forked from AUTOMATIC1111/stable-diffusion-webui
-
Notifications
You must be signed in to change notification settings - Fork 1
/
webui.py
2036 lines (1544 loc) · 76.8 KB
/
webui.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
import argparse
import os
import sys
script_path = os.path.dirname(os.path.realpath(__file__))
sd_path = os.path.dirname(script_path)
# add parent directory to path; this is where Stable diffusion repo should be
path_dirs = [
(sd_path, 'ldm', 'Stable Diffusion'),
('../../taming-transformers', 'taming', 'Taming Transformers')
]
for d, must_exist, what in path_dirs:
must_exist_path = os.path.abspath(os.path.join(script_path, d, must_exist))
if not os.path.exists(must_exist_path):
print(f"Warning: {what} not found at path {must_exist_path}", file=sys.stderr)
else:
sys.path.append(os.path.join(script_path, d))
import torch
import torch.nn as nn
import numpy as np
import gradio as gr
import gradio.utils
from omegaconf import OmegaConf
from PIL import Image, ImageFont, ImageDraw, PngImagePlugin, ImageFilter, ImageOps
from torch import autocast
import mimetypes
import random
import math
import html
import time
import json
import traceback
from collections import namedtuple
from contextlib import nullcontext
import signal
import tqdm
import re
import threading
import k_diffusion.sampling
from ldm.util import instantiate_from_config
from ldm.models.diffusion.ddim import DDIMSampler
from ldm.models.diffusion.plms import PLMSSampler
# 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')
# some of those options should not be changed at all because they would break the model, so I removed them from options.
opt_C = 4
opt_f = 8
LANCZOS = (Image.Resampling.LANCZOS if hasattr(Image, 'Resampling') else Image.LANCZOS)
invalid_filename_chars = '<>:"/\\|?*\n'
config_filename = "config.json"
parser = argparse.ArgumentParser()
parser.add_argument("--config", type=str, default=os.path.join(sd_path, "configs/stable-diffusion/v1-inference.yaml"), help="path to config which constructs model",)
parser.add_argument("--ckpt", type=str, default=os.path.join(sd_path, "models/ldm/stable-diffusion-v1/model.ckpt"), help="path to checkpoint of model",)
parser.add_argument("--gfpgan-dir", type=str, help="GFPGAN directory", default=('./src/gfpgan' if os.path.exists('./src/gfpgan') else './GFPGAN'))
parser.add_argument("--gfpgan-model", type=str, help="GFPGAN model file name", default='GFPGANv1.3.pth')
parser.add_argument("--no-half", action='store_true', help="do not switch the model to 16-bit floats")
parser.add_argument("--no-progressbar-hiding", action='store_true', help="do not hide progressbar in gradio UI (we hide it because it slows down ML if you have hardware accleration in browser)")
parser.add_argument("--max-batch-count", type=int, default=16, help="maximum batch count value for the UI")
parser.add_argument("--embeddings-dir", type=str, default='embeddings', help="embeddings dirtectory for textual inversion (default: embeddings)")
parser.add_argument("--allow-code", action='store_true', help="allow custom script execution from webui")
parser.add_argument("--medvram", action='store_true', help="enable stable diffusion model optimizations for sacrficing a little speed for low VRM usage")
parser.add_argument("--lowvram", action='store_true', help="enable stable diffusion model optimizations for sacrficing a lot of speed for very low VRM usage")
parser.add_argument("--always-batch-cond-uncond", action='store_true', help="a workaround test; may help with speed in you use --lowvram")
parser.add_argument("--precision", type=str, help="evaluate at this precision", choices=["full", "autocast"], default="autocast")
parser.add_argument("--share", action='store_true', help="use share=True for gradio and make the UI accessible through their site (doesn't work for me but you might have better luck)")
cmd_opts = parser.parse_args()
cpu = torch.device("cpu")
gpu = torch.device("cuda")
device = gpu if torch.cuda.is_available() else cpu
batch_cond_uncond = cmd_opts.always_batch_cond_uncond or not (cmd_opts.lowvram or cmd_opts.medvram)
queue_lock = threading.Lock()
def gr_show(visible=True):
return {"visible": visible, "__type__": "update"}
class State:
interrupted = False
job = ""
def interrupt(self):
self.interrupted = True
state = State()
if not cmd_opts.share:
# fix gradio phoning home
gradio.utils.version_check = lambda: None
gradio.utils.get_local_ip_address = lambda: '127.0.0.1'
css_hide_progressbar = """
.wrap .m-12 svg { display:none!important; }
.wrap .m-12::before { content:"Loading..." }
.progress-bar { display:none!important; }
.meta-text { display:none!important; }
"""
SamplerData = namedtuple('SamplerData', ['name', 'constructor'])
samplers = [
*[SamplerData(x[0], lambda funcname=x[1]: KDiffusionSampler(funcname)) for x in [
('Euler a', 'sample_euler_ancestral'),
('Euler', 'sample_euler'),
('LMS', 'sample_lms'),
('Heun', 'sample_heun'),
('DPM2', 'sample_dpm_2'),
('DPM2 a', 'sample_dpm_2_ancestral'),
] if hasattr(k_diffusion.sampling, x[1])],
SamplerData('DDIM', lambda: VanillaStableDiffusionSampler(DDIMSampler)),
SamplerData('PLMS', lambda: VanillaStableDiffusionSampler(PLMSSampler)),
]
samplers_for_img2img = [x for x in samplers if x.name != 'PLMS']
RealesrganModelInfo = namedtuple("RealesrganModelInfo", ["name", "location", "model", "netscale"])
try:
from basicsr.archs.rrdbnet_arch import RRDBNet
from realesrgan import RealESRGANer
from realesrgan.archs.srvgg_arch import SRVGGNetCompact
realesrgan_models = [
RealesrganModelInfo(
name="Real-ESRGAN 4x plus",
location="https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth",
netscale=4, model=lambda: RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=4)
),
RealesrganModelInfo(
name="Real-ESRGAN 4x plus anime 6B",
location="https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.4/RealESRGAN_x4plus_anime_6B.pth",
netscale=4, model=lambda: RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=6, num_grow_ch=32, scale=4)
),
RealesrganModelInfo(
name="Real-ESRGAN 2x plus",
location="https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.1/RealESRGAN_x2plus.pth",
netscale=2, model=lambda: RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=2)
),
]
have_realesrgan = True
except Exception:
print("Error importing Real-ESRGAN:", file=sys.stderr)
print(traceback.format_exc(), file=sys.stderr)
realesrgan_models = [RealesrganModelInfo('None', '', 0, None)]
have_realesrgan = False
sd_upscalers = {
"RealESRGAN": lambda img: upscale_with_realesrgan(img, 2, 0),
"Lanczos": lambda img: img.resize((img.width*2, img.height*2), resample=LANCZOS),
"None": lambda img: img
}
def gfpgan_model_path():
places = [script_path, '.', os.path.join(cmd_opts.gfpgan_dir, 'experiments/pretrained_models')]
files = [cmd_opts.gfpgan_model] + [os.path.join(dirname, cmd_opts.gfpgan_model) for dirname in places]
found = [x for x in files if os.path.exists(x)]
if len(found) == 0:
raise Exception("GFPGAN model not found in paths: " + ", ".join(files))
return found[0]
def gfpgan():
return GFPGANer(model_path=gfpgan_model_path(), upscale=1, arch='clean', channel_multiplier=2, bg_upsampler=None)
def gfpgan_fix_faces(gfpgan_model, np_image):
np_image_bgr = np_image[:, :, ::-1]
cropped_faces, restored_faces, gfpgan_output_bgr = gfpgan_model.enhance(np_image_bgr, has_aligned=False, only_center_face=False, paste_back=True)
np_image = gfpgan_output_bgr[:, :, ::-1]
return np_image
have_gfpgan = False
try:
model_path = gfpgan_model_path()
if os.path.exists(cmd_opts.gfpgan_dir):
sys.path.append(os.path.abspath(cmd_opts.gfpgan_dir))
from gfpgan import GFPGANer
have_gfpgan = True
except Exception:
print("Error setting up GFPGAN:", file=sys.stderr)
print(traceback.format_exc(), file=sys.stderr)
class Options:
class OptionInfo:
def __init__(self, default=None, label="", component=None, component_args=None):
self.default = default
self.label = label
self.component = component
self.component_args = component_args
data = None
data_labels = {
"outdir_samples": OptionInfo("", "Output dictectory for images; if empty, defaults to two directories below"),
"outdir_txt2img_samples": OptionInfo("outputs/txt2img-images", 'Output dictectory for txt2img images'),
"outdir_img2img_samples": OptionInfo("outputs/img2img-images", 'Output dictectory for img2img images'),
"outdir_extras_samples": OptionInfo("outputs/extras-images", 'Output dictectory for images from extras tab'),
"outdir_grids": OptionInfo("", "Output dictectory for grids; if empty, defaults to two directories below"),
"outdir_txt2img_grids": OptionInfo("outputs/txt2img-grids", 'Output dictectory for txt2img grids'),
"outdir_img2img_grids": OptionInfo("outputs/img2img-grids", 'Output dictectory for img2img grids'),
"save_to_dirs": OptionInfo(False, "When writing images/grids, create a directory with name derived from the prompt"),
"save_to_dirs_prompt_len": OptionInfo(10, "When using above, how many words from prompt to put into directory name", gr.Slider, {"minimum": 1, "maximum": 32, "step": 1}),
"outdir_save": OptionInfo("log/images", "Directory for saving images using the Save button"),
"samples_save": OptionInfo(True, "Save indiviual samples"),
"samples_format": OptionInfo('png', 'File format for indiviual samples'),
"grid_save": OptionInfo(True, "Save image grids"),
"return_grid": OptionInfo(True, "Show grid in results for web"),
"grid_format": OptionInfo('png', 'File format for grids'),
"grid_extended_filename": OptionInfo(False, "Add extended info (seed, prompt) to filename when saving grid"),
"grid_only_if_multiple": OptionInfo(True, "Do not save grids consisting of one picture"),
"n_rows": OptionInfo(-1, "Grid row count; use -1 for autodetect and 0 for it to be same as batch size", gr.Slider, {"minimum": -1, "maximum": 16, "step": 1}),
"jpeg_quality": OptionInfo(80, "Quality for saved jpeg images", gr.Slider, {"minimum": 1, "maximum": 100, "step": 1}),
"export_for_4chan": OptionInfo(True, "If PNG image is larger than 4MB or any dimension is larger than 4000, downscale and save copy as JPG"),
"enable_pnginfo": OptionInfo(True, "Save text information about generation parameters as chunks to png files"),
"font": OptionInfo("arial.ttf", "Font for image grids that have text"),
"prompt_matrix_add_to_start": OptionInfo(True, "In prompt matrix, add the variable combination of text to the start of the prompt, rather than the end"),
}
def __init__(self):
self.data = {k: v.default for k, v in self.data_labels.items()}
def __setattr__(self, key, value):
if self.data is not None:
if key in self.data:
self.data[key] = value
return super(Options, self).__setattr__(key, value)
def __getattr__(self, item):
if self.data is not None:
if item in self.data:
return self.data[item]
if item in self.data_labels:
return self.data_labels[item].default
return super(Options, self).__getattribute__(item)
def save(self, filename):
with open(filename, "w", encoding="utf8") as file:
json.dump(self.data, file)
def load(self, filename):
with open(filename, "r", encoding="utf8") as file:
self.data = json.load(file)
def load_model_from_config(config, ckpt, verbose=False):
print(f"Loading model from {ckpt}")
pl_sd = torch.load(ckpt, map_location="cpu")
if "global_step" in pl_sd:
print(f"Global Step: {pl_sd['global_step']}")
sd = pl_sd["state_dict"]
model = instantiate_from_config(config.model)
m, u = model.load_state_dict(sd, strict=False)
if len(m) > 0 and verbose:
print("missing keys:")
print(m)
if len(u) > 0 and verbose:
print("unexpected keys:")
print(u)
model.eval()
return model
module_in_gpu = None
def setup_for_low_vram(sd_model):
parents = {}
def send_me_to_gpu(module, _):
"""send this module to GPU; send whatever tracked module was previous in GPU to CPU;
we add this as forward_pre_hook to a lot of modules and this way all but one of them will
be in CPU
"""
global module_in_gpu
module = parents.get(module, module)
if module_in_gpu == module:
return
if module_in_gpu is not None:
module_in_gpu.to(cpu)
module.to(gpu)
module_in_gpu = module
# see below for register_forward_pre_hook;
# first_stage_model does not use forward(), it uses encode/decode, so register_forward_pre_hook is
# useless here, and we just replace those methods
def first_stage_model_encode_wrap(self, encoder, x):
send_me_to_gpu(self, None)
return encoder(x)
def first_stage_model_decode_wrap(self, decoder, z):
send_me_to_gpu(self, None)
return decoder(z)
# remove three big modules, cond, first_stage, and unet from the model and then
# send the model to GPU. Then put modules back. the modules will be in CPU.
stored = sd_model.cond_stage_model.transformer, sd_model.first_stage_model, sd_model.model
sd_model.cond_stage_model.transformer, sd_model.first_stage_model, sd_model.model = None, None, None
sd_model.to(device)
sd_model.cond_stage_model.transformer, sd_model.first_stage_model, sd_model.model = stored
# register hooks for those the first two models
sd_model.cond_stage_model.transformer.register_forward_pre_hook(send_me_to_gpu)
sd_model.first_stage_model.register_forward_pre_hook(send_me_to_gpu)
sd_model.first_stage_model.encode = lambda x, en=sd_model.first_stage_model.encode: first_stage_model_encode_wrap(sd_model.first_stage_model, en, x)
sd_model.first_stage_model.decode = lambda z, de=sd_model.first_stage_model.decode: first_stage_model_decode_wrap(sd_model.first_stage_model, de, z)
parents[sd_model.cond_stage_model.transformer] = sd_model.cond_stage_model
if cmd_opts.medvram:
sd_model.model.register_forward_pre_hook(send_me_to_gpu)
else:
diff_model = sd_model.model.diffusion_model
# the third remaining model is still too big for 4GB, so we also do the same for its submodules
# so that only one of them is in GPU at a time
stored = diff_model.input_blocks, diff_model.middle_block, diff_model.output_blocks, diff_model.time_embed
diff_model.input_blocks, diff_model.middle_block, diff_model.output_blocks, diff_model.time_embed = None, None, None, None
sd_model.model.to(device)
diff_model.input_blocks, diff_model.middle_block, diff_model.output_blocks, diff_model.time_embed = stored
# install hooks for bits of third model
diff_model.time_embed.register_forward_pre_hook(send_me_to_gpu)
for block in diff_model.input_blocks:
block.register_forward_pre_hook(send_me_to_gpu)
diff_model.middle_block.register_forward_pre_hook(send_me_to_gpu)
for block in diff_model.output_blocks:
block.register_forward_pre_hook(send_me_to_gpu)
def create_random_tensors(shape, seeds):
xs = []
for seed in seeds:
torch.manual_seed(seed)
# randn results depend on device; gpu and cpu get different results for same seed;
# the way I see it, it's better to do this on CPU, so that everyone gets same result;
# but the original script had it like this so i do not dare change it for now because
# it will break everyone's seeds.
xs.append(torch.randn(shape, device=device))
x = torch.stack(xs)
return x
def torch_gc():
if torch.cuda.is_available():
torch.cuda.empty_cache()
torch.cuda.ipc_collect()
def save_image(image, path, basename, seed=None, prompt=None, extension='png', info=None, short_filename=False, no_prompt=False):
if short_filename or prompt is None or seed is None:
file_decoration = ""
elif opts.save_to_dirs:
file_decoration = f"-{seed}"
else:
file_decoration = f"-{seed}-{sanitize_filename_part(prompt)[:128]}"
if extension == 'png' and opts.enable_pnginfo and info is not None:
pnginfo = PngImagePlugin.PngInfo()
pnginfo.add_text("parameters", info)
else:
pnginfo = None
if opts.save_to_dirs and not no_prompt:
words = re.findall(r'\w+', prompt or "")
if len(words) == 0:
words = ["empty"]
dirname = " ".join(words[0:opts.save_to_dirs_prompt_len])
path = os.path.join(path, dirname)
os.makedirs(path, exist_ok=True)
filecount = len(os.listdir(path))
fullfn = "a.png"
fullfn_without_extension = "a"
for i in range(100):
fn = f"{filecount:05}" if basename == '' else f"{basename}-{filecount:04}"
fullfn = os.path.join(path, f"{fn}{file_decoration}.{extension}")
fullfn_without_extension = os.path.join(path, f"{fn}{file_decoration}")
if not os.path.exists(fullfn):
break
image.save(fullfn, quality=opts.jpeg_quality, pnginfo=pnginfo)
target_side_length = 4000
oversize = image.width > target_side_length or image.height > target_side_length
if opts.export_for_4chan and (oversize or os.stat(fullfn).st_size > 4 * 1024 * 1024):
ratio = image.width / image.height
if oversize and ratio > 1:
image = image.resize((target_side_length, image.height * target_side_length // image.width), LANCZOS)
elif oversize:
image = image.resize((image.width * target_side_length // image.height, target_side_length), LANCZOS)
image.save(f"{fullfn_without_extension}.jpg", quality=opts.jpeg_quality, pnginfo=pnginfo)
def sanitize_filename_part(text):
return text.replace(' ', '_').translate({ord(x): '' for x in invalid_filename_chars})[:128]
def plaintext_to_html(text):
text = "".join([f"<p>{html.escape(x)}</p>\n" for x in text.split('\n')])
return text
def image_grid(imgs, batch_size=1, rows=None):
if rows is None:
if opts.n_rows > 0:
rows = opts.n_rows
elif opts.n_rows == 0:
rows = batch_size
else:
rows = math.sqrt(len(imgs))
rows = round(rows)
cols = math.ceil(len(imgs) / rows)
w, h = imgs[0].size
grid = Image.new('RGB', size=(cols * w, rows * h), color='black')
for i, img in enumerate(imgs):
grid.paste(img, box=(i % cols * w, i // cols * h))
return grid
Grid = namedtuple("Grid", ["tiles", "tile_w", "tile_h", "image_w", "image_h", "overlap"])
def split_grid(image, tile_w=512, tile_h=512, overlap=64):
w = image.width
h = image.height
now = tile_w - overlap # non-overlap width
noh = tile_h - overlap
cols = math.ceil((w - overlap) / now)
rows = math.ceil((h - overlap) / noh)
grid = Grid([], tile_w, tile_h, w, h, overlap)
for row in range(rows):
row_images = []
y = row * noh
if y + tile_h >= h:
y = h - tile_h
for col in range(cols):
x = col * now
if x+tile_w >= w:
x = w - tile_w
tile = image.crop((x, y, x + tile_w, y + tile_h))
row_images.append([x, tile_w, tile])
grid.tiles.append([y, tile_h, row_images])
return grid
def combine_grid(grid):
def make_mask_image(r):
r = r * 255 / grid.overlap
r = r.astype(np.uint8)
return Image.fromarray(r, 'L')
mask_w = make_mask_image(np.arange(grid.overlap, dtype=np.float32).reshape((1, grid.overlap)).repeat(grid.tile_h, axis=0))
mask_h = make_mask_image(np.arange(grid.overlap, dtype=np.float32).reshape((grid.overlap, 1)).repeat(grid.image_w, axis=1))
combined_image = Image.new("RGB", (grid.image_w, grid.image_h))
for y, h, row in grid.tiles:
combined_row = Image.new("RGB", (grid.image_w, h))
for x, w, tile in row:
if x == 0:
combined_row.paste(tile, (0, 0))
continue
combined_row.paste(tile.crop((0, 0, grid.overlap, h)), (x, 0), mask=mask_w)
combined_row.paste(tile.crop((grid.overlap, 0, w, h)), (x + grid.overlap, 0))
if y == 0:
combined_image.paste(combined_row, (0, 0))
continue
combined_image.paste(combined_row.crop((0, 0, combined_row.width, grid.overlap)), (0, y), mask=mask_h)
combined_image.paste(combined_row.crop((0, grid.overlap, combined_row.width, h)), (0, y + grid.overlap))
return combined_image
class GridAnnotation:
def __init__(self, text='', is_active=True):
self.text = text
self.is_active = is_active
self.size = None
def draw_grid_annotations(im, width, height, hor_texts, ver_texts):
def wrap(drawing, text, font, line_length):
lines = ['']
for word in text.split():
line = f'{lines[-1]} {word}'.strip()
if drawing.textlength(line, font=font) <= line_length:
lines[-1] = line
else:
lines.append(word)
return lines
def draw_texts(drawing, draw_x, draw_y, lines):
for i, line in enumerate(lines):
drawing.multiline_text((draw_x, draw_y + line.size[1] / 2), line.text, font=fnt, fill=color_active if line.is_active else color_inactive, anchor="mm", align="center")
if not line.is_active:
drawing.line((draw_x - line.size[0]//2, draw_y + line.size[1]//2, draw_x + line.size[0]//2, draw_y + line.size[1]//2), fill=color_inactive, width=4)
draw_y += line.size[1] + line_spacing
fontsize = (width + height) // 25
line_spacing = fontsize // 2
fnt = ImageFont.truetype(opts.font, fontsize)
color_active = (0, 0, 0)
color_inactive = (153, 153, 153)
pad_left = width * 3 // 4 if len(ver_texts) > 0 else 0
cols = im.width // width
rows = im.height // height
assert cols == len(hor_texts), f'bad number of horizontal texts: {len(hor_texts)}; must be {cols}'
assert rows == len(ver_texts), f'bad number of vertical texts: {len(ver_texts)}; must be {rows}'
calc_img = Image.new("RGB", (1, 1), "white")
calc_d = ImageDraw.Draw(calc_img)
for texts, allowed_width in zip(hor_texts + ver_texts, [width] * len(hor_texts) + [pad_left] * len(ver_texts)):
items = [] + texts
texts.clear()
for line in items:
wrapped = wrap(calc_d, line.text, fnt, allowed_width)
texts += [GridAnnotation(x, line.is_active) for x in wrapped]
for line in texts:
bbox = calc_d.multiline_textbbox((0, 0), line.text, font=fnt)
line.size = (bbox[2] - bbox[0], bbox[3] - bbox[1])
hor_text_heights = [sum([line.size[1] + line_spacing for line in lines]) - line_spacing for lines in hor_texts]
ver_text_heights = [sum([line.size[1] + line_spacing for line in lines]) - line_spacing * len(lines) for lines in ver_texts]
pad_top = max(hor_text_heights) + line_spacing * 2
result = Image.new("RGB", (im.width + pad_left, im.height + pad_top), "white")
result.paste(im, (pad_left, pad_top))
d = ImageDraw.Draw(result)
for col in range(cols):
x = pad_left + width * col + width / 2
y = pad_top / 2 - hor_text_heights[col] / 2
draw_texts(d, x, y, hor_texts[col])
for row in range(rows):
x = pad_left / 2
y = pad_top + height * row + height / 2 - ver_text_heights[row] / 2
draw_texts(d, x, y, ver_texts[row])
return result
def draw_prompt_matrix(im, width, height, all_prompts):
prompts = all_prompts[1:]
boundary = math.ceil(len(prompts) / 2)
prompts_horiz = prompts[:boundary]
prompts_vert = prompts[boundary:]
hor_texts = [[GridAnnotation(x, is_active=pos & (1 << i) != 0) for i, x in enumerate(prompts_horiz)] for pos in range(1 << len(prompts_horiz))]
ver_texts = [[GridAnnotation(x, is_active=pos & (1 << i) != 0) for i, x in enumerate(prompts_vert)] for pos in range(1 << len(prompts_vert))]
return draw_grid_annotations(im, width, height, hor_texts, ver_texts)
def draw_xy_grid(xs, ys, x_label, y_label, cell):
res = []
ver_texts = [[GridAnnotation(y_label(y))] for y in ys]
hor_texts = [[GridAnnotation(x_label(x))] for x in xs]
for y in ys:
for x in xs:
res.append(cell(x, y))
grid = image_grid(res, rows=len(ys))
grid = draw_grid_annotations(grid, res[0].width, res[0].height, hor_texts, ver_texts)
return grid
def resize_image(resize_mode, im, width, height):
if resize_mode == 0:
res = im.resize((width, height), resample=LANCZOS)
elif resize_mode == 1:
ratio = width / height
src_ratio = im.width / im.height
src_w = width if ratio > src_ratio else im.width * height // im.height
src_h = height if ratio <= src_ratio else im.height * width // im.width
resized = im.resize((src_w, src_h), resample=LANCZOS)
res = Image.new("RGB", (width, height))
res.paste(resized, box=(width // 2 - src_w // 2, height // 2 - src_h // 2))
else:
ratio = width / height
src_ratio = im.width / im.height
src_w = width if ratio < src_ratio else im.width * height // im.height
src_h = height if ratio >= src_ratio else im.height * width // im.width
resized = im.resize((src_w, src_h), resample=LANCZOS)
res = Image.new("RGB", (width, height))
res.paste(resized, box=(width // 2 - src_w // 2, height // 2 - src_h // 2))
if ratio < src_ratio:
fill_height = height // 2 - src_h // 2
res.paste(resized.resize((width, fill_height), box=(0, 0, width, 0)), box=(0, 0))
res.paste(resized.resize((width, fill_height), box=(0, resized.height, width, resized.height)), box=(0, fill_height + src_h))
elif ratio > src_ratio:
fill_width = width // 2 - src_w // 2
res.paste(resized.resize((fill_width, height), box=(0, 0, 0, height)), box=(0, 0))
res.paste(resized.resize((fill_width, height), box=(resized.width, 0, resized.width, height)), box=(fill_width + src_w, 0))
return res
def wrap_gradio_gpu_call(func):
def f(*args, **kwargs):
with queue_lock:
res = func(*args, **kwargs)
return res
return wrap_gradio_call(f)
def wrap_gradio_call(func):
def f(*args, **kwargs):
t = time.perf_counter()
try:
res = list(func(*args, **kwargs))
except Exception as e:
print("Error completing request", file=sys.stderr)
print("Arguments:", args, kwargs, file=sys.stderr)
print(traceback.format_exc(), file=sys.stderr)
res = [None, f"<div class='error'>{plaintext_to_html(type(e).__name__+': '+str(e))}</div>"]
elapsed = time.perf_counter() - t
# last item is always HTML
res[-1] = res[-1] + f"<p class='performance'>Time taken: {elapsed:.2f}s</p>"
state.interrupted = False
return tuple(res)
return f
class StableDiffusionModelHijack:
ids_lookup = {}
word_embeddings = {}
word_embeddings_checksums = {}
fixes = None
comments = None
dir_mtime = None
def load_textual_inversion_embeddings(self, dirname, model):
mt = os.path.getmtime(dirname)
if self.dir_mtime is not None and mt <= self.dir_mtime:
return
self.dir_mtime = mt
self.ids_lookup.clear()
self.word_embeddings.clear()
tokenizer = model.cond_stage_model.tokenizer
def const_hash(a):
r = 0
for v in a:
r = (r * 281 ^ int(v) * 997) & 0xFFFFFFFF
return r
def process_file(path, filename):
name = os.path.splitext(filename)[0]
data = torch.load(path)
param_dict = data['string_to_param']
assert len(param_dict) == 1, 'embedding file has multiple terms in it'
emb = next(iter(param_dict.items()))[1].reshape(768)
self.word_embeddings[name] = emb
self.word_embeddings_checksums[name] = f'{const_hash(emb)&0xffff:04x}'
ids = tokenizer([name], add_special_tokens=False)['input_ids'][0]
first_id = ids[0]
if first_id not in self.ids_lookup:
self.ids_lookup[first_id] = []
self.ids_lookup[first_id].append((ids, name))
for fn in os.listdir(dirname):
try:
process_file(os.path.join(dirname, fn), fn)
except Exception:
print(f"Error loading emedding {fn}:", file=sys.stderr)
print(traceback.format_exc(), file=sys.stderr)
continue
print(f"Loaded a total of {len(self.word_embeddings)} text inversion embeddings.")
def hijack(self, m):
model_embeddings = m.cond_stage_model.transformer.text_model.embeddings
model_embeddings.token_embedding = EmbeddingsWithFixes(model_embeddings.token_embedding, self)
m.cond_stage_model = FrozenCLIPEmbedderWithCustomWords(m.cond_stage_model, self)
class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
def __init__(self, wrapped, hijack):
super().__init__()
self.wrapped = wrapped
self.hijack = hijack
self.tokenizer = wrapped.tokenizer
self.max_length = wrapped.max_length
self.token_mults = {}
tokens_with_parens = [(k, v) for k, v in self.tokenizer.get_vocab().items() if '(' in k or ')' in k or '[' in k or ']' in k]
for text, ident in tokens_with_parens:
mult = 1.0
for c in text:
if c == '[':
mult /= 1.1
if c == ']':
mult *= 1.1
if c == '(':
mult *= 1.1
if c == ')':
mult /= 1.1
if mult != 1.0:
self.token_mults[ident] = mult
def forward(self, text):
self.hijack.fixes = []
self.hijack.comments = []
remade_batch_tokens = []
id_start = self.wrapped.tokenizer.bos_token_id
id_end = self.wrapped.tokenizer.eos_token_id
maxlen = self.wrapped.max_length - 2
used_custom_terms = []
cache = {}
batch_tokens = self.wrapped.tokenizer(text, truncation=False, add_special_tokens=False)["input_ids"]
batch_multipliers = []
for tokens in batch_tokens:
tuple_tokens = tuple(tokens)
if tuple_tokens in cache:
remade_tokens, fixes, multipliers = cache[tuple_tokens]
else:
fixes = []
remade_tokens = []
multipliers = []
mult = 1.0
i = 0
while i < len(tokens):
token = tokens[i]
possible_matches = self.hijack.ids_lookup.get(token, None)
mult_change = self.token_mults.get(token)
if mult_change is not None:
mult *= mult_change
elif possible_matches is None:
remade_tokens.append(token)
multipliers.append(mult)
else:
found = False
for ids, word in possible_matches:
if tokens[i:i+len(ids)] == ids:
fixes.append((len(remade_tokens), word))
remade_tokens.append(777)
multipliers.append(mult)
i += len(ids) - 1
found = True
used_custom_terms.append((word, self.hijack.word_embeddings_checksums[word]))
break
if not found:
remade_tokens.append(token)
multipliers.append(mult)
i += 1
if len(remade_tokens) > maxlen - 2:
vocab = {v: k for k, v in self.wrapped.tokenizer.get_vocab().items()}
ovf = remade_tokens[maxlen - 2:]
overflowing_words = [vocab.get(int(x), "") for x in ovf]
overflowing_text = self.wrapped.tokenizer.convert_tokens_to_string(''.join(overflowing_words))
self.hijack.comments.append(f"Warning: too many input tokens; some ({len(overflowing_words)}) have been truncated:\n{overflowing_text}\n")
remade_tokens = remade_tokens + [id_end] * (maxlen - 2 - len(remade_tokens))
remade_tokens = [id_start] + remade_tokens[0:maxlen-2] + [id_end]
cache[tuple_tokens] = (remade_tokens, fixes, multipliers)
multipliers = multipliers + [1.0] * (maxlen - 2 - len(multipliers))
multipliers = [1.0] + multipliers[0:maxlen - 2] + [1.0]
remade_batch_tokens.append(remade_tokens)
self.hijack.fixes.append(fixes)
batch_multipliers.append(multipliers)
if len(used_custom_terms) > 0:
self.hijack.comments.append("Used custom terms: " + ", ".join([f'{word} [{checksum}]' for word, checksum in used_custom_terms]))
tokens = torch.asarray(remade_batch_tokens).to(device)
outputs = self.wrapped.transformer(input_ids=tokens)
z = outputs.last_hidden_state
# restoring original mean is likely not correct, but it seems to work well to prevent artifacts that happen otherwise
batch_multipliers = torch.asarray(np.array(batch_multipliers)).to(device)
original_mean = z.mean()
z *= batch_multipliers.reshape(batch_multipliers.shape + (1,)).expand(z.shape)
new_mean = z.mean()
z *= original_mean / new_mean
return z
class EmbeddingsWithFixes(nn.Module):
def __init__(self, wrapped, embeddings):
super().__init__()
self.wrapped = wrapped
self.embeddings = embeddings
def forward(self, input_ids):
batch_fixes = self.embeddings.fixes
self.embeddings.fixes = None
inputs_embeds = self.wrapped(input_ids)
if batch_fixes is not None:
for fixes, tensor in zip(batch_fixes, inputs_embeds):
for offset, word in fixes:
tensor[offset] = self.embeddings.word_embeddings[word]
return inputs_embeds
class StableDiffusionProcessing:
def __init__(self, outpath_samples=None, outpath_grids=None, prompt="", seed=-1, sampler_index=0, batch_size=1, n_iter=1, steps=50, cfg_scale=7.0, width=512, height=512, prompt_matrix=False, use_GFPGAN=False, do_not_save_samples=False, do_not_save_grid=False, extra_generation_params=None, overlay_images=None, negative_prompt=None):
self.outpath_samples: str = outpath_samples
self.outpath_grids: str = outpath_grids
self.prompt: str = prompt
self.negative_prompt: str = (negative_prompt or "")
self.seed: int = seed
self.sampler_index: int = sampler_index
self.batch_size: int = batch_size
self.n_iter: int = n_iter
self.steps: int = steps
self.cfg_scale: float = cfg_scale
self.width: int = width
self.height: int = height
self.prompt_matrix: bool = prompt_matrix
self.use_GFPGAN: bool = use_GFPGAN
self.do_not_save_samples: bool = do_not_save_samples
self.do_not_save_grid: bool = do_not_save_grid
self.extra_generation_params: dict = extra_generation_params
self.overlay_images = overlay_images
self.paste_to = None
def init(self):
pass
def sample(self, x, conditioning, unconditional_conditioning):
raise NotImplementedError()
def p_sample_ddim_hook(sampler_wrapper, x_dec, cond, ts, *args, **kwargs):
if sampler_wrapper.mask is not None:
img_orig = sampler_wrapper.sampler.model.q_sample(sampler_wrapper.init_latent, ts)
x_dec = img_orig * sampler_wrapper.mask + sampler_wrapper.nmask * x_dec
return sampler_wrapper.orig_p_sample_ddim(x_dec, cond, ts, *args, **kwargs)
class VanillaStableDiffusionSampler:
def __init__(self, constructor):
self.sampler = constructor(sd_model)
self.orig_p_sample_ddim = self.sampler.p_sample_ddim if hasattr(self.sampler, 'p_sample_ddim') else None
self.mask = None
self.nmask = None
self.init_latent = None
def sample_img2img(self, p, x, noise, conditioning, unconditional_conditioning):
t_enc = int(min(p.denoising_strength, 0.999) * p.steps)
# existing code fails with cetin step counts, like 9
try:
self.sampler.make_schedule(ddim_num_steps=p.steps, verbose=False)
except Exception:
self.sampler.make_schedule(ddim_num_steps=p.steps+1, verbose=False)
x1 = self.sampler.stochastic_encode(x, torch.tensor([t_enc] * int(x.shape[0])).to(device), noise=noise)
self.sampler.p_sample_ddim = lambda x_dec, cond, ts, *args, **kwargs: p_sample_ddim_hook(self, x_dec, cond, ts, *args, **kwargs)
self.mask = p.mask
self.nmask = p.nmask
self.init_latent = p.init_latent
samples = self.sampler.decode(x1, conditioning, t_enc, unconditional_guidance_scale=p.cfg_scale, unconditional_conditioning=unconditional_conditioning)
return samples
def sample(self, p: StableDiffusionProcessing, x, conditioning, unconditional_conditioning):
samples_ddim, _ = self.sampler.sample(S=p.steps, conditioning=conditioning, batch_size=int(x.shape[0]), shape=x[0].shape, verbose=False, unconditional_guidance_scale=p.cfg_scale, unconditional_conditioning=unconditional_conditioning, x_T=x)
return samples_ddim
class CFGDenoiser(nn.Module):
def __init__(self, model):
super().__init__()
self.inner_model = model
self.mask = None
self.nmask = None
self.init_latent = None
def forward(self, x, sigma, uncond, cond, cond_scale):
if batch_cond_uncond:
x_in = torch.cat([x] * 2)
sigma_in = torch.cat([sigma] * 2)
cond_in = torch.cat([uncond, cond])
uncond, cond = self.inner_model(x_in, sigma_in, cond=cond_in).chunk(2)
denoised = uncond + (cond - uncond) * cond_scale
else:
uncond = self.inner_model(x, sigma, cond=uncond)
cond = self.inner_model(x, sigma, cond=cond)
denoised = uncond + (cond - uncond) * cond_scale
if self.mask is not None:
denoised = self.init_latent * self.mask + self.nmask * denoised
return denoised
def extended_trange(*args, **kwargs):
for x in tqdm.trange(*args, desc=state.job, **kwargs):
if state.interrupted:
break
yield x
class KDiffusionSampler:
def __init__(self, funcname):
self.model_wrap = k_diffusion.external.CompVisDenoiser(sd_model)