@ class ControlAnimationModel:
|
|
445 |
smooth_bg_strength: float = 0.4,
|
446 |
path: str = None,
|
447 |
):
|
448 |
-
|
449 |
-
|
450 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
import gc
|
4 |
import numpy as np
|
5 |
import jax.numpy as jnp
|
|
|
6 |
import jax
|
7 |
|
8 |
from PIL import Image
|
|
|
19 |
FlaxAutoencoderKL,
|
20 |
FlaxStableDiffusionControlNetPipeline,
|
21 |
StableDiffusionPipeline,
|
22 |
+
FlaxUNet2DConditionModel,
|
23 |
)
|
24 |
+
from text_to_animation.models.unet_2d_condition_flax import (
|
25 |
+
FlaxUNet2DConditionModel as CustomFlaxUNet2DConditionModel,
|
26 |
+
)
|
27 |
+
from diffusers import FlaxControlNetModel
|
28 |
|
29 |
from text_to_animation.pipelines.text_to_video_pipeline_flax import (
|
30 |
FlaxTextToVideoPipeline,
|
|
|
50 |
|
51 |
|
52 |
class ControlAnimationModel:
|
53 |
+
def __init__(self, dtype, **kwargs):
|
|
|
54 |
self.dtype = dtype
|
55 |
self.rng = jax.random.PRNGKey(0)
|
|
|
|
|
|
|
|
|
56 |
self.pipe = None
|
57 |
self.model_type = None
|
58 |
|
59 |
self.states = {}
|
60 |
self.model_name = ""
|
61 |
|
|
|
|
|
62 |
def set_model(
|
63 |
self,
|
|
|
64 |
model_id: str,
|
|
|
|
|
|
|
|
|
|
|
65 |
**kwargs,
|
66 |
):
|
67 |
if hasattr(self, "pipe") and self.pipe is not None:
|
68 |
del self.pipe
|
69 |
self.pipe = None
|
70 |
gc.collect()
|
71 |
+
|
72 |
+
controlnet, controlnet_params = FlaxControlNetModel.from_pretrained(
|
73 |
+
"fusing/stable-diffusion-v1-5-controlnet-openpose",
|
74 |
+
from_pt=True,
|
75 |
+
dtype=jnp.float16,
|
76 |
+
)
|
77 |
+
|
78 |
scheduler, scheduler_state = FlaxDDIMScheduler.from_pretrained(
|
79 |
model_id, subfolder="scheduler", from_pt=True
|
80 |
)
|
|
|
82 |
feature_extractor = CLIPFeatureExtractor.from_pretrained(
|
83 |
model_id, subfolder="feature_extractor"
|
84 |
)
|
85 |
+
unet, unet_params = CustomFlaxUNet2DConditionModel.from_pretrained(
|
86 |
+
model_id, subfolder="unet", from_pt=True, dtype=self.dtype
|
87 |
+
)
|
88 |
+
unet_vanilla, _ = FlaxUNet2DConditionModel.from_pretrained(
|
89 |
+
model_id, subfolder="unet", from_pt=True, dtype=self.dtype
|
90 |
+
)
|
|
|
|
|
|
|
|
|
|
|
91 |
vae, vae_params = FlaxAutoencoderKL.from_pretrained(
|
92 |
model_id, subfolder="vae", from_pt=True, dtype=self.dtype
|
93 |
)
|
|
|
99 |
text_encoder=text_encoder,
|
100 |
tokenizer=tokenizer,
|
101 |
unet=unet,
|
102 |
+
unet_vanilla=unet_vanilla,
|
103 |
controlnet=controlnet,
|
104 |
scheduler=scheduler,
|
105 |
safety_checker=None,
|
|
|
113 |
"text_encoder": text_encoder.params,
|
114 |
}
|
115 |
self.p_params = jax_utils.replicate(self.params)
|
|
|
|
|
116 |
self.model_name = model_id
|
117 |
|
118 |
+
def generate_initial_frames(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
119 |
self,
|
120 |
+
prompt: str,
|
121 |
+
video_path: str,
|
122 |
+
n_prompt: str = "",
|
123 |
+
num_imgs: int = 4,
|
124 |
+
resolution: int = 512,
|
125 |
+
model_id: str = "runwayml/stable-diffusion-v1-5",
|
126 |
+
) -> List[Image.Image]:
|
127 |
+
self.set_model(model_id=model_id)
|
128 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
129 |
video_path = gradio_utils.motion_to_video_path(video_path)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
130 |
|
131 |
+
added_prompt = "high quality, best quality, HD, clay stop-motion, claymation, HQ, masterpiece, art, smooth"
|
132 |
+
prompts = added_prompt + ", " + prompt
|
|
|
|
|
|
|
133 |
|
134 |
+
added_n_prompt = "longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer difits, cropped, worst quality, low quality, deformed body, bloated, ugly"
|
135 |
+
negative_prompts = added_n_prompt + ", " + n_prompt
|
136 |
|
137 |
video, fps = utils.prepare_video(
|
138 |
+
video_path, resolution, None, self.dtype, False, output_fps=4
|
|
|
|
|
|
|
|
|
|
|
139 |
)
|
140 |
+
control = utils.pre_process_pose(video, apply_pose_detect=False)
|
141 |
+
|
142 |
+
seeds = [seed for seed in jax.random.randint(self.rng, [num_imgs], 0, 65536)]
|
143 |
+
prngs = [jax.random.PRNGKey(seed) for seed in seeds]
|
144 |
+
images = self.pipe.generate_starting_frames(
|
145 |
+
params=self.params,
|
146 |
+
prngs=prngs,
|
147 |
+
controlnet_image=control,
|
148 |
+
prompt=prompts,
|
149 |
+
neg_prompt=negative_prompts,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
150 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
151 |
|
152 |
+
images = [np.array(images[i]) for i in range(images.shape[0])]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
153 |
|
154 |
+
return images
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
155 |
|
156 |
def generate_animation(
|
157 |
self,
|
158 |
prompt: str,
|
159 |
+
initial_frame_index: int,
|
160 |
+
input_video_path: str,
|
161 |
model_link: str = "dreamlike-art/dreamlike-photoreal-2.0",
|
|
|
162 |
motion_field_strength_x: int = 12,
|
163 |
motion_field_strength_y: int = 12,
|
164 |
t0: int = 44,
|
|
|
176 |
smooth_bg_strength: float = 0.4,
|
177 |
path: str = None,
|
178 |
):
|
179 |
+
video_path = gradio_utils.motion_to_video_path(video_path)
|
180 |
+
|
181 |
+
# added_prompt = 'best quality, HD, clay stop-motion, claymation, HQ, masterpiece, art, smooth'
|
182 |
+
# added_prompt = 'high quality, anatomically correct, clay stop-motion, aardman, claymation, smooth'
|
183 |
+
added_n_prompt = "longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer difits, cropped, worst quality, low quality, deformed body, bloated, ugly"
|
184 |
+
negative_prompts = added_n_prompt + ", " + n_prompt
|
185 |
+
|
186 |
+
video, fps = utils.prepare_video(
|
187 |
+
video_path, resolution, None, self.dtype, False, output_fps=4
|
188 |
+
)
|
189 |
+
control = utils.pre_process_pose(video, apply_pose_detect=False)
|
190 |
+
f, _, h, w = video.shape
|
191 |
+
|
192 |
+
prng_seed = jax.random.PRNGKey(seed)
|
193 |
+
vid = self.pipe.generate_video(
|
194 |
+
prompt,
|
195 |
+
image=control,
|
196 |
+
params=self.params,
|
197 |
+
prng_seed=prng_seed,
|
198 |
+
neg_prompt="",
|
199 |
+
controlnet_conditioning_scale=1.0,
|
200 |
+
motion_field_strength_x=3,
|
201 |
+
motion_field_strength_y=4,
|
202 |
+
jit=True,
|
203 |
+
).image
|
204 |
+
return utils.create_gif(np.array(vid), 4, path=None, watermark=None)
|
text_to_animation/models/cross_frame_attention_flax.py
CHANGED
@@ -50,7 +50,6 @@ class FlaxCrossFrameAttention(nn.Module):
|
|
50 |
batch_size: The number that represents actual batch size, other than the frames.
|
51 |
For example, using calling unet with a single prompt and num_images_per_prompt=1, batch_size should be
|
52 |
equal to 2, due to classifier-free guidance.
|
53 |
-
|
54 |
"""
|
55 |
query_dim: int
|
56 |
heads: int = 8
|
|
|
50 |
batch_size: The number that represents actual batch size, other than the frames.
|
51 |
For example, using calling unet with a single prompt and num_images_per_prompt=1, batch_size should be
|
52 |
equal to 2, due to classifier-free guidance.
|
|
|
53 |
"""
|
54 |
query_dim: int
|
55 |
heads: int = 8
|
text_to_animation/models/unet_3d_blocks_flax.py
ADDED
@@ -0,0 +1,717 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright 2023 The HuggingFace Team. All rights reserved.
|
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 |
+
import torch
|
16 |
+
from torch import nn
|
17 |
+
|
18 |
+
# from .resnet import Downsample2D, ResnetBlock2D, TemporalConvLayer, Upsample2D
|
19 |
+
# from diffusers.models.transformer_2d import Transformer2DModel
|
20 |
+
# from .transformer_temporal import TransformerTemporalModel
|
21 |
+
|
22 |
+
from diffusers.models.resnet_flax import (
|
23 |
+
FlaxDownsample2D,
|
24 |
+
FlaxResnetBlock2D,
|
25 |
+
FlaxUpsample2D,
|
26 |
+
)
|
27 |
+
from diffusers.models.attention_flax import FlaxTransformer2DModel
|
28 |
+
from diffusers.models.transformer_temporal import (
|
29 |
+
TransformerTemporalModel,
|
30 |
+
) # TODO: convert to flax
|
31 |
+
|
32 |
+
|
33 |
+
def get_down_block(
|
34 |
+
down_block_type,
|
35 |
+
num_layers,
|
36 |
+
in_channels,
|
37 |
+
out_channels,
|
38 |
+
temb_channels,
|
39 |
+
add_downsample,
|
40 |
+
resnet_eps,
|
41 |
+
resnet_act_fn,
|
42 |
+
attn_num_head_channels,
|
43 |
+
resnet_groups=None,
|
44 |
+
cross_attention_dim=None,
|
45 |
+
downsample_padding=None,
|
46 |
+
dual_cross_attention=False,
|
47 |
+
use_linear_projection=True,
|
48 |
+
only_cross_attention=False,
|
49 |
+
upcast_attention=False,
|
50 |
+
resnet_time_scale_shift="default",
|
51 |
+
):
|
52 |
+
if down_block_type == "DownBlock3D":
|
53 |
+
return DownBlock3D(
|
54 |
+
num_layers=num_layers,
|
55 |
+
in_channels=in_channels,
|
56 |
+
out_channels=out_channels,
|
57 |
+
temb_channels=temb_channels,
|
58 |
+
add_downsample=add_downsample,
|
59 |
+
resnet_eps=resnet_eps,
|
60 |
+
resnet_act_fn=resnet_act_fn,
|
61 |
+
resnet_groups=resnet_groups,
|
62 |
+
downsample_padding=downsample_padding,
|
63 |
+
resnet_time_scale_shift=resnet_time_scale_shift,
|
64 |
+
)
|
65 |
+
elif down_block_type == "CrossAttnDownBlock3D":
|
66 |
+
if cross_attention_dim is None:
|
67 |
+
raise ValueError(
|
68 |
+
"cross_attention_dim must be specified for CrossAttnDownBlock3D"
|
69 |
+
)
|
70 |
+
return CrossAttnDownBlock3D(
|
71 |
+
num_layers=num_layers,
|
72 |
+
in_channels=in_channels,
|
73 |
+
out_channels=out_channels,
|
74 |
+
temb_channels=temb_channels,
|
75 |
+
add_downsample=add_downsample,
|
76 |
+
resnet_eps=resnet_eps,
|
77 |
+
resnet_act_fn=resnet_act_fn,
|
78 |
+
resnet_groups=resnet_groups,
|
79 |
+
downsample_padding=downsample_padding,
|
80 |
+
cross_attention_dim=cross_attention_dim,
|
81 |
+
attn_num_head_channels=attn_num_head_channels,
|
82 |
+
dual_cross_attention=dual_cross_attention,
|
83 |
+
use_linear_projection=use_linear_projection,
|
84 |
+
only_cross_attention=only_cross_attention,
|
85 |
+
upcast_attention=upcast_attention,
|
86 |
+
resnet_time_scale_shift=resnet_time_scale_shift,
|
87 |
+
)
|
88 |
+
raise ValueError(f"{down_block_type} does not exist.")
|
89 |
+
|
90 |
+
|
91 |
+
def get_up_block(
|
92 |
+
up_block_type,
|
93 |
+
num_layers,
|
94 |
+
in_channels,
|
95 |
+
out_channels,
|
96 |
+
prev_output_channel,
|
97 |
+
temb_channels,
|
98 |
+
add_upsample,
|
99 |
+
resnet_eps,
|
100 |
+
resnet_act_fn,
|
101 |
+
attn_num_head_channels,
|
102 |
+
resnet_groups=None,
|
103 |
+
cross_attention_dim=None,
|
104 |
+
dual_cross_attention=False,
|
105 |
+
use_linear_projection=True,
|
106 |
+
only_cross_attention=False,
|
107 |
+
upcast_attention=False,
|
108 |
+
resnet_time_scale_shift="default",
|
109 |
+
):
|
110 |
+
if up_block_type == "UpBlock3D":
|
111 |
+
return UpBlock3D(
|
112 |
+
num_layers=num_layers,
|
113 |
+
in_channels=in_channels,
|
114 |
+
out_channels=out_channels,
|
115 |
+
prev_output_channel=prev_output_channel,
|
116 |
+
temb_channels=temb_channels,
|
117 |
+
add_upsample=add_upsample,
|
118 |
+
resnet_eps=resnet_eps,
|
119 |
+
resnet_act_fn=resnet_act_fn,
|
120 |
+
resnet_groups=resnet_groups,
|
121 |
+
resnet_time_scale_shift=resnet_time_scale_shift,
|
122 |
+
)
|
123 |
+
elif up_block_type == "CrossAttnUpBlock3D":
|
124 |
+
if cross_attention_dim is None:
|
125 |
+
raise ValueError(
|
126 |
+
"cross_attention_dim must be specified for CrossAttnUpBlock3D"
|
127 |
+
)
|
128 |
+
return CrossAttnUpBlock3D(
|
129 |
+
num_layers=num_layers,
|
130 |
+
in_channels=in_channels,
|
131 |
+
out_channels=out_channels,
|
132 |
+
prev_output_channel=prev_output_channel,
|
133 |
+
temb_channels=temb_channels,
|
134 |
+
add_upsample=add_upsample,
|
135 |
+
resnet_eps=resnet_eps,
|
136 |
+
resnet_act_fn=resnet_act_fn,
|
137 |
+
resnet_groups=resnet_groups,
|
138 |
+
cross_attention_dim=cross_attention_dim,
|
139 |
+
attn_num_head_channels=attn_num_head_channels,
|
140 |
+
dual_cross_attention=dual_cross_attention,
|
141 |
+
use_linear_projection=use_linear_projection,
|
142 |
+
only_cross_attention=only_cross_attention,
|
143 |
+
upcast_attention=upcast_attention,
|
144 |
+
resnet_time_scale_shift=resnet_time_scale_shift,
|
145 |
+
)
|
146 |
+
raise ValueError(f"{up_block_type} does not exist.")
|
147 |
+
|
148 |
+
|
149 |
+
class FlaxUNetMidBlock3DCrossAttn(nn.Module):
|
150 |
+
def __init__(
|
151 |
+
self,
|
152 |
+
in_channels: int,
|
153 |
+
temb_channels: int,
|
154 |
+
dropout: float = 0.0,
|
155 |
+
num_layers: int = 1,
|
156 |
+
resnet_eps: float = 1e-6,
|
157 |
+
resnet_time_scale_shift: str = "default",
|
158 |
+
resnet_act_fn: str = "swish",
|
159 |
+
resnet_groups: int = 32,
|
160 |
+
resnet_pre_norm: bool = True,
|
161 |
+
attn_num_head_channels=1,
|
162 |
+
output_scale_factor=1.0,
|
163 |
+
cross_attention_dim=1280,
|
164 |
+
dual_cross_attention=False,
|
165 |
+
use_linear_projection=True,
|
166 |
+
upcast_attention=False,
|
167 |
+
):
|
168 |
+
super().__init__()
|
169 |
+
|
170 |
+
self.has_cross_attention = True
|
171 |
+
self.attn_num_head_channels = attn_num_head_channels
|
172 |
+
resnet_groups = (
|
173 |
+
resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
|
174 |
+
)
|
175 |
+
|
176 |
+
# there is always at least one resnet
|
177 |
+
resnets = [
|
178 |
+
FlaxResnetBlock2D(
|
179 |
+
in_channels=in_channels,
|
180 |
+
out_channels=in_channels,
|
181 |
+
temb_channels=temb_channels,
|
182 |
+
eps=resnet_eps,
|
183 |
+
groups=resnet_groups,
|
184 |
+
dropout=dropout,
|
185 |
+
time_embedding_norm=resnet_time_scale_shift,
|
186 |
+
non_linearity=resnet_act_fn,
|
187 |
+
output_scale_factor=output_scale_factor,
|
188 |
+
pre_norm=resnet_pre_norm,
|
189 |
+
)
|
190 |
+
]
|
191 |
+
temp_convs = [
|
192 |
+
TemporalConvLayer(
|
193 |
+
in_channels,
|
194 |
+
in_channels,
|
195 |
+
dropout=0.1,
|
196 |
+
)
|
197 |
+
]
|
198 |
+
attentions = []
|
199 |
+
temp_attentions = []
|
200 |
+
|
201 |
+
for _ in range(num_layers):
|
202 |
+
attentions.append(
|
203 |
+
Transformer2DModel(
|
204 |
+
in_channels // attn_num_head_channels,
|
205 |
+
attn_num_head_channels,
|
206 |
+
in_channels=in_channels,
|
207 |
+
num_layers=1,
|
208 |
+
cross_attention_dim=cross_attention_dim,
|
209 |
+
norm_num_groups=resnet_groups,
|
210 |
+
use_linear_projection=use_linear_projection,
|
211 |
+
upcast_attention=upcast_attention,
|
212 |
+
)
|
213 |
+
)
|
214 |
+
temp_attentions.append(
|
215 |
+
TransformerTemporalModel(
|
216 |
+
in_channels // attn_num_head_channels,
|
217 |
+
attn_num_head_channels,
|
218 |
+
in_channels=in_channels,
|
219 |
+
num_layers=1,
|
220 |
+
cross_attention_dim=cross_attention_dim,
|
221 |
+
norm_num_groups=resnet_groups,
|
222 |
+
)
|
223 |
+
)
|
224 |
+
resnets.append(
|
225 |
+
ResnetBlock2D(
|
226 |
+
in_channels=in_channels,
|
227 |
+
out_channels=in_channels,
|
228 |
+
temb_channels=temb_channels,
|
229 |
+
eps=resnet_eps,
|
230 |
+
groups=resnet_groups,
|
231 |
+
dropout=dropout,
|
232 |
+
time_embedding_norm=resnet_time_scale_shift,
|
233 |
+
non_linearity=resnet_act_fn,
|
234 |
+
output_scale_factor=output_scale_factor,
|
235 |
+
pre_norm=resnet_pre_norm,
|
236 |
+
)
|
237 |
+
)
|
238 |
+
temp_convs.append(
|
239 |
+
TemporalConvLayer(
|
240 |
+
in_channels,
|
241 |
+
in_channels,
|
242 |
+
dropout=0.1,
|
243 |
+
)
|
244 |
+
)
|
245 |
+
|
246 |
+
self.resnets = nn.ModuleList(resnets)
|
247 |
+
self.temp_convs = nn.ModuleList(temp_convs)
|
248 |
+
self.attentions = nn.ModuleList(attentions)
|
249 |
+
self.temp_attentions = nn.ModuleList(temp_attentions)
|
250 |
+
|
251 |
+
def forward(
|
252 |
+
self,
|
253 |
+
hidden_states,
|
254 |
+
temb=None,
|
255 |
+
encoder_hidden_states=None,
|
256 |
+
attention_mask=None,
|
257 |
+
num_frames=1,
|
258 |
+
cross_attention_kwargs=None,
|
259 |
+
):
|
260 |
+
hidden_states = self.resnets[0](hidden_states, temb)
|
261 |
+
hidden_states = self.temp_convs[0](hidden_states, num_frames=num_frames)
|
262 |
+
for attn, temp_attn, resnet, temp_conv in zip(
|
263 |
+
self.attentions, self.temp_attentions, self.resnets[1:], self.temp_convs[1:]
|
264 |
+
):
|
265 |
+
hidden_states = attn(
|
266 |
+
hidden_states,
|
267 |
+
encoder_hidden_states=encoder_hidden_states,
|
268 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
269 |
+
).sample
|
270 |
+
hidden_states = temp_attn(
|
271 |
+
hidden_states,
|
272 |
+
num_frames=num_frames,
|
273 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
274 |
+
).sample
|
275 |
+
hidden_states = resnet(hidden_states, temb)
|
276 |
+
hidden_states = temp_conv(hidden_states, num_frames=num_frames)
|
277 |
+
|
278 |
+
return hidden_states
|
279 |
+
|
280 |
+
|
281 |
+
class CrossAttnDownBlock3D(nn.Module):
|
282 |
+
def __init__(
|
283 |
+
self,
|
284 |
+
in_channels: int,
|
285 |
+
out_channels: int,
|
286 |
+
temb_channels: int,
|
287 |
+
dropout: float = 0.0,
|
288 |
+
num_layers: int = 1,
|
289 |
+
resnet_eps: float = 1e-6,
|
290 |
+
resnet_time_scale_shift: str = "default",
|
291 |
+
resnet_act_fn: str = "swish",
|
292 |
+
resnet_groups: int = 32,
|
293 |
+
resnet_pre_norm: bool = True,
|
294 |
+
attn_num_head_channels=1,
|
295 |
+
cross_attention_dim=1280,
|
296 |
+
output_scale_factor=1.0,
|
297 |
+
downsample_padding=1,
|
298 |
+
add_downsample=True,
|
299 |
+
dual_cross_attention=False,
|
300 |
+
use_linear_projection=False,
|
301 |
+
only_cross_attention=False,
|
302 |
+
upcast_attention=False,
|
303 |
+
):
|
304 |
+
super().__init__()
|
305 |
+
resnets = []
|
306 |
+
attentions = []
|
307 |
+
temp_attentions = []
|
308 |
+
temp_convs = []
|
309 |
+
|
310 |
+
self.has_cross_attention = True
|
311 |
+
self.attn_num_head_channels = attn_num_head_channels
|
312 |
+
|
313 |
+
for i in range(num_layers):
|
314 |
+
in_channels = in_channels if i == 0 else out_channels
|
315 |
+
resnets.append(
|
316 |
+
ResnetBlock2D(
|
317 |
+
in_channels=in_channels,
|
318 |
+
out_channels=out_channels,
|
319 |
+
temb_channels=temb_channels,
|
320 |
+
eps=resnet_eps,
|
321 |
+
groups=resnet_groups,
|
322 |
+
dropout=dropout,
|
323 |
+
time_embedding_norm=resnet_time_scale_shift,
|
324 |
+
non_linearity=resnet_act_fn,
|
325 |
+
output_scale_factor=output_scale_factor,
|
326 |
+
pre_norm=resnet_pre_norm,
|
327 |
+
)
|
328 |
+
)
|
329 |
+
temp_convs.append(
|
330 |
+
TemporalConvLayer(
|
331 |
+
out_channels,
|
332 |
+
out_channels,
|
333 |
+
dropout=0.1,
|
334 |
+
)
|
335 |
+
)
|
336 |
+
attentions.append(
|
337 |
+
Transformer2DModel(
|
338 |
+
out_channels // attn_num_head_channels,
|
339 |
+
attn_num_head_channels,
|
340 |
+
in_channels=out_channels,
|
341 |
+
num_layers=1,
|
342 |
+
cross_attention_dim=cross_attention_dim,
|
343 |
+
norm_num_groups=resnet_groups,
|
344 |
+
use_linear_projection=use_linear_projection,
|
345 |
+
only_cross_attention=only_cross_attention,
|
346 |
+
upcast_attention=upcast_attention,
|
347 |
+
)
|
348 |
+
)
|
349 |
+
temp_attentions.append(
|
350 |
+
TransformerTemporalModel(
|
351 |
+
out_channels // attn_num_head_channels,
|
352 |
+
attn_num_head_channels,
|
353 |
+
in_channels=out_channels,
|
354 |
+
num_layers=1,
|
355 |
+
cross_attention_dim=cross_attention_dim,
|
356 |
+
norm_num_groups=resnet_groups,
|
357 |
+
)
|
358 |
+
)
|
359 |
+
self.resnets = nn.ModuleList(resnets)
|
360 |
+
self.temp_convs = nn.ModuleList(temp_convs)
|
361 |
+
self.attentions = nn.ModuleList(attentions)
|
362 |
+
self.temp_attentions = nn.ModuleList(temp_attentions)
|
363 |
+
|
364 |
+
if add_downsample:
|
365 |
+
self.downsamplers = nn.ModuleList(
|
366 |
+
[
|
367 |
+
Downsample2D(
|
368 |
+
out_channels,
|
369 |
+
use_conv=True,
|
370 |
+
out_channels=out_channels,
|
371 |
+
padding=downsample_padding,
|
372 |
+
name="op",
|
373 |
+
)
|
374 |
+
]
|
375 |
+
)
|
376 |
+
else:
|
377 |
+
self.downsamplers = None
|
378 |
+
|
379 |
+
self.gradient_checkpointing = False
|
380 |
+
|
381 |
+
def forward(
|
382 |
+
self,
|
383 |
+
hidden_states,
|
384 |
+
temb=None,
|
385 |
+
encoder_hidden_states=None,
|
386 |
+
attention_mask=None,
|
387 |
+
num_frames=1,
|
388 |
+
cross_attention_kwargs=None,
|
389 |
+
):
|
390 |
+
# TODO(Patrick, William) - attention mask is not used
|
391 |
+
output_states = ()
|
392 |
+
|
393 |
+
for resnet, temp_conv, attn, temp_attn in zip(
|
394 |
+
self.resnets, self.temp_convs, self.attentions, self.temp_attentions
|
395 |
+
):
|
396 |
+
hidden_states = resnet(hidden_states, temb)
|
397 |
+
hidden_states = temp_conv(hidden_states, num_frames=num_frames)
|
398 |
+
hidden_states = attn(
|
399 |
+
hidden_states,
|
400 |
+
encoder_hidden_states=encoder_hidden_states,
|
401 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
402 |
+
).sample
|
403 |
+
hidden_states = temp_attn(
|
404 |
+
hidden_states,
|
405 |
+
num_frames=num_frames,
|
406 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
407 |
+
).sample
|
408 |
+
|
409 |
+
output_states += (hidden_states,)
|
410 |
+
|
411 |
+
if self.downsamplers is not None:
|
412 |
+
for downsampler in self.downsamplers:
|
413 |
+
hidden_states = downsampler(hidden_states)
|
414 |
+
|
415 |
+
output_states += (hidden_states,)
|
416 |
+
|
417 |
+
return hidden_states, output_states
|
418 |
+
|
419 |
+
|
420 |
+
class DownBlock3D(nn.Module):
|
421 |
+
def __init__(
|
422 |
+
self,
|
423 |
+
in_channels: int,
|
424 |
+
out_channels: int,
|
425 |
+
temb_channels: int,
|
426 |
+
dropout: float = 0.0,
|
427 |
+
num_layers: int = 1,
|
428 |
+
resnet_eps: float = 1e-6,
|
429 |
+
resnet_time_scale_shift: str = "default",
|
430 |
+
resnet_act_fn: str = "swish",
|
431 |
+
resnet_groups: int = 32,
|
432 |
+
resnet_pre_norm: bool = True,
|
433 |
+
output_scale_factor=1.0,
|
434 |
+
add_downsample=True,
|
435 |
+
downsample_padding=1,
|
436 |
+
):
|
437 |
+
super().__init__()
|
438 |
+
resnets = []
|
439 |
+
temp_convs = []
|
440 |
+
|
441 |
+
for i in range(num_layers):
|
442 |
+
in_channels = in_channels if i == 0 else out_channels
|
443 |
+
resnets.append(
|
444 |
+
ResnetBlock2D(
|
445 |
+
in_channels=in_channels,
|
446 |
+
out_channels=out_channels,
|
447 |
+
temb_channels=temb_channels,
|
448 |
+
eps=resnet_eps,
|
449 |
+
groups=resnet_groups,
|
450 |
+
dropout=dropout,
|
451 |
+
time_embedding_norm=resnet_time_scale_shift,
|
452 |
+
non_linearity=resnet_act_fn,
|
453 |
+
output_scale_factor=output_scale_factor,
|
454 |
+
pre_norm=resnet_pre_norm,
|
455 |
+
)
|
456 |
+
)
|
457 |
+
temp_convs.append(
|
458 |
+
TemporalConvLayer(
|
459 |
+
out_channels,
|
460 |
+
out_channels,
|
461 |
+
dropout=0.1,
|
462 |
+
)
|
463 |
+
)
|
464 |
+
|
465 |
+
self.resnets = nn.ModuleList(resnets)
|
466 |
+
self.temp_convs = nn.ModuleList(temp_convs)
|
467 |
+
|
468 |
+
if add_downsample:
|
469 |
+
self.downsamplers = nn.ModuleList(
|
470 |
+
[
|
471 |
+
Downsample2D(
|
472 |
+
out_channels,
|
473 |
+
use_conv=True,
|
474 |
+
out_channels=out_channels,
|
475 |
+
padding=downsample_padding,
|
476 |
+
name="op",
|
477 |
+
)
|
478 |
+
]
|
479 |
+
)
|
480 |
+
else:
|
481 |
+
self.downsamplers = None
|
482 |
+
|
483 |
+
self.gradient_checkpointing = False
|
484 |
+
|
485 |
+
def forward(self, hidden_states, temb=None, num_frames=1):
|
486 |
+
output_states = ()
|
487 |
+
|
488 |
+
for resnet, temp_conv in zip(self.resnets, self.temp_convs):
|
489 |
+
hidden_states = resnet(hidden_states, temb)
|
490 |
+
hidden_states = temp_conv(hidden_states, num_frames=num_frames)
|
491 |
+
|
492 |
+
output_states += (hidden_states,)
|
493 |
+
|
494 |
+
if self.downsamplers is not None:
|
495 |
+
for downsampler in self.downsamplers:
|
496 |
+
hidden_states = downsampler(hidden_states)
|
497 |
+
|
498 |
+
output_states += (hidden_states,)
|
499 |
+
|
500 |
+
return hidden_states, output_states
|
501 |
+
|
502 |
+
|
503 |
+
class CrossAttnUpBlock3D(nn.Module):
|
504 |
+
def __init__(
|
505 |
+
self,
|
506 |
+
in_channels: int,
|
507 |
+
out_channels: int,
|
508 |
+
prev_output_channel: int,
|
509 |
+
temb_channels: int,
|
510 |
+
dropout: float = 0.0,
|
511 |
+
num_layers: int = 1,
|
512 |
+
resnet_eps: float = 1e-6,
|
513 |
+
resnet_time_scale_shift: str = "default",
|
514 |
+
resnet_act_fn: str = "swish",
|
515 |
+
resnet_groups: int = 32,
|
516 |
+
resnet_pre_norm: bool = True,
|
517 |
+
attn_num_head_channels=1,
|
518 |
+
cross_attention_dim=1280,
|
519 |
+
output_scale_factor=1.0,
|
520 |
+
add_upsample=True,
|
521 |
+
dual_cross_attention=False,
|
522 |
+
use_linear_projection=False,
|
523 |
+
only_cross_attention=False,
|
524 |
+
upcast_attention=False,
|
525 |
+
):
|
526 |
+
super().__init__()
|
527 |
+
resnets = []
|
528 |
+
temp_convs = []
|
529 |
+
attentions = []
|
530 |
+
temp_attentions = []
|
531 |
+
|
532 |
+
self.has_cross_attention = True
|
533 |
+
self.attn_num_head_channels = attn_num_head_channels
|
534 |
+
|
535 |
+
for i in range(num_layers):
|
536 |
+
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
|
537 |
+
resnet_in_channels = prev_output_channel if i == 0 else out_channels
|
538 |
+
|
539 |
+
resnets.append(
|
540 |
+
ResnetBlock2D(
|
541 |
+
in_channels=resnet_in_channels + res_skip_channels,
|
542 |
+
out_channels=out_channels,
|
543 |
+
temb_channels=temb_channels,
|
544 |
+
eps=resnet_eps,
|
545 |
+
groups=resnet_groups,
|
546 |
+
dropout=dropout,
|
547 |
+
time_embedding_norm=resnet_time_scale_shift,
|
548 |
+
non_linearity=resnet_act_fn,
|
549 |
+
output_scale_factor=output_scale_factor,
|
550 |
+
pre_norm=resnet_pre_norm,
|
551 |
+
)
|
552 |
+
)
|
553 |
+
temp_convs.append(
|
554 |
+
TemporalConvLayer(
|
555 |
+
out_channels,
|
556 |
+
out_channels,
|
557 |
+
dropout=0.1,
|
558 |
+
)
|
559 |
+
)
|
560 |
+
attentions.append(
|
561 |
+
Transformer2DModel(
|
562 |
+
out_channels // attn_num_head_channels,
|
563 |
+
attn_num_head_channels,
|
564 |
+
in_channels=out_channels,
|
565 |
+
num_layers=1,
|
566 |
+
cross_attention_dim=cross_attention_dim,
|
567 |
+
norm_num_groups=resnet_groups,
|
568 |
+
use_linear_projection=use_linear_projection,
|
569 |
+
only_cross_attention=only_cross_attention,
|
570 |
+
upcast_attention=upcast_attention,
|
571 |
+
)
|
572 |
+
)
|
573 |
+
temp_attentions.append(
|
574 |
+
TransformerTemporalModel(
|
575 |
+
out_channels // attn_num_head_channels,
|
576 |
+
attn_num_head_channels,
|
577 |
+
in_channels=out_channels,
|
578 |
+
num_layers=1,
|
579 |
+
cross_attention_dim=cross_attention_dim,
|
580 |
+
norm_num_groups=resnet_groups,
|
581 |
+
)
|
582 |
+
)
|
583 |
+
self.resnets = nn.ModuleList(resnets)
|
584 |
+
self.temp_convs = nn.ModuleList(temp_convs)
|
585 |
+
self.attentions = nn.ModuleList(attentions)
|
586 |
+
self.temp_attentions = nn.ModuleList(temp_attentions)
|
587 |
+
|
588 |
+
if add_upsample:
|
589 |
+
self.upsamplers = nn.ModuleList(
|
590 |
+
[Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]
|
591 |
+
)
|
592 |
+
else:
|
593 |
+
self.upsamplers = None
|
594 |
+
|
595 |
+
self.gradient_checkpointing = False
|
596 |
+
|
597 |
+
def forward(
|
598 |
+
self,
|
599 |
+
hidden_states,
|
600 |
+
res_hidden_states_tuple,
|
601 |
+
temb=None,
|
602 |
+
encoder_hidden_states=None,
|
603 |
+
upsample_size=None,
|
604 |
+
attention_mask=None,
|
605 |
+
num_frames=1,
|
606 |
+
cross_attention_kwargs=None,
|
607 |
+
):
|
608 |
+
# TODO(Patrick, William) - attention mask is not used
|
609 |
+
for resnet, temp_conv, attn, temp_attn in zip(
|
610 |
+
self.resnets, self.temp_convs, self.attentions, self.temp_attentions
|
611 |
+
):
|
612 |
+
# pop res hidden states
|
613 |
+
res_hidden_states = res_hidden_states_tuple[-1]
|
614 |
+
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
|
615 |
+
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
|
616 |
+
|
617 |
+
hidden_states = resnet(hidden_states, temb)
|
618 |
+
hidden_states = temp_conv(hidden_states, num_frames=num_frames)
|
619 |
+
hidden_states = attn(
|
620 |
+
hidden_states,
|
621 |
+
encoder_hidden_states=encoder_hidden_states,
|
622 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
623 |
+
).sample
|
624 |
+
hidden_states = temp_attn(
|
625 |
+
hidden_states,
|
626 |
+
num_frames=num_frames,
|
627 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
628 |
+
).sample
|
629 |
+
|
630 |
+
if self.upsamplers is not None:
|
631 |
+
for upsampler in self.upsamplers:
|
632 |
+
hidden_states = upsampler(hidden_states, upsample_size)
|
633 |
+
|
634 |
+
return hidden_states
|
635 |
+
|
636 |
+
|
637 |
+
class UpBlock3D(nn.Module):
|
638 |
+
def __init__(
|
639 |
+
self,
|
640 |
+
in_channels: int,
|
641 |
+
prev_output_channel: int,
|
642 |
+
out_channels: int,
|
643 |
+
temb_channels: int,
|
644 |
+
dropout: float = 0.0,
|
645 |
+
num_layers: int = 1,
|
646 |
+
resnet_eps: float = 1e-6,
|
647 |
+
resnet_time_scale_shift: str = "default",
|
648 |
+
resnet_act_fn: str = "swish",
|
649 |
+
resnet_groups: int = 32,
|
650 |
+
resnet_pre_norm: bool = True,
|
651 |
+
output_scale_factor=1.0,
|
652 |
+
add_upsample=True,
|
653 |
+
):
|
654 |
+
super().__init__()
|
655 |
+
resnets = []
|
656 |
+
temp_convs = []
|
657 |
+
|
658 |
+
for i in range(num_layers):
|
659 |
+
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
|
660 |
+
resnet_in_channels = prev_output_channel if i == 0 else out_channels
|
661 |
+
|
662 |
+
resnets.append(
|
663 |
+
ResnetBlock2D(
|
664 |
+
in_channels=resnet_in_channels + res_skip_channels,
|
665 |
+
out_channels=out_channels,
|
666 |
+
temb_channels=temb_channels,
|
667 |
+
eps=resnet_eps,
|
668 |
+
groups=resnet_groups,
|
669 |
+
dropout=dropout,
|
670 |
+
time_embedding_norm=resnet_time_scale_shift,
|
671 |
+
non_linearity=resnet_act_fn,
|
672 |
+
output_scale_factor=output_scale_factor,
|
673 |
+
pre_norm=resnet_pre_norm,
|
674 |
+
)
|
675 |
+
)
|
676 |
+
temp_convs.append(
|
677 |
+
TemporalConvLayer(
|
678 |
+
out_channels,
|
679 |
+
out_channels,
|
680 |
+
dropout=0.1,
|
681 |
+
)
|
682 |
+
)
|
683 |
+
|
684 |
+
self.resnets = nn.ModuleList(resnets)
|
685 |
+
self.temp_convs = nn.ModuleList(temp_convs)
|
686 |
+
|
687 |
+
if add_upsample:
|
688 |
+
self.upsamplers = nn.ModuleList(
|
689 |
+
[Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]
|
690 |
+
)
|
691 |
+
else:
|
692 |
+
self.upsamplers = None
|
693 |
+
|
694 |
+
self.gradient_checkpointing = False
|
695 |
+
|
696 |
+
def forward(
|
697 |
+
self,
|
698 |
+
hidden_states,
|
699 |
+
res_hidden_states_tuple,
|
700 |
+
temb=None,
|
701 |
+
upsample_size=None,
|
702 |
+
num_frames=1,
|
703 |
+
):
|
704 |
+
for resnet, temp_conv in zip(self.resnets, self.temp_convs):
|
705 |
+
# pop res hidden states
|
706 |
+
res_hidden_states = res_hidden_states_tuple[-1]
|
707 |
+
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
|
708 |
+
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
|
709 |
+
|
710 |
+
hidden_states = resnet(hidden_states, temb)
|
711 |
+
hidden_states = temp_conv(hidden_states, num_frames=num_frames)
|
712 |
+
|
713 |
+
if self.upsamplers is not None:
|
714 |
+
for upsampler in self.upsamplers:
|
715 |
+
hidden_states = upsampler(hidden_states, upsample_size)
|
716 |
+
|
717 |
+
return hidden_states
|
text_to_animation/models/unet_3d_condition_flax.py
ADDED
@@ -0,0 +1,611 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright 2023 Alibaba DAMO-VILAB and The HuggingFace Team. All rights reserved.
|
2 |
+
# Copyright 2023 The ModelScope Team.
|
3 |
+
#
|
4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
5 |
+
# you may not use this file except in compliance with the License.
|
6 |
+
# You may obtain a copy of the License at
|
7 |
+
#
|
8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
9 |
+
#
|
10 |
+
# Unless required by applicable law or agreed to in writing, software
|
11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 |
+
# See the License for the specific language governing permissions and
|
14 |
+
# limitations under the License.
|
15 |
+
from dataclasses import dataclass
|
16 |
+
from typing import Any, Dict, List, Optional, Tuple, Union
|
17 |
+
|
18 |
+
import torch
|
19 |
+
import torch.nn as nn
|
20 |
+
import torch.utils.checkpoint
|
21 |
+
|
22 |
+
from ..configuration_utils import ConfigMixin, register_to_config
|
23 |
+
from ..loaders import UNet2DConditionLoadersMixin
|
24 |
+
from ..utils import BaseOutput, logging
|
25 |
+
from .attention_processor import AttentionProcessor, AttnProcessor
|
26 |
+
from .embeddings import TimestepEmbedding, Timesteps
|
27 |
+
from .modeling_utils import ModelMixin
|
28 |
+
from .transformer_temporal import TransformerTemporalModel
|
29 |
+
from .unet_3d_blocks import (
|
30 |
+
CrossAttnDownBlock3D,
|
31 |
+
CrossAttnUpBlock3D,
|
32 |
+
DownBlock3D,
|
33 |
+
UNetMidBlock3DCrossAttn,
|
34 |
+
UpBlock3D,
|
35 |
+
get_down_block,
|
36 |
+
get_up_block,
|
37 |
+
)
|
38 |
+
|
39 |
+
|
40 |
+
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
41 |
+
|
42 |
+
|
43 |
+
@dataclass
|
44 |
+
class UNet3DConditionOutput(BaseOutput):
|
45 |
+
"""
|
46 |
+
Args:
|
47 |
+
sample (`torch.FloatTensor` of shape `(batch_size, num_frames, num_channels, height, width)`):
|
48 |
+
Hidden states conditioned on `encoder_hidden_states` input. Output of last layer of model.
|
49 |
+
"""
|
50 |
+
|
51 |
+
sample: torch.FloatTensor
|
52 |
+
|
53 |
+
|
54 |
+
class UNet3DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
|
55 |
+
r"""
|
56 |
+
UNet3DConditionModel is a conditional 2D UNet model that takes in a noisy sample, conditional state, and a timestep
|
57 |
+
and returns sample shaped output.
|
58 |
+
|
59 |
+
This model inherits from [`ModelMixin`]. Check the superclass documentation for the generic methods the library
|
60 |
+
implements for all the models (such as downloading or saving, etc.)
|
61 |
+
|
62 |
+
Parameters:
|
63 |
+
sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
|
64 |
+
Height and width of input/output sample.
|
65 |
+
in_channels (`int`, *optional*, defaults to 4): The number of channels in the input sample.
|
66 |
+
out_channels (`int`, *optional*, defaults to 4): The number of channels in the output.
|
67 |
+
down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
|
68 |
+
The tuple of downsample blocks to use.
|
69 |
+
up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D",)`):
|
70 |
+
The tuple of upsample blocks to use.
|
71 |
+
block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
|
72 |
+
The tuple of output channels for each block.
|
73 |
+
layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
|
74 |
+
downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
|
75 |
+
mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
|
76 |
+
act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
|
77 |
+
norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
|
78 |
+
If `None`, it will skip the normalization and activation layers in post-processing
|
79 |
+
norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
|
80 |
+
cross_attention_dim (`int`, *optional*, defaults to 1280): The dimension of the cross attention features.
|
81 |
+
attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
|
82 |
+
"""
|
83 |
+
|
84 |
+
_supports_gradient_checkpointing = False
|
85 |
+
|
86 |
+
@register_to_config
|
87 |
+
def __init__(
|
88 |
+
self,
|
89 |
+
sample_size: Optional[int] = None,
|
90 |
+
in_channels: int = 4,
|
91 |
+
out_channels: int = 4,
|
92 |
+
down_block_types: Tuple[str] = (
|
93 |
+
"CrossAttnDownBlock3D",
|
94 |
+
"CrossAttnDownBlock3D",
|
95 |
+
"CrossAttnDownBlock3D",
|
96 |
+
"DownBlock3D",
|
97 |
+
),
|
98 |
+
up_block_types: Tuple[str] = (
|
99 |
+
"UpBlock3D",
|
100 |
+
"CrossAttnUpBlock3D",
|
101 |
+
"CrossAttnUpBlock3D",
|
102 |
+
"CrossAttnUpBlock3D",
|
103 |
+
),
|
104 |
+
block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
|
105 |
+
layers_per_block: int = 2,
|
106 |
+
downsample_padding: int = 1,
|
107 |
+
mid_block_scale_factor: float = 1,
|
108 |
+
act_fn: str = "silu",
|
109 |
+
norm_num_groups: Optional[int] = 32,
|
110 |
+
norm_eps: float = 1e-5,
|
111 |
+
cross_attention_dim: int = 1024,
|
112 |
+
attention_head_dim: Union[int, Tuple[int]] = 64,
|
113 |
+
):
|
114 |
+
super().__init__()
|
115 |
+
|
116 |
+
self.sample_size = sample_size
|
117 |
+
|
118 |
+
# Check inputs
|
119 |
+
if len(down_block_types) != len(up_block_types):
|
120 |
+
raise ValueError(
|
121 |
+
f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
|
122 |
+
)
|
123 |
+
|
124 |
+
if len(block_out_channels) != len(down_block_types):
|
125 |
+
raise ValueError(
|
126 |
+
f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
|
127 |
+
)
|
128 |
+
|
129 |
+
if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(
|
130 |
+
down_block_types
|
131 |
+
):
|
132 |
+
raise ValueError(
|
133 |
+
f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
|
134 |
+
)
|
135 |
+
|
136 |
+
# input
|
137 |
+
conv_in_kernel = 3
|
138 |
+
conv_out_kernel = 3
|
139 |
+
conv_in_padding = (conv_in_kernel - 1) // 2
|
140 |
+
self.conv_in = nn.Conv2d(
|
141 |
+
in_channels,
|
142 |
+
block_out_channels[0],
|
143 |
+
kernel_size=conv_in_kernel,
|
144 |
+
padding=conv_in_padding,
|
145 |
+
)
|
146 |
+
|
147 |
+
# time
|
148 |
+
time_embed_dim = block_out_channels[0] * 4
|
149 |
+
self.time_proj = Timesteps(block_out_channels[0], True, 0)
|
150 |
+
timestep_input_dim = block_out_channels[0]
|
151 |
+
|
152 |
+
self.time_embedding = TimestepEmbedding(
|
153 |
+
timestep_input_dim,
|
154 |
+
time_embed_dim,
|
155 |
+
act_fn=act_fn,
|
156 |
+
)
|
157 |
+
|
158 |
+
self.transformer_in = TransformerTemporalModel(
|
159 |
+
num_attention_heads=8,
|
160 |
+
attention_head_dim=attention_head_dim,
|
161 |
+
in_channels=block_out_channels[0],
|
162 |
+
num_layers=1,
|
163 |
+
)
|
164 |
+
|
165 |
+
# class embedding
|
166 |
+
self.down_blocks = nn.ModuleList([])
|
167 |
+
self.up_blocks = nn.ModuleList([])
|
168 |
+
|
169 |
+
if isinstance(attention_head_dim, int):
|
170 |
+
attention_head_dim = (attention_head_dim,) * len(down_block_types)
|
171 |
+
|
172 |
+
# down
|
173 |
+
output_channel = block_out_channels[0]
|
174 |
+
for i, down_block_type in enumerate(down_block_types):
|
175 |
+
input_channel = output_channel
|
176 |
+
output_channel = block_out_channels[i]
|
177 |
+
is_final_block = i == len(block_out_channels) - 1
|
178 |
+
|
179 |
+
down_block = get_down_block(
|
180 |
+
down_block_type,
|
181 |
+
num_layers=layers_per_block,
|
182 |
+
in_channels=input_channel,
|
183 |
+
out_channels=output_channel,
|
184 |
+
temb_channels=time_embed_dim,
|
185 |
+
add_downsample=not is_final_block,
|
186 |
+
resnet_eps=norm_eps,
|
187 |
+
resnet_act_fn=act_fn,
|
188 |
+
resnet_groups=norm_num_groups,
|
189 |
+
cross_attention_dim=cross_attention_dim,
|
190 |
+
attn_num_head_channels=attention_head_dim[i],
|
191 |
+
downsample_padding=downsample_padding,
|
192 |
+
dual_cross_attention=False,
|
193 |
+
)
|
194 |
+
self.down_blocks.append(down_block)
|
195 |
+
|
196 |
+
# mid
|
197 |
+
self.mid_block = UNetMidBlock3DCrossAttn(
|
198 |
+
in_channels=block_out_channels[-1],
|
199 |
+
temb_channels=time_embed_dim,
|
200 |
+
resnet_eps=norm_eps,
|
201 |
+
resnet_act_fn=act_fn,
|
202 |
+
output_scale_factor=mid_block_scale_factor,
|
203 |
+
cross_attention_dim=cross_attention_dim,
|
204 |
+
attn_num_head_channels=attention_head_dim[-1],
|
205 |
+
resnet_groups=norm_num_groups,
|
206 |
+
dual_cross_attention=False,
|
207 |
+
)
|
208 |
+
|
209 |
+
# count how many layers upsample the images
|
210 |
+
self.num_upsamplers = 0
|
211 |
+
|
212 |
+
# up
|
213 |
+
reversed_block_out_channels = list(reversed(block_out_channels))
|
214 |
+
reversed_attention_head_dim = list(reversed(attention_head_dim))
|
215 |
+
|
216 |
+
output_channel = reversed_block_out_channels[0]
|
217 |
+
for i, up_block_type in enumerate(up_block_types):
|
218 |
+
is_final_block = i == len(block_out_channels) - 1
|
219 |
+
|
220 |
+
prev_output_channel = output_channel
|
221 |
+
output_channel = reversed_block_out_channels[i]
|
222 |
+
input_channel = reversed_block_out_channels[
|
223 |
+
min(i + 1, len(block_out_channels) - 1)
|
224 |
+
]
|
225 |
+
|
226 |
+
# add upsample block for all BUT final layer
|
227 |
+
if not is_final_block:
|
228 |
+
add_upsample = True
|
229 |
+
self.num_upsamplers += 1
|
230 |
+
else:
|
231 |
+
add_upsample = False
|
232 |
+
|
233 |
+
up_block = get_up_block(
|
234 |
+
up_block_type,
|
235 |
+
num_layers=layers_per_block + 1,
|
236 |
+
in_channels=input_channel,
|
237 |
+
out_channels=output_channel,
|
238 |
+
prev_output_channel=prev_output_channel,
|
239 |
+
temb_channels=time_embed_dim,
|
240 |
+
add_upsample=add_upsample,
|
241 |
+
resnet_eps=norm_eps,
|
242 |
+
resnet_act_fn=act_fn,
|
243 |
+
resnet_groups=norm_num_groups,
|
244 |
+
cross_attention_dim=cross_attention_dim,
|
245 |
+
attn_num_head_channels=reversed_attention_head_dim[i],
|
246 |
+
dual_cross_attention=False,
|
247 |
+
)
|
248 |
+
self.up_blocks.append(up_block)
|
249 |
+
prev_output_channel = output_channel
|
250 |
+
|
251 |
+
# out
|
252 |
+
if norm_num_groups is not None:
|
253 |
+
self.conv_norm_out = nn.GroupNorm(
|
254 |
+
num_channels=block_out_channels[0],
|
255 |
+
num_groups=norm_num_groups,
|
256 |
+
eps=norm_eps,
|
257 |
+
)
|
258 |
+
self.conv_act = nn.SiLU()
|
259 |
+
else:
|
260 |
+
self.conv_norm_out = None
|
261 |
+
self.conv_act = None
|
262 |
+
|
263 |
+
conv_out_padding = (conv_out_kernel - 1) // 2
|
264 |
+
self.conv_out = nn.Conv2d(
|
265 |
+
block_out_channels[0],
|
266 |
+
out_channels,
|
267 |
+
kernel_size=conv_out_kernel,
|
268 |
+
padding=conv_out_padding,
|
269 |
+
)
|
270 |
+
|
271 |
+
@property
|
272 |
+
# Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.attn_processors
|
273 |
+
def attn_processors(self) -> Dict[str, AttentionProcessor]:
|
274 |
+
r"""
|
275 |
+
Returns:
|
276 |
+
`dict` of attention processors: A dictionary containing all attention processors used in the model with
|
277 |
+
indexed by its weight name.
|
278 |
+
"""
|
279 |
+
# set recursively
|
280 |
+
processors = {}
|
281 |
+
|
282 |
+
def fn_recursive_add_processors(
|
283 |
+
name: str,
|
284 |
+
module: torch.nn.Module,
|
285 |
+
processors: Dict[str, AttentionProcessor],
|
286 |
+
):
|
287 |
+
if hasattr(module, "set_processor"):
|
288 |
+
processors[f"{name}.processor"] = module.processor
|
289 |
+
|
290 |
+
for sub_name, child in module.named_children():
|
291 |
+
fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
|
292 |
+
|
293 |
+
return processors
|
294 |
+
|
295 |
+
for name, module in self.named_children():
|
296 |
+
fn_recursive_add_processors(name, module, processors)
|
297 |
+
|
298 |
+
return processors
|
299 |
+
|
300 |
+
# Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attention_slice
|
301 |
+
def set_attention_slice(self, slice_size):
|
302 |
+
r"""
|
303 |
+
Enable sliced attention computation.
|
304 |
+
|
305 |
+
When this option is enabled, the attention module will split the input tensor in slices, to compute attention
|
306 |
+
in several steps. This is useful to save some memory in exchange for a small speed decrease.
|
307 |
+
|
308 |
+
Args:
|
309 |
+
slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
|
310 |
+
When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
|
311 |
+
`"max"`, maximum amount of memory will be saved by running only one slice at a time. If a number is
|
312 |
+
provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
|
313 |
+
must be a multiple of `slice_size`.
|
314 |
+
"""
|
315 |
+
sliceable_head_dims = []
|
316 |
+
|
317 |
+
def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
|
318 |
+
if hasattr(module, "set_attention_slice"):
|
319 |
+
sliceable_head_dims.append(module.sliceable_head_dim)
|
320 |
+
|
321 |
+
for child in module.children():
|
322 |
+
fn_recursive_retrieve_sliceable_dims(child)
|
323 |
+
|
324 |
+
# retrieve number of attention layers
|
325 |
+
for module in self.children():
|
326 |
+
fn_recursive_retrieve_sliceable_dims(module)
|
327 |
+
|
328 |
+
num_sliceable_layers = len(sliceable_head_dims)
|
329 |
+
|
330 |
+
if slice_size == "auto":
|
331 |
+
# half the attention head size is usually a good trade-off between
|
332 |
+
# speed and memory
|
333 |
+
slice_size = [dim // 2 for dim in sliceable_head_dims]
|
334 |
+
elif slice_size == "max":
|
335 |
+
# make smallest slice possible
|
336 |
+
slice_size = num_sliceable_layers * [1]
|
337 |
+
|
338 |
+
slice_size = (
|
339 |
+
num_sliceable_layers * [slice_size]
|
340 |
+
if not isinstance(slice_size, list)
|
341 |
+
else slice_size
|
342 |
+
)
|
343 |
+
|
344 |
+
if len(slice_size) != len(sliceable_head_dims):
|
345 |
+
raise ValueError(
|
346 |
+
f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
|
347 |
+
f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
|
348 |
+
)
|
349 |
+
|
350 |
+
for i in range(len(slice_size)):
|
351 |
+
size = slice_size[i]
|
352 |
+
dim = sliceable_head_dims[i]
|
353 |
+
if size is not None and size > dim:
|
354 |
+
raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
|
355 |
+
|
356 |
+
# Recursively walk through all the children.
|
357 |
+
# Any children which exposes the set_attention_slice method
|
358 |
+
# gets the message
|
359 |
+
def fn_recursive_set_attention_slice(
|
360 |
+
module: torch.nn.Module, slice_size: List[int]
|
361 |
+
):
|
362 |
+
if hasattr(module, "set_attention_slice"):
|
363 |
+
module.set_attention_slice(slice_size.pop())
|
364 |
+
|
365 |
+
for child in module.children():
|
366 |
+
fn_recursive_set_attention_slice(child, slice_size)
|
367 |
+
|
368 |
+
reversed_slice_size = list(reversed(slice_size))
|
369 |
+
for module in self.children():
|
370 |
+
fn_recursive_set_attention_slice(module, reversed_slice_size)
|
371 |
+
|
372 |
+
# Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attn_processor
|
373 |
+
def set_attn_processor(
|
374 |
+
self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]
|
375 |
+
):
|
376 |
+
r"""
|
377 |
+
Parameters:
|
378 |
+
`processor (`dict` of `AttentionProcessor` or `AttentionProcessor`):
|
379 |
+
The instantiated processor class or a dictionary of processor classes that will be set as the processor
|
380 |
+
of **all** `Attention` layers.
|
381 |
+
In case `processor` is a dict, the key needs to define the path to the corresponding cross attention processor. This is strongly recommended when setting trainable attention processors.:
|
382 |
+
|
383 |
+
"""
|
384 |
+
count = len(self.attn_processors.keys())
|
385 |
+
|
386 |
+
if isinstance(processor, dict) and len(processor) != count:
|
387 |
+
raise ValueError(
|
388 |
+
f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
|
389 |
+
f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
|
390 |
+
)
|
391 |
+
|
392 |
+
def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
|
393 |
+
if hasattr(module, "set_processor"):
|
394 |
+
if not isinstance(processor, dict):
|
395 |
+
module.set_processor(processor)
|
396 |
+
else:
|
397 |
+
module.set_processor(processor.pop(f"{name}.processor"))
|
398 |
+
|
399 |
+
for sub_name, child in module.named_children():
|
400 |
+
fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
|
401 |
+
|
402 |
+
for name, module in self.named_children():
|
403 |
+
fn_recursive_attn_processor(name, module, processor)
|
404 |
+
|
405 |
+
# Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
|
406 |
+
def set_default_attn_processor(self):
|
407 |
+
"""
|
408 |
+
Disables custom attention processors and sets the default attention implementation.
|
409 |
+
"""
|
410 |
+
self.set_attn_processor(AttnProcessor())
|
411 |
+
|
412 |
+
def _set_gradient_checkpointing(self, module, value=False):
|
413 |
+
if isinstance(
|
414 |
+
module, (CrossAttnDownBlock3D, DownBlock3D, CrossAttnUpBlock3D, UpBlock3D)
|
415 |
+
):
|
416 |
+
module.gradient_checkpointing = value
|
417 |
+
|
418 |
+
def forward(
|
419 |
+
self,
|
420 |
+
sample: torch.FloatTensor,
|
421 |
+
timestep: Union[torch.Tensor, float, int],
|
422 |
+
encoder_hidden_states: torch.Tensor,
|
423 |
+
class_labels: Optional[torch.Tensor] = None,
|
424 |
+
timestep_cond: Optional[torch.Tensor] = None,
|
425 |
+
attention_mask: Optional[torch.Tensor] = None,
|
426 |
+
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
|
427 |
+
down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
|
428 |
+
mid_block_additional_residual: Optional[torch.Tensor] = None,
|
429 |
+
return_dict: bool = True,
|
430 |
+
) -> Union[UNet3DConditionOutput, Tuple]:
|
431 |
+
r"""
|
432 |
+
Args:
|
433 |
+
sample (`torch.FloatTensor`): (batch, num_frames, channel, height, width) noisy inputs tensor
|
434 |
+
timestep (`torch.FloatTensor` or `float` or `int`): (batch) timesteps
|
435 |
+
encoder_hidden_states (`torch.FloatTensor`): (batch, sequence_length, feature_dim) encoder hidden states
|
436 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
437 |
+
Whether or not to return a [`models.unet_2d_condition.UNet3DConditionOutput`] instead of a plain tuple.
|
438 |
+
cross_attention_kwargs (`dict`, *optional*):
|
439 |
+
A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
|
440 |
+
`self.processor` in
|
441 |
+
[diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
|
442 |
+
|
443 |
+
Returns:
|
444 |
+
[`~models.unet_2d_condition.UNet3DConditionOutput`] or `tuple`:
|
445 |
+
[`~models.unet_2d_condition.UNet3DConditionOutput`] if `return_dict` is True, otherwise a `tuple`. When
|
446 |
+
returning a tuple, the first element is the sample tensor.
|
447 |
+
"""
|
448 |
+
# By default samples have to be AT least a multiple of the overall upsampling factor.
|
449 |
+
# The overall upsampling factor is equal to 2 ** (# num of upsampling layears).
|
450 |
+
# However, the upsampling interpolation output size can be forced to fit any upsampling size
|
451 |
+
# on the fly if necessary.
|
452 |
+
default_overall_up_factor = 2**self.num_upsamplers
|
453 |
+
|
454 |
+
# upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
|
455 |
+
forward_upsample_size = False
|
456 |
+
upsample_size = None
|
457 |
+
|
458 |
+
if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
|
459 |
+
logger.info("Forward upsample size to force interpolation output size.")
|
460 |
+
forward_upsample_size = True
|
461 |
+
|
462 |
+
# prepare attention_mask
|
463 |
+
if attention_mask is not None:
|
464 |
+
attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
|
465 |
+
attention_mask = attention_mask.unsqueeze(1)
|
466 |
+
|
467 |
+
# 1. time
|
468 |
+
timesteps = timestep
|
469 |
+
if not torch.is_tensor(timesteps):
|
470 |
+
# TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
|
471 |
+
# This would be a good case for the `match` statement (Python 3.10+)
|
472 |
+
is_mps = sample.device.type == "mps"
|
473 |
+
if isinstance(timestep, float):
|
474 |
+
dtype = torch.float32 if is_mps else torch.float64
|
475 |
+
else:
|
476 |
+
dtype = torch.int32 if is_mps else torch.int64
|
477 |
+
timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
|
478 |
+
elif len(timesteps.shape) == 0:
|
479 |
+
timesteps = timesteps[None].to(sample.device)
|
480 |
+
|
481 |
+
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
|
482 |
+
num_frames = sample.shape[2]
|
483 |
+
timesteps = timesteps.expand(sample.shape[0])
|
484 |
+
|
485 |
+
t_emb = self.time_proj(timesteps)
|
486 |
+
|
487 |
+
# timesteps does not contain any weights and will always return f32 tensors
|
488 |
+
# but time_embedding might actually be running in fp16. so we need to cast here.
|
489 |
+
# there might be better ways to encapsulate this.
|
490 |
+
t_emb = t_emb.to(dtype=self.dtype)
|
491 |
+
|
492 |
+
emb = self.time_embedding(t_emb, timestep_cond)
|
493 |
+
emb = emb.repeat_interleave(repeats=num_frames, dim=0)
|
494 |
+
encoder_hidden_states = encoder_hidden_states.repeat_interleave(
|
495 |
+
repeats=num_frames, dim=0
|
496 |
+
)
|
497 |
+
|
498 |
+
# 2. pre-process
|
499 |
+
sample = sample.permute(0, 2, 1, 3, 4).reshape(
|
500 |
+
(sample.shape[0] * num_frames, -1) + sample.shape[3:]
|
501 |
+
)
|
502 |
+
sample = self.conv_in(sample)
|
503 |
+
|
504 |
+
sample = self.transformer_in(
|
505 |
+
sample, num_frames=num_frames, cross_attention_kwargs=cross_attention_kwargs
|
506 |
+
).sample
|
507 |
+
|
508 |
+
# 3. down
|
509 |
+
down_block_res_samples = (sample,)
|
510 |
+
for downsample_block in self.down_blocks:
|
511 |
+
if (
|
512 |
+
hasattr(downsample_block, "has_cross_attention")
|
513 |
+
and downsample_block.has_cross_attention
|
514 |
+
):
|
515 |
+
sample, res_samples = downsample_block(
|
516 |
+
hidden_states=sample,
|
517 |
+
temb=emb,
|
518 |
+
encoder_hidden_states=encoder_hidden_states,
|
519 |
+
attention_mask=attention_mask,
|
520 |
+
num_frames=num_frames,
|
521 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
522 |
+
)
|
523 |
+
else:
|
524 |
+
sample, res_samples = downsample_block(
|
525 |
+
hidden_states=sample, temb=emb, num_frames=num_frames
|
526 |
+
)
|
527 |
+
|
528 |
+
down_block_res_samples += res_samples
|
529 |
+
|
530 |
+
if down_block_additional_residuals is not None:
|
531 |
+
new_down_block_res_samples = ()
|
532 |
+
|
533 |
+
for down_block_res_sample, down_block_additional_residual in zip(
|
534 |
+
down_block_res_samples, down_block_additional_residuals
|
535 |
+
):
|
536 |
+
down_block_res_sample = (
|
537 |
+
down_block_res_sample + down_block_additional_residual
|
538 |
+
)
|
539 |
+
new_down_block_res_samples += (down_block_res_sample,)
|
540 |
+
|
541 |
+
down_block_res_samples = new_down_block_res_samples
|
542 |
+
|
543 |
+
# 4. mid
|
544 |
+
if self.mid_block is not None:
|
545 |
+
sample = self.mid_block(
|
546 |
+
sample,
|
547 |
+
emb,
|
548 |
+
encoder_hidden_states=encoder_hidden_states,
|
549 |
+
attention_mask=attention_mask,
|
550 |
+
num_frames=num_frames,
|
551 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
552 |
+
)
|
553 |
+
|
554 |
+
if mid_block_additional_residual is not None:
|
555 |
+
sample = sample + mid_block_additional_residual
|
556 |
+
|
557 |
+
# 5. up
|
558 |
+
for i, upsample_block in enumerate(self.up_blocks):
|
559 |
+
is_final_block = i == len(self.up_blocks) - 1
|
560 |
+
|
561 |
+
res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
|
562 |
+
down_block_res_samples = down_block_res_samples[
|
563 |
+
: -len(upsample_block.resnets)
|
564 |
+
]
|
565 |
+
|
566 |
+
# if we have not reached the final block and need to forward the
|
567 |
+
# upsample size, we do it here
|
568 |
+
if not is_final_block and forward_upsample_size:
|
569 |
+
upsample_size = down_block_res_samples[-1].shape[2:]
|
570 |
+
|
571 |
+
if (
|
572 |
+
hasattr(upsample_block, "has_cross_attention")
|
573 |
+
and upsample_block.has_cross_attention
|
574 |
+
):
|
575 |
+
sample = upsample_block(
|
576 |
+
hidden_states=sample,
|
577 |
+
temb=emb,
|
578 |
+
res_hidden_states_tuple=res_samples,
|
579 |
+
encoder_hidden_states=encoder_hidden_states,
|
580 |
+
upsample_size=upsample_size,
|
581 |
+
attention_mask=attention_mask,
|
582 |
+
num_frames=num_frames,
|
583 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
584 |
+
)
|
585 |
+
else:
|
586 |
+
sample = upsample_block(
|
587 |
+
hidden_states=sample,
|
588 |
+
temb=emb,
|
589 |
+
res_hidden_states_tuple=res_samples,
|
590 |
+
upsample_size=upsample_size,
|
591 |
+
num_frames=num_frames,
|
592 |
+
)
|
593 |
+
|
594 |
+
# 6. post-process
|
595 |
+
if self.conv_norm_out:
|
596 |
+
sample = self.conv_norm_out(sample)
|
597 |
+
sample = self.conv_act(sample)
|
598 |
+
|
599 |
+
sample = self.conv_out(sample)
|
600 |
+
|
601 |
+
# reshape to (batch, channel, framerate, width, height)
|
602 |
+
sample = (
|
603 |
+
sample[None, :]
|
604 |
+
.reshape((-1, num_frames) + sample.shape[1:])
|
605 |
+
.permute(0, 2, 1, 3, 4)
|
606 |
+
)
|
607 |
+
|
608 |
+
if not return_dict:
|
609 |
+
return (sample,)
|
610 |
+
|
611 |
+
return UNet3DConditionOutput(sample=sample)
|
text_to_animation/pipelines/text_to_video_pipeline_flax.py
CHANGED
@@ -6,11 +6,16 @@ import jax.numpy as jnp
|
|
6 |
import numpy as np
|
7 |
from flax.core.frozen_dict import FrozenDict
|
8 |
from flax.jax_utils import unreplicate
|
|
|
9 |
from flax.training.common_utils import shard
|
10 |
from PIL import Image
|
11 |
from transformers import CLIPFeatureExtractor, CLIPTokenizer, FlaxCLIPTextModel
|
12 |
from einops import rearrange, repeat
|
13 |
-
from diffusers.models import
|
|
|
|
|
|
|
|
|
14 |
from diffusers.schedulers import (
|
15 |
FlaxDDIMScheduler,
|
16 |
FlaxDPMSolverMultistepScheduler,
|
@@ -20,17 +25,24 @@ from diffusers.schedulers import (
|
|
20 |
from diffusers.utils import PIL_INTERPOLATION, logging, replace_example_docstring
|
21 |
from diffusers.pipelines.pipeline_flax_utils import FlaxDiffusionPipeline
|
22 |
from diffusers.pipelines.stable_diffusion import FlaxStableDiffusionPipelineOutput
|
23 |
-
from diffusers.pipelines.stable_diffusion.safety_checker_flax import
|
|
|
|
|
|
|
24 |
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
25 |
"""
|
26 |
Text2Video-Zero:
|
27 |
- Inputs: Prompt, Pose Control via mp4/gif, First Frame (?)
|
28 |
- JAX implementation
|
29 |
- 3DUnet to replace 2DUnetConditional
|
30 |
-
|
31 |
"""
|
32 |
|
33 |
-
|
|
|
|
|
|
|
|
|
|
|
34 |
|
35 |
EXAMPLE_DOC_STRING = """
|
36 |
Examples:
|
@@ -89,16 +101,22 @@ EXAMPLE_DOC_STRING = """
|
|
89 |
>>> output_images.save("generated_image.png")
|
90 |
```
|
91 |
"""
|
|
|
|
|
92 |
class FlaxTextToVideoPipeline(FlaxDiffusionPipeline):
|
93 |
def __init__(
|
94 |
self,
|
95 |
-
vae
|
96 |
-
text_encoder
|
97 |
-
tokenizer
|
98 |
-
unet
|
99 |
-
|
|
|
100 |
scheduler: Union[
|
101 |
-
FlaxDDIMScheduler,
|
|
|
|
|
|
|
102 |
],
|
103 |
safety_checker: FlaxStableDiffusionSafetyChecker,
|
104 |
feature_extractor: CLIPFeatureExtractor,
|
@@ -122,6 +140,7 @@ class FlaxTextToVideoPipeline(FlaxDiffusionPipeline):
|
|
122 |
text_encoder=text_encoder,
|
123 |
tokenizer=tokenizer,
|
124 |
unet=unet,
|
|
|
125 |
controlnet=controlnet,
|
126 |
scheduler=scheduler,
|
127 |
safety_checker=safety_checker,
|
@@ -135,30 +154,50 @@ class FlaxTextToVideoPipeline(FlaxDiffusionPipeline):
|
|
135 |
else:
|
136 |
eps = jax.random.normal(prng, x0.shape, dtype=text_embeddings.dtype)
|
137 |
alpha_vec = jnp.prod(params["scheduler"].common.alphas[t0:tMax])
|
138 |
-
xt = jnp.sqrt(alpha_vec) * x0 +
|
139 |
-
jnp.sqrt(1-alpha_vec) * eps
|
140 |
return xt
|
141 |
-
|
142 |
-
def DDIM_backward(
|
143 |
-
|
144 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
145 |
f = latents_local.shape[2]
|
146 |
-
latents_local = rearrange(latents_local, "b c f w
|
147 |
latents = latents_local.copy()
|
148 |
x_t0_1 = None
|
149 |
x_t1_1 = None
|
150 |
-
max_timestep = len(timesteps)-1
|
151 |
timesteps = jnp.array(timesteps)
|
|
|
152 |
def while_body(args):
|
153 |
step, latents, x_t0_1, x_t1_1, scheduler_state = args
|
154 |
t = jnp.array(scheduler_state.timesteps, dtype=jnp.int32)[step]
|
155 |
-
latent_model_input =
|
156 |
-
[latents] * 2)
|
|
|
|
|
|
|
157 |
latent_model_input = self.scheduler.scale_model_input(
|
158 |
scheduler_state, latent_model_input, timestep=t
|
159 |
)
|
160 |
f = latents.shape[0]
|
161 |
-
te = jnp.stack(
|
|
|
|
|
162 |
timestep = jnp.broadcast_to(t, latent_model_input.shape[0])
|
163 |
if controlnet_image is not None:
|
164 |
down_block_res_samples, mid_block_res_sample = self.controlnet.apply(
|
@@ -185,41 +224,53 @@ class FlaxTextToVideoPipeline(FlaxDiffusionPipeline):
|
|
185 |
jnp.array(latent_model_input),
|
186 |
jnp.array(timestep, dtype=jnp.int32),
|
187 |
encoder_hidden_states=te,
|
188 |
-
|
189 |
# perform guidance
|
190 |
if do_classifier_free_guidance:
|
191 |
noise_pred_uncond, noise_pred_text = jnp.split(noise_pred, 2, axis=0)
|
192 |
-
noise_pred = noise_pred_uncond + guidance_scale *
|
193 |
-
|
|
|
194 |
# compute the previous noisy sample x_t -> x_t-1
|
195 |
-
latents, scheduler_state = self.scheduler.step(
|
196 |
-
|
197 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
198 |
return (step + 1, latents, x_t0_1, x_t1_1, scheduler_state)
|
|
|
199 |
latents_shape = latents.shape
|
200 |
x_t0_1, x_t1_1 = jnp.zeros(latents_shape), jnp.zeros(latents_shape)
|
201 |
|
202 |
def cond_fun(arg):
|
203 |
step, latents, x_t0_1, x_t1_1, scheduler_state = arg
|
204 |
return (step < skip_t) & (step < num_inference_steps)
|
205 |
-
|
206 |
if DEBUG:
|
207 |
step = 0
|
208 |
while cond_fun((step, latents, x_t0_1, x_t1_1)):
|
209 |
-
step, latents, x_t0_1, x_t1_1, scheduler_state = while_body(
|
|
|
|
|
210 |
step = step + 1
|
211 |
else:
|
212 |
-
_, latents, x_t0_1, x_t1_1, scheduler_state = jax.lax.while_loop(
|
213 |
-
|
|
|
|
|
214 |
res = {"x0": latents.copy()}
|
215 |
if x_t0_1 is not None:
|
216 |
-
x_t0_1 = rearrange(x_t0_1, "(b f) c w
|
217 |
res["x_t0_1"] = x_t0_1.copy()
|
218 |
if x_t1_1 is not None:
|
219 |
-
x_t1_1 = rearrange(x_t1_1, "(b f) c w
|
220 |
res["x_t1_1"] = x_t1_1.copy()
|
221 |
return res
|
222 |
-
|
223 |
def warp_latents_independently(self, latents, reference_flow):
|
224 |
_, _, H, W = reference_flow.shape
|
225 |
b, _, f, h, w = latents.shape
|
@@ -230,10 +281,10 @@ class FlaxTextToVideoPipeline(FlaxDiffusionPipeline):
|
|
230 |
coords_t0 = coords_t0.at[:, 1].set(coords_t0[:, 1] * h / H)
|
231 |
f, c, _, _ = coords_t0.shape
|
232 |
coords_t0 = jax.image.resize(coords_t0, (f, c, h, w), "linear")
|
233 |
-
coords_t0 = rearrange(coords_t0,
|
234 |
-
latents_0 = rearrange(latents[0],
|
235 |
warped = grid_sample(latents_0, coords_t0, "mirror")
|
236 |
-
warped = rearrange(warped,
|
237 |
return warped
|
238 |
|
239 |
def warp_vid_independently(self, vid, reference_flow):
|
@@ -245,74 +296,173 @@ class FlaxTextToVideoPipeline(FlaxDiffusionPipeline):
|
|
245 |
coords_t0 = coords_t0.at[:, 1].set(coords_t0[:, 1] * h / H)
|
246 |
f, c, _, _ = coords_t0.shape
|
247 |
coords_t0 = jax.image.resize(coords_t0, (f, c, h, w), "linear")
|
248 |
-
coords_t0 = rearrange(coords_t0,
|
249 |
# latents_0 = rearrange(vid, 'c f h w -> f c h w')
|
250 |
warped = grid_sample(vid, coords_t0, "zeropad")
|
251 |
# warped = rearrange(warped, 'f c h w -> b c f h w', f=f)
|
252 |
return warped
|
253 |
-
|
254 |
-
def create_motion_field(
|
255 |
-
|
256 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
257 |
for fr_idx, frame_id in enumerate(frame_ids):
|
258 |
-
reference_flow = reference_flow.at[fr_idx, 0, :,
|
259 |
-
|
260 |
-
|
261 |
-
|
|
|
|
|
262 |
return reference_flow
|
263 |
-
|
264 |
-
def create_motion_field_and_warp_latents(
|
265 |
-
|
266 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
267 |
for idx, latent in enumerate(latents):
|
268 |
-
latents = latents.at[idx].set(
|
269 |
-
latent[None], motion_field)[0]
|
|
|
270 |
return motion_field, latents
|
271 |
-
|
272 |
-
def text_to_video_zero(
|
273 |
-
|
274 |
-
|
275 |
-
|
276 |
-
|
277 |
-
|
278 |
-
|
279 |
-
|
280 |
-
|
281 |
-
|
282 |
-
|
283 |
-
|
284 |
-
|
285 |
-
|
286 |
-
|
287 |
-
|
288 |
-
|
289 |
-
|
|
|
|
|
|
|
290 |
frame_ids = list(range(video_length))
|
291 |
# Prepare timesteps
|
292 |
-
params["scheduler"] = self.scheduler.set_timesteps(
|
|
|
|
|
293 |
timesteps = params["scheduler"].timesteps
|
294 |
# Prepare latent variables
|
295 |
num_channels_latents = self.unet.in_channels
|
296 |
batch_size = 1
|
297 |
-
xT = prepare_latents(
|
298 |
-
|
299 |
-
|
300 |
-
|
301 |
-
|
302 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
303 |
timesteps_ddpm.reverse()
|
304 |
t0 = timesteps_ddpm[t0]
|
305 |
t1 = timesteps_ddpm[t1]
|
306 |
x_t1_1 = None
|
307 |
|
308 |
# Denoising loop
|
309 |
-
shape = (
|
310 |
-
|
|
|
|
|
|
|
|
|
|
|
311 |
|
312 |
# perform ∆t backward steps by stable diffusion
|
313 |
-
ddim_res = self.DDIM_backward(
|
314 |
-
|
315 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
316 |
x0 = ddim_res["x0"]
|
317 |
|
318 |
# apply warping functions
|
@@ -320,37 +470,524 @@ class FlaxTextToVideoPipeline(FlaxDiffusionPipeline):
|
|
320 |
x_t0_1 = ddim_res["x_t0_1"]
|
321 |
if "x_t1_1" in ddim_res:
|
322 |
x_t1_1 = ddim_res["x_t1_1"]
|
323 |
-
x_t0_k = x_t0_1[:, :, :1, :, :].repeat(video_length-1, 2)
|
324 |
reference_flow, x_t0_k = self.create_motion_field_and_warp_latents(
|
325 |
-
motion_field_strength_x=motion_field_strength_x,
|
|
|
|
|
|
|
|
|
|
|
326 |
# assuming t0=t1=1000, if t0 = 1000
|
327 |
|
328 |
# DDPM forward for more motion freedom
|
329 |
-
ddpm_fwd = partial(
|
330 |
-
|
331 |
-
|
332 |
-
|
333 |
-
|
|
|
|
|
|
|
|
|
334 |
)
|
335 |
-
|
|
|
336 |
|
337 |
# backward stepts by stable diffusion
|
338 |
|
339 |
-
#warp the controlnet image following the same flow defined for latent
|
340 |
controlnet_video = controlnet_image[:video_length]
|
341 |
-
controlnet_video = controlnet_video.at[1:].set(
|
342 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
343 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
344 |
|
345 |
-
ddim_res = self.DDIM_backward(params, num_inference_steps=num_inference_steps, timesteps=timesteps, skip_t=t1, t0=-1, t1=-1, do_classifier_free_guidance=do_classifier_free_guidance,
|
346 |
-
text_embeddings=text_embeddings, latents_local=x_t1, guidance_scale=guidance_scale,
|
347 |
-
controlnet_image=controlnet_image, controlnet_conditioning_scale=controlnet_conditioning_scale)
|
348 |
x0 = ddim_res["x0"]
|
|
|
|
|
|
|
|
|
349 |
return x0
|
350 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
351 |
def prepare_text_inputs(self, prompt: Union[str, List[str]]):
|
352 |
if not isinstance(prompt, (str, list)):
|
353 |
-
raise ValueError(
|
|
|
|
|
354 |
text_input = self.tokenizer(
|
355 |
prompt,
|
356 |
padding="max_length",
|
@@ -359,27 +996,38 @@ class FlaxTextToVideoPipeline(FlaxDiffusionPipeline):
|
|
359 |
return_tensors="np",
|
360 |
)
|
361 |
return text_input.input_ids
|
|
|
362 |
def prepare_image_inputs(self, image: Union[Image.Image, List[Image.Image]]):
|
363 |
if not isinstance(image, (Image.Image, list)):
|
364 |
-
raise ValueError(
|
|
|
|
|
365 |
if isinstance(image, Image.Image):
|
366 |
image = [image]
|
367 |
-
processed_images = jnp.concatenate(
|
|
|
|
|
368 |
return processed_images
|
|
|
369 |
def _get_has_nsfw_concepts(self, features, params):
|
370 |
has_nsfw_concepts = self.safety_checker(features, params)
|
371 |
return has_nsfw_concepts
|
|
|
372 |
def _run_safety_checker(self, images, safety_model_params, jit=False):
|
373 |
# safety_model_params should already be replicated when jit is True
|
374 |
pil_images = [Image.fromarray(image) for image in images]
|
375 |
features = self.feature_extractor(pil_images, return_tensors="np").pixel_values
|
376 |
if jit:
|
377 |
features = shard(features)
|
378 |
-
has_nsfw_concepts = _p_get_has_nsfw_concepts(
|
|
|
|
|
379 |
has_nsfw_concepts = unshard(has_nsfw_concepts)
|
380 |
safety_model_params = unreplicate(safety_model_params)
|
381 |
else:
|
382 |
-
has_nsfw_concepts = self._get_has_nsfw_concepts(
|
|
|
|
|
383 |
images_was_copied = False
|
384 |
for idx, has_nsfw_concept in enumerate(has_nsfw_concepts):
|
385 |
if has_nsfw_concept:
|
@@ -393,6 +1041,7 @@ class FlaxTextToVideoPipeline(FlaxDiffusionPipeline):
|
|
393 |
" instead. Try again with a different prompt and/or seed."
|
394 |
)
|
395 |
return images, has_nsfw_concepts
|
|
|
396 |
def _generate(
|
397 |
self,
|
398 |
prompt_ids: jnp.array,
|
@@ -404,7 +1053,8 @@ class FlaxTextToVideoPipeline(FlaxDiffusionPipeline):
|
|
404 |
latents: Optional[jnp.array] = None,
|
405 |
neg_prompt_ids: Optional[jnp.array] = None,
|
406 |
controlnet_conditioning_scale: float = 1.0,
|
407 |
-
xT
|
|
|
408 |
motion_field_strength_x: float = 12,
|
409 |
motion_field_strength_y: float = 12,
|
410 |
t0: int = 44,
|
@@ -413,7 +1063,9 @@ class FlaxTextToVideoPipeline(FlaxDiffusionPipeline):
|
|
413 |
height, width = image.shape[-2:]
|
414 |
video_length = image.shape[0]
|
415 |
if height % 64 != 0 or width % 64 != 0:
|
416 |
-
raise ValueError(
|
|
|
|
|
417 |
# get prompt text embeddings
|
418 |
prompt_embeds = self.text_encoder(prompt_ids, params=params["text_encoder"])[0]
|
419 |
# TODO: currently it is assumed `do_classifier_free_guidance = guidance_scale > 1.0`
|
@@ -422,30 +1074,47 @@ class FlaxTextToVideoPipeline(FlaxDiffusionPipeline):
|
|
422 |
max_length = prompt_ids.shape[-1]
|
423 |
if neg_prompt_ids is None:
|
424 |
uncond_input = self.tokenizer(
|
425 |
-
[""] * batch_size,
|
|
|
|
|
|
|
426 |
).input_ids
|
427 |
else:
|
428 |
uncond_input = neg_prompt_ids
|
429 |
-
negative_prompt_embeds = self.text_encoder(
|
|
|
|
|
430 |
context = jnp.concatenate([negative_prompt_embeds, prompt_embeds])
|
431 |
image = jnp.concatenate([image] * 2)
|
432 |
seed_t2vz, prng_seed = jax.random.split(prng_seed)
|
433 |
-
#get the latent following text to video zero
|
434 |
-
latents = self.text_to_video_zero(
|
435 |
-
|
436 |
-
|
437 |
-
|
438 |
-
|
439 |
-
|
440 |
-
|
441 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
442 |
# scale and decode the image latents with vae
|
443 |
latents = 1 / self.vae.config.scaling_factor * latents
|
444 |
latents = rearrange(latents, "b c f h w -> (b f) c h w")
|
445 |
-
video = self.vae.apply(
|
|
|
|
|
446 |
video = (video / 2 + 0.5).clip(0, 1).transpose(0, 2, 3, 1)
|
447 |
return video
|
448 |
-
|
449 |
@replace_example_docstring(EXAMPLE_DOC_STRING)
|
450 |
def __call__(
|
451 |
self,
|
@@ -460,7 +1129,8 @@ class FlaxTextToVideoPipeline(FlaxDiffusionPipeline):
|
|
460 |
controlnet_conditioning_scale: Union[float, jnp.array] = 1.0,
|
461 |
return_dict: bool = True,
|
462 |
jit: bool = False,
|
463 |
-
xT
|
|
|
464 |
motion_field_strength_x: float = 3,
|
465 |
motion_field_strength_y: float = 4,
|
466 |
t0: int = 44,
|
@@ -517,7 +1187,9 @@ class FlaxTextToVideoPipeline(FlaxDiffusionPipeline):
|
|
517 |
if isinstance(controlnet_conditioning_scale, float):
|
518 |
# Convert to a tensor so each device gets a copy. Follow the prompt_ids for
|
519 |
# shape information, as they may be sharded (when `jit` is `True`), or not.
|
520 |
-
controlnet_conditioning_scale = jnp.array(
|
|
|
|
|
521 |
if len(prompt_ids.shape) > 2:
|
522 |
# Assume sharded
|
523 |
controlnet_conditioning_scale = controlnet_conditioning_scale[:, None]
|
@@ -534,6 +1206,7 @@ class FlaxTextToVideoPipeline(FlaxDiffusionPipeline):
|
|
534 |
neg_prompt_ids,
|
535 |
controlnet_conditioning_scale,
|
536 |
xT,
|
|
|
537 |
motion_field_strength_x,
|
538 |
motion_field_strength_y,
|
539 |
t0,
|
@@ -551,6 +1224,7 @@ class FlaxTextToVideoPipeline(FlaxDiffusionPipeline):
|
|
551 |
neg_prompt_ids,
|
552 |
controlnet_conditioning_scale,
|
553 |
xT,
|
|
|
554 |
motion_field_strength_x,
|
555 |
motion_field_strength_y,
|
556 |
t0,
|
@@ -560,8 +1234,12 @@ class FlaxTextToVideoPipeline(FlaxDiffusionPipeline):
|
|
560 |
safety_params = params["safety_checker"]
|
561 |
images_uint8_casted = (images * 255).round().astype("uint8")
|
562 |
num_devices, batch_size = images.shape[:2]
|
563 |
-
images_uint8_casted = np.asarray(images_uint8_casted).reshape(
|
564 |
-
|
|
|
|
|
|
|
|
|
565 |
images = np.asarray(images)
|
566 |
# block images
|
567 |
if any(has_nsfw_concept):
|
@@ -574,17 +1252,21 @@ class FlaxTextToVideoPipeline(FlaxDiffusionPipeline):
|
|
574 |
has_nsfw_concept = False
|
575 |
if not return_dict:
|
576 |
return (images, has_nsfw_concept)
|
577 |
-
return FlaxStableDiffusionPipelineOutput(
|
|
|
|
|
|
|
|
|
578 |
# Static argnums are pipe, num_inference_steps. A change would trigger recompilation.
|
579 |
# Non-static args are (sharded) input tensors mapped over their first dimension (hence, `0`).
|
580 |
@partial(
|
581 |
jax.pmap,
|
582 |
-
in_axes=(None, 0, 0, 0, 0, None, 0, 0, 0, 0, 0,
|
583 |
-
static_broadcasted_argnums=(0, 5,
|
584 |
)
|
585 |
def _p_generate(
|
586 |
pipe,
|
587 |
-
prompt_ids,
|
588 |
image,
|
589 |
params,
|
590 |
prng_seed,
|
@@ -594,6 +1276,7 @@ def _p_generate(
|
|
594 |
neg_prompt_ids,
|
595 |
controlnet_conditioning_scale,
|
596 |
xT,
|
|
|
597 |
motion_field_strength_x,
|
598 |
motion_field_strength_y,
|
599 |
t0,
|
@@ -610,19 +1293,26 @@ def _p_generate(
|
|
610 |
neg_prompt_ids,
|
611 |
controlnet_conditioning_scale,
|
612 |
xT,
|
|
|
613 |
motion_field_strength_x,
|
614 |
motion_field_strength_y,
|
615 |
t0,
|
616 |
t1,
|
617 |
)
|
|
|
|
|
618 |
@partial(jax.pmap, static_broadcasted_argnums=(0,))
|
619 |
def _p_get_has_nsfw_concepts(pipe, features, params):
|
620 |
return pipe._get_has_nsfw_concepts(features, params)
|
|
|
|
|
621 |
def unshard(x: jnp.ndarray):
|
622 |
# einops.rearrange(x, 'd b ... -> (d b) ...')
|
623 |
num_devices, batch_size = x.shape[:2]
|
624 |
rest = x.shape[2:]
|
625 |
return x.reshape(num_devices * batch_size, *rest)
|
|
|
|
|
626 |
def preprocess(image, dtype):
|
627 |
image = image.convert("RGB")
|
628 |
w, h = image.size
|
@@ -632,43 +1322,98 @@ def preprocess(image, dtype):
|
|
632 |
image = image[None].transpose(0, 3, 1, 2)
|
633 |
return image
|
634 |
|
635 |
-
|
636 |
-
|
637 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
638 |
# scale the initial noise by the standard deviation required by the scheduler
|
639 |
if latents is None:
|
640 |
latents = jax.random.normal(prng, shape)
|
641 |
latents = latents * params["scheduler"].init_noise_sigma
|
642 |
return latents
|
643 |
|
|
|
644 |
def coords_grid(batch, ht, wd):
|
645 |
coords = jnp.meshgrid(jnp.arange(ht), jnp.arange(wd), indexing="ij")
|
646 |
coords = jnp.stack(coords[::-1], axis=0)
|
647 |
return coords[None].repeat(batch, 0)
|
648 |
|
|
|
649 |
def adapt_pos_mirror(x, y, W, H):
|
650 |
-
|
651 |
-
|
652 |
-
|
653 |
-
|
654 |
-
|
655 |
-
|
|
|
|
|
|
|
|
|
|
|
656 |
|
657 |
-
def
|
658 |
-
|
659 |
|
660 |
-
def safe_get_mirror(img, x,y,W,H):
|
661 |
-
return img[adapt_pos_mirror(x,y,W,H)]
|
662 |
|
663 |
@partial(jax.vmap, in_axes=(0, 0, None))
|
664 |
@partial(jax.vmap, in_axes=(0, None, None))
|
665 |
-
@partial(jax.vmap, in_axes=(None,0, None))
|
666 |
@partial(jax.vmap, in_axes=(None, 0, None))
|
667 |
def grid_sample(latents, grid, method):
|
668 |
# this is an alternative to torch.functional.nn.grid_sample in jax
|
669 |
# this implementation is following the algorithm described @ https://pytorch.org/docs/stable/generated/torch.nn.functional.grid_sample.html
|
670 |
# but with coordinates scaled to the size of the image
|
671 |
if method == "mirror":
|
672 |
-
|
673 |
-
|
674 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
6 |
import numpy as np
|
7 |
from flax.core.frozen_dict import FrozenDict
|
8 |
from flax.jax_utils import unreplicate
|
9 |
+
from flax import jax_utils
|
10 |
from flax.training.common_utils import shard
|
11 |
from PIL import Image
|
12 |
from transformers import CLIPFeatureExtractor, CLIPTokenizer, FlaxCLIPTextModel
|
13 |
from einops import rearrange, repeat
|
14 |
+
from diffusers.models import (
|
15 |
+
FlaxAutoencoderKL,
|
16 |
+
FlaxControlNetModel,
|
17 |
+
FlaxUNet2DConditionModel,
|
18 |
+
)
|
19 |
from diffusers.schedulers import (
|
20 |
FlaxDDIMScheduler,
|
21 |
FlaxDPMSolverMultistepScheduler,
|
|
|
25 |
from diffusers.utils import PIL_INTERPOLATION, logging, replace_example_docstring
|
26 |
from diffusers.pipelines.pipeline_flax_utils import FlaxDiffusionPipeline
|
27 |
from diffusers.pipelines.stable_diffusion import FlaxStableDiffusionPipelineOutput
|
28 |
+
from diffusers.pipelines.stable_diffusion.safety_checker_flax import (
|
29 |
+
FlaxStableDiffusionSafetyChecker,
|
30 |
+
)
|
31 |
+
|
32 |
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
33 |
"""
|
34 |
Text2Video-Zero:
|
35 |
- Inputs: Prompt, Pose Control via mp4/gif, First Frame (?)
|
36 |
- JAX implementation
|
37 |
- 3DUnet to replace 2DUnetConditional
|
|
|
38 |
"""
|
39 |
|
40 |
+
|
41 |
+
def replicate_devices(array):
|
42 |
+
return jnp.expand_dims(array, 0).repeat(jax.device_count(), 0)
|
43 |
+
|
44 |
+
|
45 |
+
DEBUG = False # Set to True to use python for loop instead of jax.fori_loop for easier debugging
|
46 |
|
47 |
EXAMPLE_DOC_STRING = """
|
48 |
Examples:
|
|
|
101 |
>>> output_images.save("generated_image.png")
|
102 |
```
|
103 |
"""
|
104 |
+
|
105 |
+
|
106 |
class FlaxTextToVideoPipeline(FlaxDiffusionPipeline):
|
107 |
def __init__(
|
108 |
self,
|
109 |
+
vae,
|
110 |
+
text_encoder,
|
111 |
+
tokenizer,
|
112 |
+
unet,
|
113 |
+
unet_vanilla,
|
114 |
+
controlnet,
|
115 |
scheduler: Union[
|
116 |
+
FlaxDDIMScheduler,
|
117 |
+
FlaxPNDMScheduler,
|
118 |
+
FlaxLMSDiscreteScheduler,
|
119 |
+
FlaxDPMSolverMultistepScheduler,
|
120 |
],
|
121 |
safety_checker: FlaxStableDiffusionSafetyChecker,
|
122 |
feature_extractor: CLIPFeatureExtractor,
|
|
|
140 |
text_encoder=text_encoder,
|
141 |
tokenizer=tokenizer,
|
142 |
unet=unet,
|
143 |
+
unet_vanilla=unet_vanilla,
|
144 |
controlnet=controlnet,
|
145 |
scheduler=scheduler,
|
146 |
safety_checker=safety_checker,
|
|
|
154 |
else:
|
155 |
eps = jax.random.normal(prng, x0.shape, dtype=text_embeddings.dtype)
|
156 |
alpha_vec = jnp.prod(params["scheduler"].common.alphas[t0:tMax])
|
157 |
+
xt = jnp.sqrt(alpha_vec) * x0 + jnp.sqrt(1 - alpha_vec) * eps
|
|
|
158 |
return xt
|
159 |
+
|
160 |
+
def DDIM_backward(
|
161 |
+
self,
|
162 |
+
params,
|
163 |
+
num_inference_steps,
|
164 |
+
timesteps,
|
165 |
+
skip_t,
|
166 |
+
t0,
|
167 |
+
t1,
|
168 |
+
do_classifier_free_guidance,
|
169 |
+
text_embeddings,
|
170 |
+
latents_local,
|
171 |
+
guidance_scale,
|
172 |
+
controlnet_image=None,
|
173 |
+
controlnet_conditioning_scale=None,
|
174 |
+
):
|
175 |
+
scheduler_state = self.scheduler.set_timesteps(
|
176 |
+
params["scheduler"], num_inference_steps
|
177 |
+
)
|
178 |
f = latents_local.shape[2]
|
179 |
+
latents_local = rearrange(latents_local, "b c f h w -> (b f) c h w")
|
180 |
latents = latents_local.copy()
|
181 |
x_t0_1 = None
|
182 |
x_t1_1 = None
|
183 |
+
max_timestep = len(timesteps) - 1
|
184 |
timesteps = jnp.array(timesteps)
|
185 |
+
|
186 |
def while_body(args):
|
187 |
step, latents, x_t0_1, x_t1_1, scheduler_state = args
|
188 |
t = jnp.array(scheduler_state.timesteps, dtype=jnp.int32)[step]
|
189 |
+
latent_model_input = (
|
190 |
+
jnp.concatenate([latents] * 2)
|
191 |
+
if do_classifier_free_guidance
|
192 |
+
else latents
|
193 |
+
)
|
194 |
latent_model_input = self.scheduler.scale_model_input(
|
195 |
scheduler_state, latent_model_input, timestep=t
|
196 |
)
|
197 |
f = latents.shape[0]
|
198 |
+
te = jnp.stack(
|
199 |
+
[text_embeddings[0, :, :]] * f + [text_embeddings[-1, :, :]] * f
|
200 |
+
)
|
201 |
timestep = jnp.broadcast_to(t, latent_model_input.shape[0])
|
202 |
if controlnet_image is not None:
|
203 |
down_block_res_samples, mid_block_res_sample = self.controlnet.apply(
|
|
|
224 |
jnp.array(latent_model_input),
|
225 |
jnp.array(timestep, dtype=jnp.int32),
|
226 |
encoder_hidden_states=te,
|
227 |
+
).sample
|
228 |
# perform guidance
|
229 |
if do_classifier_free_guidance:
|
230 |
noise_pred_uncond, noise_pred_text = jnp.split(noise_pred, 2, axis=0)
|
231 |
+
noise_pred = noise_pred_uncond + guidance_scale * (
|
232 |
+
noise_pred_text - noise_pred_uncond
|
233 |
+
)
|
234 |
# compute the previous noisy sample x_t -> x_t-1
|
235 |
+
latents, scheduler_state = self.scheduler.step(
|
236 |
+
scheduler_state, noise_pred, t, latents
|
237 |
+
).to_tuple()
|
238 |
+
x_t0_1 = jax.lax.select(
|
239 |
+
(step < max_timestep - 1) & (timesteps[step + 1] == t0), latents, x_t0_1
|
240 |
+
)
|
241 |
+
x_t1_1 = jax.lax.select(
|
242 |
+
(step < max_timestep - 1) & (timesteps[step + 1] == t1), latents, x_t1_1
|
243 |
+
)
|
244 |
return (step + 1, latents, x_t0_1, x_t1_1, scheduler_state)
|
245 |
+
|
246 |
latents_shape = latents.shape
|
247 |
x_t0_1, x_t1_1 = jnp.zeros(latents_shape), jnp.zeros(latents_shape)
|
248 |
|
249 |
def cond_fun(arg):
|
250 |
step, latents, x_t0_1, x_t1_1, scheduler_state = arg
|
251 |
return (step < skip_t) & (step < num_inference_steps)
|
252 |
+
|
253 |
if DEBUG:
|
254 |
step = 0
|
255 |
while cond_fun((step, latents, x_t0_1, x_t1_1)):
|
256 |
+
step, latents, x_t0_1, x_t1_1, scheduler_state = while_body(
|
257 |
+
(step, latents, x_t0_1, x_t1_1, scheduler_state)
|
258 |
+
)
|
259 |
step = step + 1
|
260 |
else:
|
261 |
+
_, latents, x_t0_1, x_t1_1, scheduler_state = jax.lax.while_loop(
|
262 |
+
cond_fun, while_body, (0, latents, x_t0_1, x_t1_1, scheduler_state)
|
263 |
+
)
|
264 |
+
latents = rearrange(latents, "(b f) c h w -> b c f h w", f=f)
|
265 |
res = {"x0": latents.copy()}
|
266 |
if x_t0_1 is not None:
|
267 |
+
x_t0_1 = rearrange(x_t0_1, "(b f) c h w -> b c f h w", f=f)
|
268 |
res["x_t0_1"] = x_t0_1.copy()
|
269 |
if x_t1_1 is not None:
|
270 |
+
x_t1_1 = rearrange(x_t1_1, "(b f) c h w -> b c f h w", f=f)
|
271 |
res["x_t1_1"] = x_t1_1.copy()
|
272 |
return res
|
273 |
+
|
274 |
def warp_latents_independently(self, latents, reference_flow):
|
275 |
_, _, H, W = reference_flow.shape
|
276 |
b, _, f, h, w = latents.shape
|
|
|
281 |
coords_t0 = coords_t0.at[:, 1].set(coords_t0[:, 1] * h / H)
|
282 |
f, c, _, _ = coords_t0.shape
|
283 |
coords_t0 = jax.image.resize(coords_t0, (f, c, h, w), "linear")
|
284 |
+
coords_t0 = rearrange(coords_t0, "f c h w -> f h w c")
|
285 |
+
latents_0 = rearrange(latents[0], "c f h w -> f c h w")
|
286 |
warped = grid_sample(latents_0, coords_t0, "mirror")
|
287 |
+
warped = rearrange(warped, "(b f) c h w -> b c f h w", f=f)
|
288 |
return warped
|
289 |
|
290 |
def warp_vid_independently(self, vid, reference_flow):
|
|
|
296 |
coords_t0 = coords_t0.at[:, 1].set(coords_t0[:, 1] * h / H)
|
297 |
f, c, _, _ = coords_t0.shape
|
298 |
coords_t0 = jax.image.resize(coords_t0, (f, c, h, w), "linear")
|
299 |
+
coords_t0 = rearrange(coords_t0, "f c h w -> f h w c")
|
300 |
# latents_0 = rearrange(vid, 'c f h w -> f c h w')
|
301 |
warped = grid_sample(vid, coords_t0, "zeropad")
|
302 |
# warped = rearrange(warped, 'f c h w -> b c f h w', f=f)
|
303 |
return warped
|
304 |
+
|
305 |
+
def create_motion_field(
|
306 |
+
self,
|
307 |
+
motion_field_strength_x,
|
308 |
+
motion_field_strength_y,
|
309 |
+
frame_ids,
|
310 |
+
video_length,
|
311 |
+
latents,
|
312 |
+
):
|
313 |
+
reference_flow = jnp.zeros((video_length - 1, 2, 512, 512), dtype=latents.dtype)
|
314 |
for fr_idx, frame_id in enumerate(frame_ids):
|
315 |
+
reference_flow = reference_flow.at[fr_idx, 0, :, :].set(
|
316 |
+
motion_field_strength_x * (frame_id)
|
317 |
+
)
|
318 |
+
reference_flow = reference_flow.at[fr_idx, 1, :, :].set(
|
319 |
+
motion_field_strength_y * (frame_id)
|
320 |
+
)
|
321 |
return reference_flow
|
322 |
+
|
323 |
+
def create_motion_field_and_warp_latents(
|
324 |
+
self,
|
325 |
+
motion_field_strength_x,
|
326 |
+
motion_field_strength_y,
|
327 |
+
frame_ids,
|
328 |
+
video_length,
|
329 |
+
latents,
|
330 |
+
):
|
331 |
+
motion_field = self.create_motion_field(
|
332 |
+
motion_field_strength_x=motion_field_strength_x,
|
333 |
+
motion_field_strength_y=motion_field_strength_y,
|
334 |
+
latents=latents,
|
335 |
+
video_length=video_length,
|
336 |
+
frame_ids=frame_ids,
|
337 |
+
)
|
338 |
for idx, latent in enumerate(latents):
|
339 |
+
latents = latents.at[idx].set(
|
340 |
+
self.warp_latents_independently(latent[None], motion_field)[0]
|
341 |
+
)
|
342 |
return motion_field, latents
|
343 |
+
|
344 |
+
def text_to_video_zero(
|
345 |
+
self,
|
346 |
+
params,
|
347 |
+
prng,
|
348 |
+
text_embeddings,
|
349 |
+
video_length: Optional[int],
|
350 |
+
do_classifier_free_guidance=True,
|
351 |
+
height: Optional[int] = None,
|
352 |
+
width: Optional[int] = None,
|
353 |
+
num_inference_steps: int = 50,
|
354 |
+
guidance_scale: float = 7.5,
|
355 |
+
num_videos_per_prompt: Optional[int] = 1,
|
356 |
+
xT=None,
|
357 |
+
smooth_bg_strength: float = 0.0,
|
358 |
+
motion_field_strength_x: float = 12,
|
359 |
+
motion_field_strength_y: float = 12,
|
360 |
+
t0: int = 44,
|
361 |
+
t1: int = 47,
|
362 |
+
controlnet_image=None,
|
363 |
+
controlnet_conditioning_scale=0,
|
364 |
+
):
|
365 |
frame_ids = list(range(video_length))
|
366 |
# Prepare timesteps
|
367 |
+
params["scheduler"] = self.scheduler.set_timesteps(
|
368 |
+
params["scheduler"], num_inference_steps
|
369 |
+
)
|
370 |
timesteps = params["scheduler"].timesteps
|
371 |
# Prepare latent variables
|
372 |
num_channels_latents = self.unet.in_channels
|
373 |
batch_size = 1
|
374 |
+
xT = prepare_latents(
|
375 |
+
params,
|
376 |
+
prng,
|
377 |
+
batch_size * num_videos_per_prompt,
|
378 |
+
num_channels_latents,
|
379 |
+
height,
|
380 |
+
width,
|
381 |
+
self.vae_scale_factor,
|
382 |
+
xT,
|
383 |
+
)
|
384 |
+
|
385 |
+
timesteps_ddpm = [
|
386 |
+
981,
|
387 |
+
961,
|
388 |
+
941,
|
389 |
+
921,
|
390 |
+
901,
|
391 |
+
881,
|
392 |
+
861,
|
393 |
+
841,
|
394 |
+
821,
|
395 |
+
801,
|
396 |
+
781,
|
397 |
+
761,
|
398 |
+
741,
|
399 |
+
721,
|
400 |
+
701,
|
401 |
+
681,
|
402 |
+
661,
|
403 |
+
641,
|
404 |
+
621,
|
405 |
+
601,
|
406 |
+
581,
|
407 |
+
561,
|
408 |
+
541,
|
409 |
+
521,
|
410 |
+
501,
|
411 |
+
481,
|
412 |
+
461,
|
413 |
+
441,
|
414 |
+
421,
|
415 |
+
401,
|
416 |
+
381,
|
417 |
+
361,
|
418 |
+
341,
|
419 |
+
321,
|
420 |
+
301,
|
421 |
+
281,
|
422 |
+
261,
|
423 |
+
241,
|
424 |
+
221,
|
425 |
+
201,
|
426 |
+
181,
|
427 |
+
161,
|
428 |
+
141,
|
429 |
+
121,
|
430 |
+
101,
|
431 |
+
81,
|
432 |
+
61,
|
433 |
+
41,
|
434 |
+
21,
|
435 |
+
1,
|
436 |
+
]
|
437 |
timesteps_ddpm.reverse()
|
438 |
t0 = timesteps_ddpm[t0]
|
439 |
t1 = timesteps_ddpm[t1]
|
440 |
x_t1_1 = None
|
441 |
|
442 |
# Denoising loop
|
443 |
+
shape = (
|
444 |
+
batch_size,
|
445 |
+
num_channels_latents,
|
446 |
+
1,
|
447 |
+
height // self.vae.scaling_factor,
|
448 |
+
width // self.vae.scaling_factor,
|
449 |
+
)
|
450 |
|
451 |
# perform ∆t backward steps by stable diffusion
|
452 |
+
ddim_res = self.DDIM_backward(
|
453 |
+
params,
|
454 |
+
num_inference_steps=num_inference_steps,
|
455 |
+
timesteps=timesteps,
|
456 |
+
skip_t=1000,
|
457 |
+
t0=t0,
|
458 |
+
t1=t1,
|
459 |
+
do_classifier_free_guidance=do_classifier_free_guidance,
|
460 |
+
text_embeddings=text_embeddings,
|
461 |
+
latents_local=xT,
|
462 |
+
guidance_scale=guidance_scale,
|
463 |
+
controlnet_image=jnp.stack([controlnet_image[0]] * 2),
|
464 |
+
controlnet_conditioning_scale=controlnet_conditioning_scale,
|
465 |
+
)
|
466 |
x0 = ddim_res["x0"]
|
467 |
|
468 |
# apply warping functions
|
|
|
470 |
x_t0_1 = ddim_res["x_t0_1"]
|
471 |
if "x_t1_1" in ddim_res:
|
472 |
x_t1_1 = ddim_res["x_t1_1"]
|
473 |
+
x_t0_k = x_t0_1[:, :, :1, :, :].repeat(video_length - 1, 2)
|
474 |
reference_flow, x_t0_k = self.create_motion_field_and_warp_latents(
|
475 |
+
motion_field_strength_x=motion_field_strength_x,
|
476 |
+
motion_field_strength_y=motion_field_strength_y,
|
477 |
+
latents=x_t0_k,
|
478 |
+
video_length=video_length,
|
479 |
+
frame_ids=frame_ids[1:],
|
480 |
+
)
|
481 |
# assuming t0=t1=1000, if t0 = 1000
|
482 |
|
483 |
# DDPM forward for more motion freedom
|
484 |
+
ddpm_fwd = partial(
|
485 |
+
self.DDPM_forward,
|
486 |
+
params=params,
|
487 |
+
prng=prng,
|
488 |
+
x0=x_t0_k,
|
489 |
+
t0=t0,
|
490 |
+
tMax=t1,
|
491 |
+
shape=shape,
|
492 |
+
text_embeddings=text_embeddings,
|
493 |
)
|
494 |
+
x_t1_k = jax.lax.cond(t1 > t0, ddpm_fwd, lambda: x_t0_k)
|
495 |
+
x_t1 = jnp.concatenate([x_t1_1, x_t1_k], axis=2)
|
496 |
|
497 |
# backward stepts by stable diffusion
|
498 |
|
499 |
+
# warp the controlnet image following the same flow defined for latent
|
500 |
controlnet_video = controlnet_image[:video_length]
|
501 |
+
controlnet_video = controlnet_video.at[1:].set(
|
502 |
+
self.warp_vid_independently(controlnet_video[1:], reference_flow)
|
503 |
+
)
|
504 |
+
controlnet_image = jnp.concatenate([controlnet_video] * 2)
|
505 |
+
smooth_bg = True
|
506 |
+
|
507 |
+
if smooth_bg:
|
508 |
+
# latent shape: "b c f h w"
|
509 |
+
M_FG = repeat(
|
510 |
+
get_mask_pose(controlnet_video),
|
511 |
+
"f h w -> b c f h w",
|
512 |
+
c=x_t1.shape[1],
|
513 |
+
b=batch_size,
|
514 |
+
)
|
515 |
+
initial_bg = repeat(
|
516 |
+
x_t1[:, :, 0] * (1 - M_FG[:, :, 0]),
|
517 |
+
"b c h w -> b c f h w",
|
518 |
+
f=video_length - 1,
|
519 |
+
)
|
520 |
+
# warp the controlnet image following the same flow defined for latent #f c h w
|
521 |
+
initial_bg_warped = self.warp_latents_independently(
|
522 |
+
initial_bg, reference_flow
|
523 |
+
)
|
524 |
+
bgs = x_t1[:, :, 1:] * (1 - M_FG[:, :, 1:]) # initial background
|
525 |
+
initial_mask_warped = 1 - self.warp_latents_independently(
|
526 |
+
repeat(M_FG[:, :, 0], "b c h w -> b c f h w", f=video_length - 1),
|
527 |
+
reference_flow,
|
528 |
+
)
|
529 |
+
# initial_mask_warped = 1 - warp_vid_independently(repeat(M_FG[:,:,0], "b c h w -> (b f) c h w", f = video_length-1), reference_flow)
|
530 |
+
# initial_mask_warped = rearrange(initial_mask_warped, "(b f) c h w -> b c f h w", b=batch_size)
|
531 |
+
mask = (1 - M_FG[:, :, 1:]) * initial_mask_warped
|
532 |
+
x_t1 = x_t1.at[:, :, 1:].set(
|
533 |
+
(1 - mask) * x_t1[:, :, 1:]
|
534 |
+
+ mask
|
535 |
+
* (
|
536 |
+
initial_bg_warped * smooth_bg_strength
|
537 |
+
+ (1 - smooth_bg_strength) * bgs
|
538 |
+
)
|
539 |
+
)
|
540 |
|
541 |
+
ddim_res = self.DDIM_backward(
|
542 |
+
params,
|
543 |
+
num_inference_steps=num_inference_steps,
|
544 |
+
timesteps=timesteps,
|
545 |
+
skip_t=t1,
|
546 |
+
t0=-1,
|
547 |
+
t1=-1,
|
548 |
+
do_classifier_free_guidance=do_classifier_free_guidance,
|
549 |
+
text_embeddings=text_embeddings,
|
550 |
+
latents_local=x_t1,
|
551 |
+
guidance_scale=guidance_scale,
|
552 |
+
controlnet_image=controlnet_image,
|
553 |
+
controlnet_conditioning_scale=controlnet_conditioning_scale,
|
554 |
+
)
|
555 |
|
|
|
|
|
|
|
556 |
x0 = ddim_res["x0"]
|
557 |
+
del ddim_res
|
558 |
+
del x_t1
|
559 |
+
del x_t1_1
|
560 |
+
del x_t1_k
|
561 |
return x0
|
562 |
|
563 |
+
def denoise_latent(
|
564 |
+
self,
|
565 |
+
params,
|
566 |
+
num_inference_steps,
|
567 |
+
timesteps,
|
568 |
+
do_classifier_free_guidance,
|
569 |
+
text_embeddings,
|
570 |
+
latents,
|
571 |
+
guidance_scale,
|
572 |
+
controlnet_image=None,
|
573 |
+
controlnet_conditioning_scale=None,
|
574 |
+
):
|
575 |
+
scheduler_state = self.scheduler.set_timesteps(
|
576 |
+
params["scheduler"], num_inference_steps
|
577 |
+
)
|
578 |
+
# f = latents_local.shape[2]
|
579 |
+
# latents_local = rearrange(latents_local, "b c f h w -> (b f) c h w")
|
580 |
+
|
581 |
+
max_timestep = len(timesteps) - 1
|
582 |
+
timesteps = jnp.array(timesteps)
|
583 |
+
|
584 |
+
def while_body(args):
|
585 |
+
step, latents, scheduler_state = args
|
586 |
+
t = jnp.array(scheduler_state.timesteps, dtype=jnp.int32)[step]
|
587 |
+
latent_model_input = (
|
588 |
+
jnp.concatenate([latents] * 2)
|
589 |
+
if do_classifier_free_guidance
|
590 |
+
else latents
|
591 |
+
)
|
592 |
+
latent_model_input = self.scheduler.scale_model_input(
|
593 |
+
scheduler_state, latent_model_input, timestep=t
|
594 |
+
)
|
595 |
+
f = latents.shape[0]
|
596 |
+
te = jnp.stack(
|
597 |
+
[text_embeddings[0, :, :]] * f + [text_embeddings[-1, :, :]] * f
|
598 |
+
)
|
599 |
+
timestep = jnp.broadcast_to(t, latent_model_input.shape[0])
|
600 |
+
if controlnet_image is not None:
|
601 |
+
down_block_res_samples, mid_block_res_sample = self.controlnet.apply(
|
602 |
+
{"params": params["controlnet"]},
|
603 |
+
jnp.array(latent_model_input),
|
604 |
+
jnp.array(timestep, dtype=jnp.int32),
|
605 |
+
encoder_hidden_states=te,
|
606 |
+
controlnet_cond=controlnet_image,
|
607 |
+
conditioning_scale=controlnet_conditioning_scale,
|
608 |
+
return_dict=False,
|
609 |
+
)
|
610 |
+
# predict the noise residual
|
611 |
+
noise_pred = self.unet_vanilla.apply(
|
612 |
+
{"params": params["unet"]},
|
613 |
+
jnp.array(latent_model_input),
|
614 |
+
jnp.array(timestep, dtype=jnp.int32),
|
615 |
+
encoder_hidden_states=te,
|
616 |
+
down_block_additional_residuals=down_block_res_samples,
|
617 |
+
mid_block_additional_residual=mid_block_res_sample,
|
618 |
+
).sample
|
619 |
+
else:
|
620 |
+
noise_pred = self.unet_vanilla.apply(
|
621 |
+
{"params": params["unet"]},
|
622 |
+
jnp.array(latent_model_input),
|
623 |
+
jnp.array(timestep, dtype=jnp.int32),
|
624 |
+
encoder_hidden_states=te,
|
625 |
+
).sample
|
626 |
+
# perform guidance
|
627 |
+
if do_classifier_free_guidance:
|
628 |
+
noise_pred_uncond, noise_pred_text = jnp.split(noise_pred, 2, axis=0)
|
629 |
+
noise_pred = noise_pred_uncond + guidance_scale * (
|
630 |
+
noise_pred_text - noise_pred_uncond
|
631 |
+
)
|
632 |
+
# compute the previous noisy sample x_t -> x_t-1
|
633 |
+
latents, scheduler_state = self.scheduler.step(
|
634 |
+
scheduler_state, noise_pred, t, latents
|
635 |
+
).to_tuple()
|
636 |
+
return (step + 1, latents, scheduler_state)
|
637 |
+
|
638 |
+
def cond_fun(arg):
|
639 |
+
step, latents, scheduler_state = arg
|
640 |
+
return step < num_inference_steps
|
641 |
+
|
642 |
+
if DEBUG:
|
643 |
+
step = 0
|
644 |
+
while cond_fun((step, latents, scheduler_state)):
|
645 |
+
step, latents, scheduler_state = while_body(
|
646 |
+
(step, latents, scheduler_state)
|
647 |
+
)
|
648 |
+
step = step + 1
|
649 |
+
else:
|
650 |
+
_, latents, scheduler_state = jax.lax.while_loop(
|
651 |
+
cond_fun, while_body, (0, latents, scheduler_state)
|
652 |
+
)
|
653 |
+
# latents = rearrange(latents, "(b f) c h w -> b c f h w", f=f)
|
654 |
+
return latents
|
655 |
+
|
656 |
+
@partial(jax.jit, static_argnums=(0, 1))
|
657 |
+
def _generate_starting_frames(
|
658 |
+
self,
|
659 |
+
num_inference_steps,
|
660 |
+
params,
|
661 |
+
timesteps,
|
662 |
+
text_embeddings,
|
663 |
+
latents,
|
664 |
+
guidance_scale,
|
665 |
+
controlnet_image,
|
666 |
+
controlnet_conditioning_scale,
|
667 |
+
):
|
668 |
+
# perform ∆t backward steps by stable diffusion
|
669 |
+
# delta_t_diffusion = jax.vmap(lambda latent : self.DDIM_backward(params, num_inference_steps=num_inference_steps, timesteps=timesteps, skip_t=1000, t0=t0, t1=t1, do_classifier_free_guidance=do_classifier_free_guidance,
|
670 |
+
# text_embeddings=text_embeddings, latents_local=latent, guidance_scale=guidance_scale,
|
671 |
+
# controlnet_image=controlnet_image, controlnet_conditioning_scale=controlnet_conditioning_scale))
|
672 |
+
# ddim_res = delta_t_diffusion(latents)
|
673 |
+
# latents = ddim_res["x0"] #output is i b c f h w
|
674 |
+
|
675 |
+
# DDPM forward for more motion freedom
|
676 |
+
# ddpm_fwd = jax.vmap(lambda prng, latent: self.DDPM_forward(params=params, prng=prng, x0=latent, t0=t0,
|
677 |
+
# tMax=t1, shape=shape, text_embeddings=text_embeddings))
|
678 |
+
# latents = ddpm_fwd(stacked_prngs, latents)
|
679 |
+
# main backward diffusion
|
680 |
+
# denoise_first_frame = lambda latent : self.DDIM_backward(params, num_inference_steps=num_inference_steps, timesteps=timesteps, skip_t=100000, t0=-1, t1=-1, do_classifier_free_guidance=do_classifier_free_guidance,
|
681 |
+
# text_embeddings=text_embeddings, latents_local=latent, guidance_scale=guidance_scale,
|
682 |
+
# controlnet_image=controlnet_image, controlnet_conditioning_scale=controlnet_conditioning_scale, use_vanilla=True)
|
683 |
+
# latents = rearrange(latents, 'i b c f h w -> (i b) c f h w')
|
684 |
+
# ddim_res = denoise_first_frame(latents)
|
685 |
+
latents = self.denoise_latent(
|
686 |
+
params,
|
687 |
+
num_inference_steps=num_inference_steps,
|
688 |
+
timesteps=timesteps,
|
689 |
+
do_classifier_free_guidance=True,
|
690 |
+
text_embeddings=text_embeddings,
|
691 |
+
latents=latents,
|
692 |
+
guidance_scale=guidance_scale,
|
693 |
+
controlnet_image=controlnet_image,
|
694 |
+
controlnet_conditioning_scale=controlnet_conditioning_scale,
|
695 |
+
)
|
696 |
+
# latents = rearrange(ddim_res["x0"], 'i b c f h w -> (i b) c f h w') #output is i b c f h w
|
697 |
+
|
698 |
+
# scale and decode the image latents with vae
|
699 |
+
latents = 1 / self.vae.config.scaling_factor * latents
|
700 |
+
# latents = rearrange(latents, "b c h w -> (b f) c h w")
|
701 |
+
imgs = self.vae.apply(
|
702 |
+
{"params": params["vae"]}, latents, method=self.vae.decode
|
703 |
+
).sample
|
704 |
+
imgs = (imgs / 2 + 0.5).clip(0, 1).transpose(0, 2, 3, 1)
|
705 |
+
return imgs
|
706 |
+
|
707 |
+
def generate_starting_frames(
|
708 |
+
self,
|
709 |
+
params,
|
710 |
+
prngs: list, # list of prngs for each img
|
711 |
+
prompt,
|
712 |
+
neg_prompt,
|
713 |
+
controlnet_image,
|
714 |
+
do_classifier_free_guidance=True,
|
715 |
+
num_inference_steps: int = 50,
|
716 |
+
guidance_scale: float = 7.5,
|
717 |
+
t0: int = 44,
|
718 |
+
t1: int = 47,
|
719 |
+
controlnet_conditioning_scale=1.0,
|
720 |
+
):
|
721 |
+
height, width = controlnet_image.shape[-2:]
|
722 |
+
if height % 64 != 0 or width % 64 != 0:
|
723 |
+
raise ValueError(
|
724 |
+
f"`height` and `width` have to be divisible by 64 but are {height} and {width}."
|
725 |
+
)
|
726 |
+
|
727 |
+
shape = (
|
728 |
+
self.unet.in_channels,
|
729 |
+
height // self.vae_scale_factor,
|
730 |
+
width // self.vae_scale_factor,
|
731 |
+
) # b c h w
|
732 |
+
# scale the initial noise by the standard deviation required by the scheduler
|
733 |
+
|
734 |
+
print(
|
735 |
+
f"Generating {len(prngs)} first frames with prompt {prompt}, for {num_inference_steps} steps. PRNG seeds are: {prngs}"
|
736 |
+
)
|
737 |
+
|
738 |
+
latents = jnp.stack(
|
739 |
+
[jax.random.normal(prng, shape) for prng in prngs]
|
740 |
+
) # b c h w
|
741 |
+
latents = latents * params["scheduler"].init_noise_sigma
|
742 |
+
|
743 |
+
timesteps = params["scheduler"].timesteps
|
744 |
+
timesteps_ddpm = [
|
745 |
+
981,
|
746 |
+
961,
|
747 |
+
941,
|
748 |
+
921,
|
749 |
+
901,
|
750 |
+
881,
|
751 |
+
861,
|
752 |
+
841,
|
753 |
+
821,
|
754 |
+
801,
|
755 |
+
781,
|
756 |
+
761,
|
757 |
+
741,
|
758 |
+
721,
|
759 |
+
701,
|
760 |
+
681,
|
761 |
+
661,
|
762 |
+
641,
|
763 |
+
621,
|
764 |
+
601,
|
765 |
+
581,
|
766 |
+
561,
|
767 |
+
541,
|
768 |
+
521,
|
769 |
+
501,
|
770 |
+
481,
|
771 |
+
461,
|
772 |
+
441,
|
773 |
+
421,
|
774 |
+
401,
|
775 |
+
381,
|
776 |
+
361,
|
777 |
+
341,
|
778 |
+
321,
|
779 |
+
301,
|
780 |
+
281,
|
781 |
+
261,
|
782 |
+
241,
|
783 |
+
221,
|
784 |
+
201,
|
785 |
+
181,
|
786 |
+
161,
|
787 |
+
141,
|
788 |
+
121,
|
789 |
+
101,
|
790 |
+
81,
|
791 |
+
61,
|
792 |
+
41,
|
793 |
+
21,
|
794 |
+
1,
|
795 |
+
]
|
796 |
+
timesteps_ddpm.reverse()
|
797 |
+
t0 = timesteps_ddpm[t0]
|
798 |
+
t1 = timesteps_ddpm[t1]
|
799 |
+
|
800 |
+
# get prompt text embeddings
|
801 |
+
prompt_ids = self.prepare_text_inputs(prompt)
|
802 |
+
prompt_embeds = self.text_encoder(prompt_ids, params=params["text_encoder"])[0]
|
803 |
+
|
804 |
+
# TODO: currently it is assumed `do_classifier_free_guidance = guidance_scale > 1.0`
|
805 |
+
# implement this conditional `do_classifier_free_guidance = guidance_scale > 1.0`
|
806 |
+
batch_size = 1
|
807 |
+
max_length = prompt_ids.shape[-1]
|
808 |
+
if neg_prompt is None:
|
809 |
+
uncond_input = self.tokenizer(
|
810 |
+
[""] * batch_size,
|
811 |
+
padding="max_length",
|
812 |
+
max_length=max_length,
|
813 |
+
return_tensors="np",
|
814 |
+
).input_ids
|
815 |
+
else:
|
816 |
+
neg_prompt_ids = self.prepare_text_inputs(neg_prompt)
|
817 |
+
uncond_input = neg_prompt_ids
|
818 |
+
|
819 |
+
negative_prompt_embeds = self.text_encoder(
|
820 |
+
uncond_input, params=params["text_encoder"]
|
821 |
+
)[0]
|
822 |
+
text_embeddings = jnp.concatenate([negative_prompt_embeds, prompt_embeds])
|
823 |
+
controlnet_image = jnp.stack([controlnet_image[0]] * 2 * len(prngs))
|
824 |
+
return self._generate_starting_frames(
|
825 |
+
num_inference_steps,
|
826 |
+
params,
|
827 |
+
timesteps,
|
828 |
+
text_embeddings,
|
829 |
+
latents,
|
830 |
+
guidance_scale,
|
831 |
+
controlnet_image,
|
832 |
+
controlnet_conditioning_scale,
|
833 |
+
)
|
834 |
+
|
835 |
+
def generate_video(
|
836 |
+
self,
|
837 |
+
prompt: str,
|
838 |
+
image: jnp.array,
|
839 |
+
params: Union[Dict, FrozenDict],
|
840 |
+
prng_seed: jax.random.KeyArray,
|
841 |
+
num_inference_steps: int = 50,
|
842 |
+
guidance_scale: Union[float, jnp.array] = 7.5,
|
843 |
+
latents: jnp.array = None,
|
844 |
+
neg_prompt: str = "",
|
845 |
+
controlnet_conditioning_scale: Union[float, jnp.array] = 1.0,
|
846 |
+
return_dict: bool = True,
|
847 |
+
jit: bool = False,
|
848 |
+
xT=None,
|
849 |
+
smooth_bg_strength: float = 0.0,
|
850 |
+
motion_field_strength_x: float = 3,
|
851 |
+
motion_field_strength_y: float = 4,
|
852 |
+
t0: int = 44,
|
853 |
+
t1: int = 47,
|
854 |
+
):
|
855 |
+
r"""
|
856 |
+
Function invoked when calling the pipeline for generation.
|
857 |
+
Args:
|
858 |
+
prompt_ids (`jnp.array`):
|
859 |
+
The prompt or prompts to guide the image generation.
|
860 |
+
image (`jnp.array`):
|
861 |
+
Array representing the ControlNet input condition. ControlNet use this input condition to generate
|
862 |
+
guidance to Unet.
|
863 |
+
params (`Dict` or `FrozenDict`): Dictionary containing the model parameters/weights
|
864 |
+
prng_seed (`jax.random.KeyArray` or `jax.Array`): Array containing random number generator key
|
865 |
+
num_inference_steps (`int`, *optional*, defaults to 50):
|
866 |
+
The number of denoising steps. More denoising steps usually lead to a higher quality image at the
|
867 |
+
expense of slower inference.
|
868 |
+
guidance_scale (`float`, *optional*, defaults to 7.5):
|
869 |
+
Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
|
870 |
+
`guidance_scale` is defined as `w` of equation 2. of [Imagen
|
871 |
+
Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
|
872 |
+
1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
|
873 |
+
usually at the expense of lower image quality.
|
874 |
+
latents (`jnp.array`, *optional*):
|
875 |
+
Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
|
876 |
+
generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
|
877 |
+
tensor will ge generated by sampling using the supplied random `generator`.
|
878 |
+
controlnet_conditioning_scale (`float` or `jnp.array`, *optional*, defaults to 1.0):
|
879 |
+
The outputs of the controlnet are multiplied by `controlnet_conditioning_scale` before they are added
|
880 |
+
to the residual in the original unet.
|
881 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
882 |
+
Whether or not to return a [`~pipelines.stable_diffusion.FlaxStableDiffusionPipelineOutput`] instead of
|
883 |
+
a plain tuple.
|
884 |
+
jit (`bool`, defaults to `False`):
|
885 |
+
Whether to run `pmap` versions of the generation and safety scoring functions. NOTE: This argument
|
886 |
+
exists because `__call__` is not yet end-to-end pmap-able. It will be removed in a future release.
|
887 |
+
Examples:
|
888 |
+
Returns:
|
889 |
+
[`~pipelines.stable_diffusion.FlaxStableDiffusionPipelineOutput`] or `tuple`:
|
890 |
+
[`~pipelines.stable_diffusion.FlaxStableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a
|
891 |
+
`tuple. When returning a tuple, the first element is a list with the generated images, and the second
|
892 |
+
element is a list of `bool`s denoting whether the corresponding generated image likely represents
|
893 |
+
"not-safe-for-work" (nsfw) content, according to the `safety_checker`.
|
894 |
+
"""
|
895 |
+
height, width = image.shape[-2:]
|
896 |
+
vid_length = image.shape[0]
|
897 |
+
# get prompt text embeddings
|
898 |
+
prompt_ids = self.prepare_text_inputs([prompt] * vid_length)
|
899 |
+
neg_prompt_ids = self.prepare_text_inputs([neg_prompt] * vid_length)
|
900 |
+
|
901 |
+
# TODO: currently it is assumed `do_classifier_free_guidance = guidance_scale > 1.0`
|
902 |
+
# implement this conditional `do_classifier_free_guidance = guidance_scale > 1.0`
|
903 |
+
batch_size = 1
|
904 |
+
|
905 |
+
if isinstance(guidance_scale, float):
|
906 |
+
# Convert to a tensor so each device gets a copy. Follow the prompt_ids for
|
907 |
+
# shape information, as they may be sharded (when `jit` is `True`), or not.
|
908 |
+
guidance_scale = jnp.array([guidance_scale] * prompt_ids.shape[0])
|
909 |
+
if len(prompt_ids.shape) > 2:
|
910 |
+
# Assume sharded
|
911 |
+
guidance_scale = guidance_scale[:, None]
|
912 |
+
if isinstance(controlnet_conditioning_scale, float):
|
913 |
+
# Convert to a tensor so each device gets a copy. Follow the prompt_ids for
|
914 |
+
# shape information, as they may be sharded (when `jit` is `True`), or not.
|
915 |
+
controlnet_conditioning_scale = jnp.array(
|
916 |
+
[controlnet_conditioning_scale] * prompt_ids.shape[0]
|
917 |
+
)
|
918 |
+
if len(prompt_ids.shape) > 2:
|
919 |
+
# Assume sharded
|
920 |
+
controlnet_conditioning_scale = controlnet_conditioning_scale[:, None]
|
921 |
+
if jit:
|
922 |
+
images = _p_generate(
|
923 |
+
self,
|
924 |
+
replicate_devices(prompt_ids),
|
925 |
+
replicate_devices(image),
|
926 |
+
jax_utils.replicate(params),
|
927 |
+
replicate_devices(prng_seed),
|
928 |
+
num_inference_steps,
|
929 |
+
replicate_devices(guidance_scale),
|
930 |
+
replicate_devices(latents) if latents is not None else None,
|
931 |
+
replicate_devices(neg_prompt_ids)
|
932 |
+
if neg_prompt_ids is not None
|
933 |
+
else None,
|
934 |
+
replicate_devices(controlnet_conditioning_scale),
|
935 |
+
replicate_devices(xT) if xT is not None else None,
|
936 |
+
replicate_devices(smooth_bg_strength),
|
937 |
+
replicate_devices(motion_field_strength_x),
|
938 |
+
replicate_devices(motion_field_strength_y),
|
939 |
+
t0,
|
940 |
+
t1,
|
941 |
+
)
|
942 |
+
else:
|
943 |
+
images = self._generate(
|
944 |
+
prompt_ids,
|
945 |
+
image,
|
946 |
+
params,
|
947 |
+
prng_seed,
|
948 |
+
num_inference_steps,
|
949 |
+
guidance_scale,
|
950 |
+
latents,
|
951 |
+
neg_prompt_ids,
|
952 |
+
controlnet_conditioning_scale,
|
953 |
+
xT,
|
954 |
+
smooth_bg_strength,
|
955 |
+
motion_field_strength_x,
|
956 |
+
motion_field_strength_y,
|
957 |
+
t0,
|
958 |
+
t1,
|
959 |
+
)
|
960 |
+
if self.safety_checker is not None:
|
961 |
+
safety_params = params["safety_checker"]
|
962 |
+
images_uint8_casted = (images * 255).round().astype("uint8")
|
963 |
+
num_devices, batch_size = images.shape[:2]
|
964 |
+
images_uint8_casted = np.asarray(images_uint8_casted).reshape(
|
965 |
+
num_devices * batch_size, height, width, 3
|
966 |
+
)
|
967 |
+
images_uint8_casted, has_nsfw_concept = self._run_safety_checker(
|
968 |
+
images_uint8_casted, safety_params, jit
|
969 |
+
)
|
970 |
+
images = np.asarray(images)
|
971 |
+
# block images
|
972 |
+
if any(has_nsfw_concept):
|
973 |
+
for i, is_nsfw in enumerate(has_nsfw_concept):
|
974 |
+
if is_nsfw:
|
975 |
+
images[i] = np.asarray(images_uint8_casted[i])
|
976 |
+
images = images.reshape(num_devices, batch_size, height, width, 3)
|
977 |
+
else:
|
978 |
+
images = np.asarray(images)
|
979 |
+
has_nsfw_concept = False
|
980 |
+
if not return_dict:
|
981 |
+
return (images, has_nsfw_concept)
|
982 |
+
return FlaxStableDiffusionPipelineOutput(
|
983 |
+
images=images, nsfw_content_detected=has_nsfw_concept
|
984 |
+
)
|
985 |
+
|
986 |
def prepare_text_inputs(self, prompt: Union[str, List[str]]):
|
987 |
if not isinstance(prompt, (str, list)):
|
988 |
+
raise ValueError(
|
989 |
+
f"`prompt` has to be of type `str` or `list` but is {type(prompt)}"
|
990 |
+
)
|
991 |
text_input = self.tokenizer(
|
992 |
prompt,
|
993 |
padding="max_length",
|
|
|
996 |
return_tensors="np",
|
997 |
)
|
998 |
return text_input.input_ids
|
999 |
+
|
1000 |
def prepare_image_inputs(self, image: Union[Image.Image, List[Image.Image]]):
|
1001 |
if not isinstance(image, (Image.Image, list)):
|
1002 |
+
raise ValueError(
|
1003 |
+
f"image has to be of type `PIL.Image.Image` or list but is {type(image)}"
|
1004 |
+
)
|
1005 |
if isinstance(image, Image.Image):
|
1006 |
image = [image]
|
1007 |
+
processed_images = jnp.concatenate(
|
1008 |
+
[preprocess(img, jnp.float32) for img in image]
|
1009 |
+
)
|
1010 |
return processed_images
|
1011 |
+
|
1012 |
def _get_has_nsfw_concepts(self, features, params):
|
1013 |
has_nsfw_concepts = self.safety_checker(features, params)
|
1014 |
return has_nsfw_concepts
|
1015 |
+
|
1016 |
def _run_safety_checker(self, images, safety_model_params, jit=False):
|
1017 |
# safety_model_params should already be replicated when jit is True
|
1018 |
pil_images = [Image.fromarray(image) for image in images]
|
1019 |
features = self.feature_extractor(pil_images, return_tensors="np").pixel_values
|
1020 |
if jit:
|
1021 |
features = shard(features)
|
1022 |
+
has_nsfw_concepts = _p_get_has_nsfw_concepts(
|
1023 |
+
self, features, safety_model_params
|
1024 |
+
)
|
1025 |
has_nsfw_concepts = unshard(has_nsfw_concepts)
|
1026 |
safety_model_params = unreplicate(safety_model_params)
|
1027 |
else:
|
1028 |
+
has_nsfw_concepts = self._get_has_nsfw_concepts(
|
1029 |
+
features, safety_model_params
|
1030 |
+
)
|
1031 |
images_was_copied = False
|
1032 |
for idx, has_nsfw_concept in enumerate(has_nsfw_concepts):
|
1033 |
if has_nsfw_concept:
|
|
|
1041 |
" instead. Try again with a different prompt and/or seed."
|
1042 |
)
|
1043 |
return images, has_nsfw_concepts
|
1044 |
+
|
1045 |
def _generate(
|
1046 |
self,
|
1047 |
prompt_ids: jnp.array,
|
|
|
1053 |
latents: Optional[jnp.array] = None,
|
1054 |
neg_prompt_ids: Optional[jnp.array] = None,
|
1055 |
controlnet_conditioning_scale: float = 1.0,
|
1056 |
+
xT=None,
|
1057 |
+
smooth_bg_strength: float = 0.0,
|
1058 |
motion_field_strength_x: float = 12,
|
1059 |
motion_field_strength_y: float = 12,
|
1060 |
t0: int = 44,
|
|
|
1063 |
height, width = image.shape[-2:]
|
1064 |
video_length = image.shape[0]
|
1065 |
if height % 64 != 0 or width % 64 != 0:
|
1066 |
+
raise ValueError(
|
1067 |
+
f"`height` and `width` have to be divisible by 64 but are {height} and {width}."
|
1068 |
+
)
|
1069 |
# get prompt text embeddings
|
1070 |
prompt_embeds = self.text_encoder(prompt_ids, params=params["text_encoder"])[0]
|
1071 |
# TODO: currently it is assumed `do_classifier_free_guidance = guidance_scale > 1.0`
|
|
|
1074 |
max_length = prompt_ids.shape[-1]
|
1075 |
if neg_prompt_ids is None:
|
1076 |
uncond_input = self.tokenizer(
|
1077 |
+
[""] * batch_size,
|
1078 |
+
padding="max_length",
|
1079 |
+
max_length=max_length,
|
1080 |
+
return_tensors="np",
|
1081 |
).input_ids
|
1082 |
else:
|
1083 |
uncond_input = neg_prompt_ids
|
1084 |
+
negative_prompt_embeds = self.text_encoder(
|
1085 |
+
uncond_input, params=params["text_encoder"]
|
1086 |
+
)[0]
|
1087 |
context = jnp.concatenate([negative_prompt_embeds, prompt_embeds])
|
1088 |
image = jnp.concatenate([image] * 2)
|
1089 |
seed_t2vz, prng_seed = jax.random.split(prng_seed)
|
1090 |
+
# get the latent following text to video zero
|
1091 |
+
latents = self.text_to_video_zero(
|
1092 |
+
params,
|
1093 |
+
seed_t2vz,
|
1094 |
+
text_embeddings=context,
|
1095 |
+
video_length=video_length,
|
1096 |
+
height=height,
|
1097 |
+
width=width,
|
1098 |
+
num_inference_steps=num_inference_steps,
|
1099 |
+
guidance_scale=guidance_scale,
|
1100 |
+
controlnet_image=image,
|
1101 |
+
xT=xT,
|
1102 |
+
smooth_bg_strength=smooth_bg_strength,
|
1103 |
+
t0=t0,
|
1104 |
+
t1=t1,
|
1105 |
+
motion_field_strength_x=motion_field_strength_x,
|
1106 |
+
motion_field_strength_y=motion_field_strength_y,
|
1107 |
+
controlnet_conditioning_scale=controlnet_conditioning_scale,
|
1108 |
+
)
|
1109 |
# scale and decode the image latents with vae
|
1110 |
latents = 1 / self.vae.config.scaling_factor * latents
|
1111 |
latents = rearrange(latents, "b c f h w -> (b f) c h w")
|
1112 |
+
video = self.vae.apply(
|
1113 |
+
{"params": params["vae"]}, latents, method=self.vae.decode
|
1114 |
+
).sample
|
1115 |
video = (video / 2 + 0.5).clip(0, 1).transpose(0, 2, 3, 1)
|
1116 |
return video
|
1117 |
+
|
1118 |
@replace_example_docstring(EXAMPLE_DOC_STRING)
|
1119 |
def __call__(
|
1120 |
self,
|
|
|
1129 |
controlnet_conditioning_scale: Union[float, jnp.array] = 1.0,
|
1130 |
return_dict: bool = True,
|
1131 |
jit: bool = False,
|
1132 |
+
xT=None,
|
1133 |
+
smooth_bg_strength: float = 0.0,
|
1134 |
motion_field_strength_x: float = 3,
|
1135 |
motion_field_strength_y: float = 4,
|
1136 |
t0: int = 44,
|
|
|
1187 |
if isinstance(controlnet_conditioning_scale, float):
|
1188 |
# Convert to a tensor so each device gets a copy. Follow the prompt_ids for
|
1189 |
# shape information, as they may be sharded (when `jit` is `True`), or not.
|
1190 |
+
controlnet_conditioning_scale = jnp.array(
|
1191 |
+
[controlnet_conditioning_scale] * prompt_ids.shape[0]
|
1192 |
+
)
|
1193 |
if len(prompt_ids.shape) > 2:
|
1194 |
# Assume sharded
|
1195 |
controlnet_conditioning_scale = controlnet_conditioning_scale[:, None]
|
|
|
1206 |
neg_prompt_ids,
|
1207 |
controlnet_conditioning_scale,
|
1208 |
xT,
|
1209 |
+
smooth_bg_strength,
|
1210 |
motion_field_strength_x,
|
1211 |
motion_field_strength_y,
|
1212 |
t0,
|
|
|
1224 |
neg_prompt_ids,
|
1225 |
controlnet_conditioning_scale,
|
1226 |
xT,
|
1227 |
+
smooth_bg_strength,
|
1228 |
motion_field_strength_x,
|
1229 |
motion_field_strength_y,
|
1230 |
t0,
|
|
|
1234 |
safety_params = params["safety_checker"]
|
1235 |
images_uint8_casted = (images * 255).round().astype("uint8")
|
1236 |
num_devices, batch_size = images.shape[:2]
|
1237 |
+
images_uint8_casted = np.asarray(images_uint8_casted).reshape(
|
1238 |
+
num_devices * batch_size, height, width, 3
|
1239 |
+
)
|
1240 |
+
images_uint8_casted, has_nsfw_concept = self._run_safety_checker(
|
1241 |
+
images_uint8_casted, safety_params, jit
|
1242 |
+
)
|
1243 |
images = np.asarray(images)
|
1244 |
# block images
|
1245 |
if any(has_nsfw_concept):
|
|
|
1252 |
has_nsfw_concept = False
|
1253 |
if not return_dict:
|
1254 |
return (images, has_nsfw_concept)
|
1255 |
+
return FlaxStableDiffusionPipelineOutput(
|
1256 |
+
images=images, nsfw_content_detected=has_nsfw_concept
|
1257 |
+
)
|
1258 |
+
|
1259 |
+
|
1260 |
# Static argnums are pipe, num_inference_steps. A change would trigger recompilation.
|
1261 |
# Non-static args are (sharded) input tensors mapped over their first dimension (hence, `0`).
|
1262 |
@partial(
|
1263 |
jax.pmap,
|
1264 |
+
in_axes=(None, 0, 0, 0, 0, None, 0, 0, 0, 0, 0, 0, 0, 0, None, None),
|
1265 |
+
static_broadcasted_argnums=(0, 5, 14, 15),
|
1266 |
)
|
1267 |
def _p_generate(
|
1268 |
pipe,
|
1269 |
+
prompt_ids,
|
1270 |
image,
|
1271 |
params,
|
1272 |
prng_seed,
|
|
|
1276 |
neg_prompt_ids,
|
1277 |
controlnet_conditioning_scale,
|
1278 |
xT,
|
1279 |
+
smooth_bg_strength,
|
1280 |
motion_field_strength_x,
|
1281 |
motion_field_strength_y,
|
1282 |
t0,
|
|
|
1293 |
neg_prompt_ids,
|
1294 |
controlnet_conditioning_scale,
|
1295 |
xT,
|
1296 |
+
smooth_bg_strength,
|
1297 |
motion_field_strength_x,
|
1298 |
motion_field_strength_y,
|
1299 |
t0,
|
1300 |
t1,
|
1301 |
)
|
1302 |
+
|
1303 |
+
|
1304 |
@partial(jax.pmap, static_broadcasted_argnums=(0,))
|
1305 |
def _p_get_has_nsfw_concepts(pipe, features, params):
|
1306 |
return pipe._get_has_nsfw_concepts(features, params)
|
1307 |
+
|
1308 |
+
|
1309 |
def unshard(x: jnp.ndarray):
|
1310 |
# einops.rearrange(x, 'd b ... -> (d b) ...')
|
1311 |
num_devices, batch_size = x.shape[:2]
|
1312 |
rest = x.shape[2:]
|
1313 |
return x.reshape(num_devices * batch_size, *rest)
|
1314 |
+
|
1315 |
+
|
1316 |
def preprocess(image, dtype):
|
1317 |
image = image.convert("RGB")
|
1318 |
w, h = image.size
|
|
|
1322 |
image = image[None].transpose(0, 3, 1, 2)
|
1323 |
return image
|
1324 |
|
1325 |
+
|
1326 |
+
def prepare_latents(
|
1327 |
+
params,
|
1328 |
+
prng,
|
1329 |
+
batch_size,
|
1330 |
+
num_channels_latents,
|
1331 |
+
height,
|
1332 |
+
width,
|
1333 |
+
vae_scale_factor,
|
1334 |
+
latents=None,
|
1335 |
+
):
|
1336 |
+
shape = (
|
1337 |
+
batch_size,
|
1338 |
+
num_channels_latents,
|
1339 |
+
1,
|
1340 |
+
height // vae_scale_factor,
|
1341 |
+
width // vae_scale_factor,
|
1342 |
+
) # b c f h w
|
1343 |
# scale the initial noise by the standard deviation required by the scheduler
|
1344 |
if latents is None:
|
1345 |
latents = jax.random.normal(prng, shape)
|
1346 |
latents = latents * params["scheduler"].init_noise_sigma
|
1347 |
return latents
|
1348 |
|
1349 |
+
|
1350 |
def coords_grid(batch, ht, wd):
|
1351 |
coords = jnp.meshgrid(jnp.arange(ht), jnp.arange(wd), indexing="ij")
|
1352 |
coords = jnp.stack(coords[::-1], axis=0)
|
1353 |
return coords[None].repeat(batch, 0)
|
1354 |
|
1355 |
+
|
1356 |
def adapt_pos_mirror(x, y, W, H):
|
1357 |
+
# adapt the position, with mirror padding
|
1358 |
+
x_w_mirror = ((x + W - 1) % (2 * (W - 1))) - W + 1
|
1359 |
+
x_adapted = jnp.where(x_w_mirror > 0, x_w_mirror, -(x_w_mirror))
|
1360 |
+
y_w_mirror = ((y + H - 1) % (2 * (H - 1))) - H + 1
|
1361 |
+
y_adapted = jnp.where(y_w_mirror > 0, y_w_mirror, -(y_w_mirror))
|
1362 |
+
return y_adapted, x_adapted
|
1363 |
+
|
1364 |
+
|
1365 |
+
def safe_get_zeropad(img, x, y, W, H):
|
1366 |
+
return jnp.where((x < W) & (x > 0) & (y < H) & (y > 0), img[y, x], 0.0)
|
1367 |
+
|
1368 |
|
1369 |
+
def safe_get_mirror(img, x, y, W, H):
|
1370 |
+
return img[adapt_pos_mirror(x, y, W, H)]
|
1371 |
|
|
|
|
|
1372 |
|
1373 |
@partial(jax.vmap, in_axes=(0, 0, None))
|
1374 |
@partial(jax.vmap, in_axes=(0, None, None))
|
1375 |
+
@partial(jax.vmap, in_axes=(None, 0, None))
|
1376 |
@partial(jax.vmap, in_axes=(None, 0, None))
|
1377 |
def grid_sample(latents, grid, method):
|
1378 |
# this is an alternative to torch.functional.nn.grid_sample in jax
|
1379 |
# this implementation is following the algorithm described @ https://pytorch.org/docs/stable/generated/torch.nn.functional.grid_sample.html
|
1380 |
# but with coordinates scaled to the size of the image
|
1381 |
if method == "mirror":
|
1382 |
+
return safe_get_mirror(
|
1383 |
+
latents,
|
1384 |
+
jnp.array(grid[0], dtype=jnp.int16),
|
1385 |
+
jnp.array(grid[1], dtype=jnp.int16),
|
1386 |
+
latents.shape[0],
|
1387 |
+
latents.shape[1],
|
1388 |
+
)
|
1389 |
+
else: # default is zero padding
|
1390 |
+
return safe_get_zeropad(
|
1391 |
+
latents,
|
1392 |
+
jnp.array(grid[0], dtype=jnp.int16),
|
1393 |
+
jnp.array(grid[1], dtype=jnp.int16),
|
1394 |
+
latents.shape[0],
|
1395 |
+
latents.shape[1],
|
1396 |
+
)
|
1397 |
+
|
1398 |
+
|
1399 |
+
def bandw_vid(vid, threshold):
|
1400 |
+
vid = jnp.max(vid, axis=1)
|
1401 |
+
return jnp.where(vid > threshold, 1, 0)
|
1402 |
+
|
1403 |
+
|
1404 |
+
def mean_blur(vid, k):
|
1405 |
+
window = jnp.ones((vid.shape[0], k, k)) / (k * k)
|
1406 |
+
convolve = jax.vmap(
|
1407 |
+
lambda img, kernel: jax.scipy.signal.convolve(img, kernel, mode="same")
|
1408 |
+
)
|
1409 |
+
smooth_vid = convolve(vid, window)
|
1410 |
+
return smooth_vid
|
1411 |
+
|
1412 |
+
|
1413 |
+
def get_mask_pose(vid):
|
1414 |
+
vid = bandw_vid(vid, 0.4)
|
1415 |
+
l, h, w = vid.shape
|
1416 |
+
vid = jax.image.resize(vid, (l, h // 8, w // 8), "nearest")
|
1417 |
+
vid = bandw_vid(mean_blur(vid, 7)[:, None], threshold=0.01)
|
1418 |
+
return vid / (jnp.max(vid) + 1e-4)
|
1419 |
+
# return jax.image.resize(vid/(jnp.max(vid) + 1e-4), (l, h, w), "nearest")
|
webui/app_control_animation.py
CHANGED
@@ -19,112 +19,46 @@ examples = [
|
|
19 |
]
|
20 |
|
21 |
|
22 |
-
|
23 |
-
|
24 |
-
animation_model = None
|
25 |
|
26 |
|
27 |
-
def
|
28 |
-
|
29 |
-
model_link,
|
30 |
-
is_safetensor,
|
31 |
-
frames_n_prompt,
|
32 |
-
width,
|
33 |
-
height,
|
34 |
-
cfg_scale,
|
35 |
-
seed,
|
36 |
-
):
|
37 |
-
global images
|
38 |
|
39 |
-
if not model_link:
|
40 |
-
model_link = "dreamlike-art/dreamlike-photoreal-2.0"
|
41 |
|
42 |
-
|
43 |
-
|
44 |
-
model_link,
|
45 |
-
is_safetensor,
|
46 |
-
frames_n_prompt,
|
47 |
-
width,
|
48 |
-
height,
|
49 |
-
cfg_scale,
|
50 |
-
seed,
|
51 |
-
)
|
52 |
-
|
53 |
-
return images
|
54 |
-
|
55 |
-
|
56 |
-
def select_initial_frame(evt: gr.SelectData):
|
57 |
-
global initial_frame
|
58 |
-
|
59 |
-
if evt.index < len(images):
|
60 |
-
initial_frame = images[evt.index]
|
61 |
-
print(initial_frame)
|
62 |
|
63 |
|
64 |
def create_demo(model: ControlAnimationModel):
|
65 |
-
global animation_model
|
66 |
-
animation_model = model
|
67 |
-
|
68 |
with gr.Blocks() as demo:
|
69 |
-
with gr.Column(
|
70 |
with gr.Row():
|
71 |
with gr.Column():
|
72 |
-
|
73 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
74 |
)
|
75 |
-
|
76 |
placeholder="Negative Prompt (optional)",
|
77 |
show_label=False,
|
78 |
lines=2,
|
79 |
)
|
80 |
|
81 |
-
with gr.Column():
|
82 |
-
model_link = gr.Textbox(
|
83 |
-
label="Model Link",
|
84 |
-
placeholder="dreamlike-art/dreamlike-photoreal-2.0",
|
85 |
-
info="Give the hugging face model name or URL link to safetensor.",
|
86 |
-
)
|
87 |
-
is_safetensor = gr.Checkbox(label="Safetensors")
|
88 |
gen_frames_button = gr.Button(
|
89 |
value="Generate Initial Frames", variant="primary"
|
90 |
)
|
91 |
|
92 |
-
with gr.Row():
|
93 |
-
with gr.Column(scale=2):
|
94 |
-
width = gr.Slider(32, 2048, value=512, label="Width")
|
95 |
-
height = gr.Slider(32, 2048, value=512, label="Height")
|
96 |
-
cfg_scale = gr.Slider(1, 20, value=7.0, step=0.1, label="CFG scale")
|
97 |
-
seed = gr.Slider(
|
98 |
-
label="Seed",
|
99 |
-
info="-1 for random seed on each run. Otherwise, the seed will be fixed.",
|
100 |
-
minimum=-1,
|
101 |
-
maximum=65536,
|
102 |
-
value=0,
|
103 |
-
step=1,
|
104 |
-
)
|
105 |
-
|
106 |
-
with gr.Column(scale=3):
|
107 |
-
initial_frames = gr.Gallery(
|
108 |
-
label="Initial Frames", show_label=False
|
109 |
-
).style(columns=4, object_fit="contain")
|
110 |
-
initial_frames.select(select_initial_frame)
|
111 |
-
select_frame_button = gr.Button(
|
112 |
-
value="Select Initial Frame", variant="secondary"
|
113 |
-
)
|
114 |
-
|
115 |
-
with gr.Column(visible=False) as gen_animation_col:
|
116 |
-
with gr.Row():
|
117 |
-
with gr.Column():
|
118 |
-
prompt = gr.Textbox(label="Prompt")
|
119 |
-
gen_animation_button = gr.Button(
|
120 |
-
value="Generate Animation", variant="primary"
|
121 |
-
)
|
122 |
-
|
123 |
with gr.Accordion("Advanced options", open=False):
|
124 |
-
n_prompt = gr.Textbox(
|
125 |
-
label="Negative Prompt (optional)", value=""
|
126 |
-
)
|
127 |
-
|
128 |
if on_huggingspace:
|
129 |
video_length = gr.Slider(
|
130 |
label="Video length", minimum=8, maximum=16, step=1
|
@@ -197,68 +131,101 @@ def create_demo(model: ControlAnimationModel):
|
|
197 |
)
|
198 |
|
199 |
with gr.Column():
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
200 |
result = gr.Video(label="Generated Video")
|
201 |
|
202 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
203 |
prompt,
|
|
|
|
|
204 |
model_link,
|
205 |
-
is_safetensor,
|
206 |
motion_field_strength_x,
|
207 |
motion_field_strength_y,
|
208 |
t0,
|
209 |
t1,
|
210 |
-
|
211 |
chunk_size,
|
212 |
video_length,
|
213 |
merging_ratio,
|
214 |
seed,
|
215 |
]
|
216 |
|
217 |
-
|
218 |
-
|
219 |
-
# outputs=result,
|
220 |
-
# fn=None,
|
221 |
-
# run_on_click=False,
|
222 |
-
# cache_examples=on_huggingspace,
|
223 |
-
# )
|
224 |
-
|
225 |
-
frame_inputs = [
|
226 |
-
frames_prompt,
|
227 |
-
model_link,
|
228 |
-
is_safetensor,
|
229 |
-
frames_n_prompt,
|
230 |
-
width,
|
231 |
-
height,
|
232 |
-
cfg_scale,
|
233 |
-
seed,
|
234 |
-
]
|
235 |
-
|
236 |
-
def submit_select():
|
237 |
-
show = True
|
238 |
-
if initial_frame is not None: # More to next step
|
239 |
return {
|
240 |
-
|
241 |
-
|
242 |
}
|
243 |
|
244 |
return {
|
245 |
-
|
246 |
-
|
247 |
}
|
248 |
|
249 |
gen_frames_button.click(
|
250 |
-
generate_initial_frames,
|
251 |
inputs=frame_inputs,
|
252 |
outputs=initial_frames,
|
253 |
)
|
254 |
-
select_frame_button.click(
|
255 |
-
submit_select, inputs=None, outputs=[frame_selection_col, gen_animation_col]
|
256 |
-
)
|
257 |
|
258 |
gen_animation_button.click(
|
259 |
-
fn=
|
260 |
-
inputs=
|
|
|
|
|
|
|
|
|
261 |
outputs=result,
|
262 |
)
|
263 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
264 |
return demo
|
|
|
19 |
]
|
20 |
|
21 |
|
22 |
+
def on_video_path_update(evt: gr.EventData):
|
23 |
+
return f"Selection: **{evt._data}**"
|
|
|
24 |
|
25 |
|
26 |
+
def pose_gallery_callback(evt: gr.SelectData):
|
27 |
+
return f"Motion {evt.index+1}"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
28 |
|
|
|
|
|
29 |
|
30 |
+
def get_frame_index(evt: gr.SelectData):
|
31 |
+
return evt.index
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
32 |
|
33 |
|
34 |
def create_demo(model: ControlAnimationModel):
|
|
|
|
|
|
|
35 |
with gr.Blocks() as demo:
|
36 |
+
with gr.Column():
|
37 |
with gr.Row():
|
38 |
with gr.Column():
|
39 |
+
# TODO: update so that model_link is customizable
|
40 |
+
model_link = gr.Dropdown(
|
41 |
+
label="Model Link",
|
42 |
+
choices=["runwayml/stable-diffusion-v1-5"],
|
43 |
+
value="runwayml/stable-diffusion-v1-5",
|
44 |
+
)
|
45 |
+
prompt = gr.Textbox(
|
46 |
+
placeholder="Prompt",
|
47 |
+
show_label=False,
|
48 |
+
lines=2,
|
49 |
+
info="Give a prompt for an animation you would like to generate. The prompt will be used to create the first initial frame and then the animation.",
|
50 |
)
|
51 |
+
negative_prompt = gr.Textbox(
|
52 |
placeholder="Negative Prompt (optional)",
|
53 |
show_label=False,
|
54 |
lines=2,
|
55 |
)
|
56 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
57 |
gen_frames_button = gr.Button(
|
58 |
value="Generate Initial Frames", variant="primary"
|
59 |
)
|
60 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
61 |
with gr.Accordion("Advanced options", open=False):
|
|
|
|
|
|
|
|
|
62 |
if on_huggingspace:
|
63 |
video_length = gr.Slider(
|
64 |
label="Video length", minimum=8, maximum=16, step=1
|
|
|
131 |
)
|
132 |
|
133 |
with gr.Column():
|
134 |
+
gallery_pose_sequence = gr.Gallery(
|
135 |
+
label="Pose Sequence",
|
136 |
+
value=[
|
137 |
+
("__assets__/dance1.gif", "Motion 1"),
|
138 |
+
("__assets__/dance2.gif", "Motion 2"),
|
139 |
+
("__assets__/dance3.gif", "Motion 3"),
|
140 |
+
("__assets__/dance4.gif", "Motion 4"),
|
141 |
+
("__assets__/dance5.gif", "Motion 5"),
|
142 |
+
],
|
143 |
+
).style(columns=3)
|
144 |
+
input_video_path = gr.Textbox(
|
145 |
+
label="Pose Sequence", visible=False, value="Motion 1"
|
146 |
+
)
|
147 |
+
pose_sequence_selector = gr.Markdown("Pose Sequence: **Motion 1**")
|
148 |
+
|
149 |
+
with gr.Row():
|
150 |
+
with gr.Column(visible=True) as frame_selection_view:
|
151 |
+
initial_frames = gr.Gallery(
|
152 |
+
label="Initial Frames", show_label=False
|
153 |
+
).style(columns=4, rows=1, object_fit="contain", preview=True)
|
154 |
+
|
155 |
+
gr.Markdown("Select an initial frame to start your animation with.")
|
156 |
+
gen_animation_button = gr.Button(
|
157 |
+
value="Select Initial Frame & Generate Animation",
|
158 |
+
variant="secondary",
|
159 |
+
)
|
160 |
+
|
161 |
+
with gr.Column(visible=False) as animation_view:
|
162 |
result = gr.Video(label="Generated Video")
|
163 |
|
164 |
+
with gr.Box(visible=False):
|
165 |
+
initial_frame_index = gr.Number(
|
166 |
+
label="Selected Initial Frame Index", value=-1, precision=0
|
167 |
+
)
|
168 |
+
|
169 |
+
input_video_path.change(on_video_path_update, None, pose_sequence_selector)
|
170 |
+
gallery_pose_sequence.select(pose_gallery_callback, None, input_video_path)
|
171 |
+
initial_frames.select(fn=get_frame_index, outputs=initial_frame_index)
|
172 |
+
|
173 |
+
frame_inputs = [
|
174 |
+
prompt,
|
175 |
+
input_video_path,
|
176 |
+
negative_prompt,
|
177 |
+
]
|
178 |
+
|
179 |
+
animation_inputs = [
|
180 |
prompt,
|
181 |
+
initial_frame_index,
|
182 |
+
input_video_path,
|
183 |
model_link,
|
|
|
184 |
motion_field_strength_x,
|
185 |
motion_field_strength_y,
|
186 |
t0,
|
187 |
t1,
|
188 |
+
negative_prompt,
|
189 |
chunk_size,
|
190 |
video_length,
|
191 |
merging_ratio,
|
192 |
seed,
|
193 |
]
|
194 |
|
195 |
+
def submit_select(initial_frame_index: int):
|
196 |
+
if initial_frame_index != -1: # More to next step
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
197 |
return {
|
198 |
+
frame_selection_view: gr.update(visible=False),
|
199 |
+
animation_view: gr.update(visible=True),
|
200 |
}
|
201 |
|
202 |
return {
|
203 |
+
frame_selection_view: gr.update(visible=True),
|
204 |
+
animation_view: gr.update(visible=False),
|
205 |
}
|
206 |
|
207 |
gen_frames_button.click(
|
208 |
+
fn=model.generate_initial_frames,
|
209 |
inputs=frame_inputs,
|
210 |
outputs=initial_frames,
|
211 |
)
|
|
|
|
|
|
|
212 |
|
213 |
gen_animation_button.click(
|
214 |
+
fn=submit_select,
|
215 |
+
inputs=initial_frame_index,
|
216 |
+
outputs=[frame_selection_view, animation_view],
|
217 |
+
).then(
|
218 |
+
fn=None,
|
219 |
+
inputs=animation_inputs,
|
220 |
outputs=result,
|
221 |
)
|
222 |
|
223 |
+
# gr.Examples(examples=examples,
|
224 |
+
# inputs=inputs,
|
225 |
+
# outputs=result,
|
226 |
+
# fn=None,
|
227 |
+
# run_on_click=False,
|
228 |
+
# cache_examples=on_huggingspace,
|
229 |
+
# )
|
230 |
+
|
231 |
return demo
|