-
-
Notifications
You must be signed in to change notification settings - Fork 890
/
txt2vid.py
2498 lines (2179 loc) · 106 KB
/
txt2vid.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
# This file is part of sygil-webui (https://github.com/Sygil-Dev/sygil-webui/).
# Copyright 2022 Sygil-Dev team.
# This program is free software: you can redistribute it and/or modify
# it under the terms of the GNU Affero General Public License as published by
# the Free Software Foundation, either version 3 of the License, or
# (at your option) any later version.
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU Affero General Public License for more details.
# You should have received a copy of the GNU Affero General Public License
# along with this program. If not, see <http://www.gnu.org/licenses/>.
# base webui import and utils.
"""
Implementation of Text to Video based on the
https://github.com/nateraw/stable-diffusion-videos
repo and the original gist script from
https://gist.github.com/karpathy/00103b0037c5aaea32fe1da1af553355
"""
from sd_utils import (
st,
MemUsageMonitor,
server_state,
no_rerun,
torch_gc,
custom_models_available,
RealESRGAN_available,
GFPGAN_available,
LDSR_available,
hc,
seed_to_int,
logger,
slerp,
optimize_update_preview_frequency,
load_learned_embed_in_clip,
load_GFPGAN,
RealESRGANModel,
set_page_title,
)
# streamlit imports
from streamlit.runtime.scriptrunner import StopException
# from streamlit.elements import image as STImage
# streamlit components section
from streamlit_server_state import server_state, server_state_lock
# from streamlitextras.threader import lock, trigger_rerun, \
# streamlit_thread, get_thread, \
# last_trigger_time
# other imports
import os, sys, json, re, random, datetime, time, warnings, mimetypes
from PIL import Image
import torch
import numpy as np
import time, inspect, timeit
import torch
from torch import autocast
# from io import BytesIO
import imageio
from slugify import slugify
from diffusers import StableDiffusionPipeline, DiffusionPipeline
# from stable_diffusion_videos import StableDiffusionWalkPipeline
from diffusers.schedulers import (
DDIMScheduler,
LMSDiscreteScheduler,
PNDMScheduler,
DDPMScheduler,
)
from diffusers.configuration_utils import FrozenDict
from diffusers.models import AutoencoderKL, UNet2DConditionModel
from diffusers.pipelines.stable_diffusion.safety_checker import (
StableDiffusionSafetyChecker,
)
from diffusers.utils import deprecate
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
from typing import Callable, List, Optional, Union
from pathlib import Path
from torchvision.transforms.functional import pil_to_tensor
from torchvision import transforms
import librosa
from PIL import Image
from torchvision.io import write_video
from torchvision import transforms
import torch.nn as nn
from uuid import uuid4
# streamlit components
from custom_components import sygil_suggestions
# Temp imports
# end of imports
# ---------------------------------------------------------------------------------------------------------------
sygil_suggestions.init()
try:
# this silences the annoying "Some weights of the model checkpoint were not used when initializing..." message at start.
from transformers import logging
logging.set_verbosity_error()
except:
pass
# remove some annoying deprecation warnings that show every now and then.
warnings.filterwarnings("ignore", category=DeprecationWarning)
warnings.filterwarnings("ignore", category=UserWarning)
# this is a fix for Windows users. Without it, javascript files will be served with text/html content-type and the bowser will not show any UI
mimetypes.init()
mimetypes.add_type("application/javascript", ".js")
class plugin_info:
plugname = "txt2vid"
description = "Text to Image"
isTab = True
displayPriority = 1
#
# -----------------------------------------------------------------------------
def txt2vid_generation_callback(step: int, timestep: int, latents: torch.FloatTensor):
# print ("test")
# scale and decode the image latents with vae
cond_latents_2 = 1 / 0.18215 * latents
image = server_state["pipe"].vae.decode(cond_latents_2)
# generate output numpy image as uint8
image = torch.clamp((image["sample"] + 1.0) / 2.0, min=0.0, max=1.0)
image2 = transforms.ToPILImage()(image.squeeze_(0))
st.session_state["preview_image"].image(image2)
def get_timesteps_arr(audio_filepath, offset, duration, fps=30, margin=1.0, smooth=0.0):
y, sr = librosa.load(audio_filepath, offset=offset, duration=duration)
# librosa.stft hardcoded defaults...
# n_fft defaults to 2048
# hop length is win_length // 4
# win_length defaults to n_fft
D = librosa.stft(y, n_fft=2048, hop_length=2048 // 4, win_length=2048)
# Extract percussive elements
D_harmonic, D_percussive = librosa.decompose.hpss(D, margin=margin)
y_percussive = librosa.istft(D_percussive, length=len(y))
# Get normalized melspectrogram
spec_raw = librosa.feature.melspectrogram(y=y_percussive, sr=sr)
spec_max = np.amax(spec_raw, axis=0)
spec_norm = (spec_max - np.min(spec_max)) / np.ptp(spec_max)
# Resize cumsum of spec norm to our desired number of interpolation frames
x_norm = np.linspace(0, spec_norm.shape[-1], spec_norm.shape[-1])
y_norm = np.cumsum(spec_norm)
y_norm /= y_norm[-1]
x_resize = np.linspace(0, y_norm.shape[-1], int(duration * fps))
T = np.interp(x_resize, x_norm, y_norm)
# Apply smoothing
return T * (1 - smooth) + np.linspace(0.0, 1.0, T.shape[0]) * smooth
#
def make_video_pyav(
frames_or_frame_dir: Union[str, Path, torch.Tensor],
audio_filepath: Union[str, Path] = None,
fps: int = 30,
audio_offset: int = 0,
audio_duration: int = 2,
sr: int = 22050,
output_filepath: Union[str, Path] = "output.mp4",
glob_pattern: str = "*.png",
):
"""
TODO - docstring here
frames_or_frame_dir: (Union[str, Path, torch.Tensor]):
Either a directory of images, or a tensor of shape (T, C, H, W) in range [0, 255].
"""
# Torchvision write_video doesn't support pathlib paths
output_filepath = str(output_filepath)
if isinstance(frames_or_frame_dir, (str, Path)):
frames = None
for img in sorted(Path(frames_or_frame_dir).glob(glob_pattern)):
frame = pil_to_tensor(Image.open(img)).unsqueeze(0)
frames = frame if frames is None else torch.cat([frames, frame])
else:
frames = frames_or_frame_dir
# TCHW -> THWC
frames = frames.permute(0, 2, 3, 1)
if audio_filepath:
# Read audio, convert to tensor
audio, sr = librosa.load(
audio_filepath,
sr=sr,
mono=True,
offset=audio_offset,
duration=audio_duration,
)
audio_tensor = torch.tensor(audio).unsqueeze(0)
write_video(
output_filepath,
frames,
fps=fps,
audio_array=audio_tensor,
audio_fps=sr,
audio_codec="aac",
options={"crf": "10", "pix_fmt": "yuv420p"},
)
else:
write_video(
output_filepath,
frames,
fps=fps,
options={"crf": "10", "pix_fmt": "yuv420p"},
)
return output_filepath
class StableDiffusionWalkPipeline(DiffusionPipeline):
r"""
Pipeline for generating videos by interpolating Stable Diffusion's latent space.
This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
Args:
vae ([`AutoencoderKL`]):
Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
text_encoder ([`CLIPTextModel`]):
Frozen text-encoder. Stable Diffusion uses the text portion of
[CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
tokenizer (`CLIPTokenizer`):
Tokenizer of class
[CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
scheduler ([`SchedulerMixin`]):
A scheduler to be used in combination with `unet` to denoise the encoded image latens. Can be one of
[`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
safety_checker ([`StableDiffusionSafetyChecker`]):
Classification module that estimates whether generated images could be considered offensive or harmful.
Please, refer to the [model card](https://huggingface.co/CompVis/stable-diffusion-v1-4) for details.
feature_extractor ([`CLIPFeatureExtractor`]):
Model that extracts features from generated images to be used as inputs for the `safety_checker`.
"""
def __init__(
self,
vae: AutoencoderKL,
text_encoder: CLIPTextModel,
tokenizer: CLIPTokenizer,
unet: UNet2DConditionModel,
scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
safety_checker: StableDiffusionSafetyChecker,
feature_extractor: CLIPFeatureExtractor,
):
super().__init__()
if (
hasattr(scheduler.config, "steps_offset")
and scheduler.config.steps_offset != 1
):
deprecation_message = (
f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
"to update the config accordingly as leaving `steps_offset` might led to incorrect results"
" in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
" it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
" file"
)
deprecate(
"steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False
)
new_config = dict(scheduler.config)
new_config["steps_offset"] = 1
scheduler._internal_dict = FrozenDict(new_config)
self.register_modules(
vae=vae,
text_encoder=text_encoder,
tokenizer=tokenizer,
unet=unet,
scheduler=scheduler,
safety_checker=safety_checker,
feature_extractor=feature_extractor,
)
def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
r"""
Enable sliced attention computation.
When this option is enabled, the attention module will split the input tensor in slices, to compute attention
in several steps. This is useful to save some memory in exchange for a small speed decrease.
Args:
slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
`attention_head_dim` must be a multiple of `slice_size`.
"""
if slice_size == "auto":
# half the attention head size is usually a good trade-off between
# speed and memory
slice_size = self.unet.config.attention_head_dim // 2
self.unet.set_attention_slice(slice_size)
def disable_attention_slicing(self):
r"""
Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
back to computing attention in one step.
"""
# set slice_size = `None` to disable `attention slicing`
self.enable_attention_slicing(None)
@torch.no_grad()
def __call__(
self,
prompt: Optional[Union[str, List[str]]] = None,
height: int = 512,
width: int = 512,
num_inference_steps: int = 50,
guidance_scale: float = 7.5,
negative_prompt: Optional[Union[str, List[str]]] = None,
num_images_per_prompt: Optional[int] = 1,
eta: float = 0.0,
generator: Optional[torch.Generator] = None,
latents: Optional[torch.FloatTensor] = None,
output_type: Optional[str] = "pil",
return_dict: bool = True,
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
callback_steps: Optional[int] = 1,
text_embeddings: Optional[torch.FloatTensor] = None,
**kwargs,
):
r"""
Function invoked when calling the pipeline for generation.
Args:
prompt (`str` or `List[str]`, *optional*, defaults to `None`):
The prompt or prompts to guide the image generation. If not provided, `text_embeddings` is required.
height (`int`, *optional*, defaults to 512):
The height in pixels of the generated image.
width (`int`, *optional*, defaults to 512):
The width in pixels of the generated image.
num_inference_steps (`int`, *optional*, defaults to 50):
The number of denoising steps. More denoising steps usually lead to a higher quality image at the
expense of slower inference.
guidance_scale (`float`, *optional*, defaults to 7.5):
Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
`guidance_scale` is defined as `w` of equation 2. of [Imagen
Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
usually at the expense of lower image quality.
negative_prompt (`str` or `List[str]`, *optional*):
The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
if `guidance_scale` is less than `1`).
num_images_per_prompt (`int`, *optional*, defaults to 1):
The number of images to generate per prompt.
eta (`float`, *optional*, defaults to 0.0):
Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
[`schedulers.DDIMScheduler`], will be ignored for others.
generator (`torch.Generator`, *optional*):
A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
deterministic.
latents (`torch.FloatTensor`, *optional*):
Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
tensor will ge generated by sampling using the supplied random `generator`.
output_type (`str`, *optional*, defaults to `"pil"`):
The output format of the generate image. Choose between
[PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
return_dict (`bool`, *optional*, defaults to `True`):
Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
plain tuple.
callback (`Callable`, *optional*):
A function that will be called every `callback_steps` steps during inference. The function will be
called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
callback_steps (`int`, *optional*, defaults to 1):
The frequency at which the `callback` function will be called. If not specified, the callback will be
called at every step.
text_embeddings (`torch.FloatTensor`, *optional*, defaults to `None`):
Pre-generated text embeddings to be used as inputs for image generation. Can be used in place of
`prompt` to avoid re-computing the embeddings. If not provided, the embeddings will be generated from
the supplied `prompt`.
Returns:
[`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
[`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
When returning a tuple, the first element is a list with the generated images, and the second element is a
list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
(nsfw) content, according to the `safety_checker`.
"""
if height % 8 != 0 or width % 8 != 0:
raise ValueError(
f"`height` and `width` have to be divisible by 8 but are {height} and {width}."
)
if (callback_steps is None) or (
callback_steps is not None
and (not isinstance(callback_steps, int) or callback_steps <= 0)
):
raise ValueError(
f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
f" {type(callback_steps)}."
)
if text_embeddings is None:
if isinstance(prompt, str):
batch_size = 1
elif isinstance(prompt, list):
batch_size = len(prompt)
else:
raise ValueError(
f"`prompt` has to be of type `str` or `list` but is {type(prompt)}"
)
# get prompt text embeddings
text_inputs = self.tokenizer(
prompt,
padding="max_length",
max_length=self.tokenizer.model_max_length,
return_tensors="pt",
)
text_input_ids = text_inputs.input_ids
if text_input_ids.shape[-1] > self.tokenizer.model_max_length:
removed_text = self.tokenizer.batch_decode(
text_input_ids[:, self.tokenizer.model_max_length :]
)
print(
"The following part of your input was truncated because CLIP can only handle sequences up to"
f" {self.tokenizer.model_max_length} tokens: {removed_text}"
)
text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length]
text_embeddings = self.text_encoder(text_input_ids.to(self.device))[0]
else:
batch_size = text_embeddings.shape[0]
# duplicate text embeddings for each generation per prompt, using mps friendly method
bs_embed, seq_len, _ = text_embeddings.shape
text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1)
text_embeddings = text_embeddings.view(
bs_embed * num_images_per_prompt, seq_len, -1
)
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
# of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
# corresponds to doing no classifier free guidance.
do_classifier_free_guidance = guidance_scale > 1.0
# get unconditional embeddings for classifier free guidance
if do_classifier_free_guidance:
uncond_tokens: List[str]
if negative_prompt is None:
uncond_tokens = [""]
elif type(prompt) is not type(negative_prompt):
raise TypeError(
f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
f" {type(prompt)}."
)
elif isinstance(negative_prompt, str):
uncond_tokens = [negative_prompt]
elif batch_size != len(negative_prompt):
raise ValueError(
f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
" the batch size of `prompt`."
)
else:
uncond_tokens = negative_prompt
max_length = self.tokenizer.model_max_length
uncond_input = self.tokenizer(
uncond_tokens,
padding="max_length",
max_length=max_length,
truncation=True,
return_tensors="pt",
)
uncond_embeddings = self.text_encoder(
uncond_input.input_ids.to(self.device)
)[0]
# duplicate unconditional embeddings for each generation per prompt, using mps friendly method
seq_len = uncond_embeddings.shape[1]
uncond_embeddings = uncond_embeddings.repeat(
batch_size, num_images_per_prompt, 1
)
uncond_embeddings = uncond_embeddings.view(
batch_size * num_images_per_prompt, seq_len, -1
)
# For classifier free guidance, we need to do two forward passes.
# Here we concatenate the unconditional and text embeddings into a single batch
# to avoid doing two forward passes
text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
# get the initial random noise unless the user supplied it
# Unlike in other pipelines, latents need to be generated in the target device
# for 1-to-1 results reproducibility with the CompVis implementation.
# However this currently doesn't work in `mps`.
latents_shape = (
batch_size * num_images_per_prompt,
self.unet.in_channels,
height // 8,
width // 8,
)
latents_dtype = text_embeddings.dtype
if latents is None:
if self.device.type == "mps":
# randn does not exist on mps
latents = torch.randn(
latents_shape,
generator=generator,
device="cpu",
dtype=latents_dtype,
).to(self.device)
else:
latents = torch.randn(
latents_shape,
generator=generator,
device=self.device,
dtype=latents_dtype,
)
else:
if latents.shape != latents_shape:
raise ValueError(
f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}"
)
latents = latents.to(self.device)
# set timesteps
self.scheduler.set_timesteps(num_inference_steps)
# Some schedulers like PNDM have timesteps as arrays
# It's more optimized to move all timesteps to correct device beforehand
timesteps_tensor = self.scheduler.timesteps.to(self.device)
# scale the initial noise by the standard deviation required by the scheduler
latents = latents * self.scheduler.init_noise_sigma
# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
# eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
# eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
# and should be between [0, 1]
accepts_eta = "eta" in set(
inspect.signature(self.scheduler.step).parameters.keys()
)
extra_step_kwargs = {}
if accepts_eta:
extra_step_kwargs["eta"] = eta
for i, t in enumerate(self.progress_bar(timesteps_tensor)):
# expand the latents if we are doing classifier free guidance
latent_model_input = (
torch.cat([latents] * 2) if do_classifier_free_guidance else latents
)
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
# predict the noise residual
noise_pred = self.unet(
latent_model_input, t, encoder_hidden_states=text_embeddings
).sample
# perform guidance
if do_classifier_free_guidance:
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
noise_pred = noise_pred_uncond + guidance_scale * (
noise_pred_text - noise_pred_uncond
)
# compute the previous noisy sample x_t -> x_t-1
latents = self.scheduler.step(
noise_pred, t, latents, **extra_step_kwargs
).prev_sample
# call the callback, if provided
if callback is not None and i % callback_steps == 0:
callback(i, t, latents)
print("test")
latents = 1 / 0.18215 * latents
image = self.vae.decode(latents).sample
image = (image / 2 + 0.5).clamp(0, 1)
# we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
image = image.cpu().permute(0, 2, 3, 1).float().numpy()
if self.safety_checker is not None:
safety_checker_input = self.feature_extractor(
self.numpy_to_pil(image), return_tensors="pt"
).to(self.device)
image, has_nsfw_concept = self.safety_checker(
images=image,
clip_input=safety_checker_input.pixel_values.to(text_embeddings.dtype),
)
else:
has_nsfw_concept = None
if output_type == "pil":
image = self.numpy_to_pil(image)
if not return_dict:
return (image, has_nsfw_concept)
return StableDiffusionPipelineOutput(
images=image, nsfw_content_detected=has_nsfw_concept
)
def generate_inputs(
self, prompt_a, prompt_b, seed_a, seed_b, noise_shape, T, batch_size
):
embeds_a = self.embed_text(prompt_a)
embeds_b = self.embed_text(prompt_b)
latents_a = self.init_noise(seed_a, noise_shape)
latents_b = self.init_noise(seed_b, noise_shape)
batch_idx = 0
embeds_batch, noise_batch = None, None
for i, t in enumerate(T):
embeds = torch.lerp(embeds_a, embeds_b, t)
noise = slerp(
device="cuda",
t=float(t),
v0=latents_a,
v1=latents_b,
DOT_THRESHOLD=0.9995,
)
embeds_batch = (
embeds if embeds_batch is None else torch.cat([embeds_batch, embeds])
)
noise_batch = (
noise if noise_batch is None else torch.cat([noise_batch, noise])
)
batch_is_ready = embeds_batch.shape[0] == batch_size or i + 1 == T.shape[0]
if not batch_is_ready:
continue
yield batch_idx, embeds_batch, noise_batch
batch_idx += 1
del embeds_batch, noise_batch
torch.cuda.empty_cache()
embeds_batch, noise_batch = None, None
def make_clip_frames(
self,
prompt_a: str,
prompt_b: str,
seed_a: int,
seed_b: int,
num_interpolation_steps: int = 5,
save_path: Union[str, Path] = "outputs/",
num_inference_steps: int = 50,
guidance_scale: float = 7.5,
eta: float = 0.0,
height: int = 512,
width: int = 512,
upsample: bool = False,
batch_size: int = 1,
image_file_ext: str = ".png",
T: np.ndarray = None,
skip: int = 0,
callback=None,
callback_steps: int = 1,
):
save_path = Path(save_path)
save_path.mkdir(parents=True, exist_ok=True)
T = T if T is not None else np.linspace(0.0, 1.0, num_interpolation_steps)
if T.shape[0] != num_interpolation_steps:
raise ValueError(
f"Unexpected T shape, got {T.shape}, expected dim 0 to be {num_interpolation_steps}"
)
if upsample:
if getattr(self, "upsampler", None) is None:
self.upsampler = RealESRGANModel.from_pretrained("nateraw/real-esrgan")
self.upsampler.to(self.device)
batch_generator = self.generate_inputs(
prompt_a,
prompt_b,
seed_a,
seed_b,
(1, self.unet.in_channels, height // 8, width // 8),
T[skip:],
batch_size,
)
frame_index = skip
for _, embeds_batch, noise_batch in batch_generator:
with torch.autocast("cuda"):
outputs = self(
latents=noise_batch,
text_embeddings=embeds_batch,
height=height,
width=width,
guidance_scale=guidance_scale,
eta=eta,
num_inference_steps=num_inference_steps,
output_type="pil" if not upsample else "numpy",
callback=callback,
callback_steps=callback_steps,
)["images"]
for image in outputs:
frame_filepath = save_path / (
f"frame%06d{image_file_ext}" % frame_index
)
image = image if not upsample else self.upsampler(image)
image.save(frame_filepath)
frame_index += 1
def walk(
self,
prompt: Optional[List[str]] = None,
seeds: Optional[List[int]] = None,
num_interpolation_steps: Optional[
Union[int, List[int]]
] = 5, # int or list of int
output_dir: Optional[str] = "./dreams",
name: Optional[str] = None,
image_file_ext: Optional[str] = ".png",
fps: Optional[int] = 30,
num_inference_steps: Optional[int] = 50,
guidance_scale: Optional[float] = 7.5,
eta: Optional[float] = 0.0,
height: Optional[int] = 512,
width: Optional[int] = 512,
upsample: Optional[bool] = False,
batch_size: Optional[int] = 1,
resume: Optional[bool] = False,
audio_filepath: str = None,
audio_start_sec: Optional[Union[int, float]] = None,
margin: Optional[float] = 1.0,
smooth: Optional[float] = 0.0,
callback=None,
callback_steps=1,
):
"""Generate a video from a sequence of prompts and seeds. Optionally, add audio to the
video to interpolate to the intensity of the audio.
Args:
prompts (Optional[List[str]], optional):
list of text prompts. Defaults to None.
seeds (Optional[List[int]], optional):
list of random seeds corresponding to prompts. Defaults to None.
num_interpolation_steps (Union[int, List[int]], *optional*):
How many interpolation steps between each prompt. Defaults to None.
output_dir (Optional[str], optional):
Where to save the video. Defaults to './dreams'.
name (Optional[str], optional):
Name of the subdirectory of output_dir. Defaults to None.
image_file_ext (Optional[str], *optional*, defaults to '.png'):
The extension to use when writing video frames.
fps (Optional[int], *optional*, defaults to 30):
The frames per second in the resulting output videos.
num_inference_steps (Optional[int], *optional*, defaults to 50):
The number of denoising steps. More denoising steps usually lead to a higher quality image at the
expense of slower inference.
guidance_scale (Optional[float], *optional*, defaults to 7.5):
Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
`guidance_scale` is defined as `w` of equation 2. of [Imagen
Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
usually at the expense of lower image quality.
eta (Optional[float], *optional*, defaults to 0.0):
Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
[`schedulers.DDIMScheduler`], will be ignored for others.
height (Optional[int], *optional*, defaults to 512):
height of the images to generate.
width (Optional[int], *optional*, defaults to 512):
width of the images to generate.
upsample (Optional[bool], *optional*, defaults to False):
When True, upsamples images with realesrgan.
batch_size (Optional[int], *optional*, defaults to 1):
Number of images to generate at once.
resume (Optional[bool], *optional*, defaults to False):
When True, resumes from the last frame in the output directory based
on available prompt config. Requires you to provide the `name` argument.
audio_filepath (str, *optional*, defaults to None):
Optional path to an audio file to influence the interpolation rate.
audio_start_sec (Optional[Union[int, float]], *optional*, defaults to 0):
Global start time of the provided audio_filepath.
margin (Optional[float], *optional*, defaults to 1.0):
Margin from librosa hpss to use for audio interpolation.
smooth (Optional[float], *optional*, defaults to 0.0):
Smoothness of the audio interpolation. 1.0 means linear interpolation.
This function will create sub directories for each prompt and seed pair.
For example, if you provide the following prompts and seeds:
```
prompts = ['a dog', 'a cat', 'a bird']
seeds = [1, 2, 3]
num_interpolation_steps = 5
output_dir = 'output_dir'
name = 'name'
fps = 5
```
Then the following directories will be created:
```
output_dir
├── name
│ ├── name_000000
│ │ ├── frame000000.png
│ │ ├── ...
│ │ ├── frame000004.png
│ │ ├── name_000000.mp4
│ ├── name_000001
│ │ ├── frame000000.png
│ │ ├── ...
│ │ ├── frame000004.png
│ │ ├── name_000001.mp4
│ ├── ...
│ ├── name.mp4
| |── prompt_config.json
```
Returns:
str: The resulting video filepath. This video includes all sub directories' video clips.
"""
if (callback_steps is None) or (
callback_steps is not None
and (not isinstance(callback_steps, int) or callback_steps <= 0)
):
raise ValueError(
f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
f" {type(callback_steps)}."
)
# init the output dir
if type(prompts) == str:
sanitized_prompt = slugify(prompts)
else:
sanitized_prompt = slugify(prompts[0])
full_path = os.path.join(str(output_dir), str(sanitized_prompt))
if len(full_path) > 220:
sanitized_prompt = sanitized_prompt[: 220 - len(full_path)]
full_path = os.path.join(output_dir, sanitized_prompt)
os.makedirs(full_path, exist_ok=True)
# Where the final video of all the clips combined will be saved
output_filepath = os.path.join(full_path, f"{sanitized_prompt}.mp4")
# If using same number of interpolation steps between, we turn into list
if not resume and isinstance(num_interpolation_steps, int):
num_interpolation_steps = [num_interpolation_steps] * (len(prompts) - 1)
if not resume:
audio_start_sec = audio_start_sec or 0
# Save/reload prompt config
prompt_config_path = Path(os.path.join(full_path, "prompt_config.json"))
if not resume:
prompt_config_path.write_text(
json.dumps(
dict(
prompts=prompts,
seeds=seeds,
num_interpolation_steps=num_interpolation_steps,
fps=fps,
num_inference_steps=num_inference_steps,
guidance_scale=guidance_scale,
eta=eta,
upsample=upsample,
height=height,
width=width,
audio_filepath=audio_filepath,
audio_start_sec=audio_start_sec,
),
indent=2,
sort_keys=False,
)
)
else:
data = json.load(open(prompt_config_path))
prompts = data["prompts"]
seeds = data["seeds"]
num_interpolation_steps = data["num_interpolation_steps"]
fps = data["fps"]
num_inference_steps = data["num_inference_steps"]
guidance_scale = data["guidance_scale"]
eta = data["eta"]
upsample = data["upsample"]
height = data["height"]
width = data["width"]
audio_filepath = data["audio_filepath"]
audio_start_sec = data["audio_start_sec"]
for i, (prompt_a, prompt_b, seed_a, seed_b, num_step) in enumerate(
zip(prompts, prompts[1:], seeds, seeds[1:], num_interpolation_steps)
):
# {name}_000000 / {name}_000001 / ...
save_path = Path(f"{full_path}/{name}_{i:06d}")
# Where the individual clips will be saved
step_output_filepath = Path(f"{save_path}/{name}_{i:06d}.mp4")
# Determine if we need to resume from a previous run
skip = 0
if resume:
if step_output_filepath.exists():
print(f"Skipping {save_path} because frames already exist")
continue
existing_frames = sorted(save_path.glob(f"*{image_file_ext}"))
if existing_frames:
skip = int(existing_frames[-1].stem[-6:]) + 1
if skip + 1 >= num_step:
print(f"Skipping {save_path} because frames already exist")
continue
print(f"Resuming {save_path.name} from frame {skip}")
audio_offset = audio_start_sec + sum(num_interpolation_steps[:i]) / fps
audio_duration = num_step / fps
self.make_clip_frames(
prompt_a,
prompt_b,
seed_a,
seed_b,
num_interpolation_steps=num_step,
save_path=save_path,
num_inference_steps=num_inference_steps,
guidance_scale=guidance_scale,
eta=eta,
height=height,
width=width,
upsample=upsample,
batch_size=batch_size,
skip=skip,
T=get_timesteps_arr(
audio_filepath,
offset=audio_offset,
duration=audio_duration,
fps=fps,
margin=margin,
smooth=smooth,
callback=callback,
callback_steps=callback_steps,
)
if audio_filepath
else None,
)
make_video_pyav(
save_path,
audio_filepath=audio_filepath,
fps=fps,
output_filepath=step_output_filepath,
glob_pattern=f"*{image_file_ext}",
audio_offset=audio_offset,
audio_duration=audio_duration,
sr=44100,
)
return make_video_pyav(
full_path,
audio_filepath=audio_filepath,
fps=fps,
audio_offset=audio_start_sec,
audio_duration=sum(num_interpolation_steps) / fps,
output_filepath=output_filepath,
glob_pattern=f"**/*{image_file_ext}",
sr=44100,
)
def embed_text(self, text):
"""Helper to embed some text"""