forked from huggingface/diffusers
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathsingle_file_utils.py
More file actions
1387 lines (1155 loc) · 55.7 KB
/
single_file_utils.py
File metadata and controls
1387 lines (1155 loc) · 55.7 KB
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
# coding=utf-8
# Copyright 2023 The HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" Conversion script for the Stable Diffusion checkpoints."""
import os
import re
from contextlib import nullcontext
from io import BytesIO
from urllib.parse import urlparse
import requests
import yaml
from ..models.modeling_utils import load_state_dict
from ..schedulers import (
DDIMScheduler,
DDPMScheduler,
DPMSolverMultistepScheduler,
EulerAncestralDiscreteScheduler,
EulerDiscreteScheduler,
HeunDiscreteScheduler,
LMSDiscreteScheduler,
PNDMScheduler,
)
from ..utils import is_accelerate_available, is_transformers_available, logging
from ..utils.hub_utils import _get_model_file
if is_transformers_available():
from transformers import (
CLIPTextConfig,
CLIPTextModel,
CLIPTextModelWithProjection,
CLIPTokenizer,
)
if is_accelerate_available():
from accelerate import init_empty_weights
from accelerate.utils import set_module_tensor_to_device
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
CONFIG_URLS = {
"v1": "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/configs/stable-diffusion/v1-inference.yaml",
"v2": "https://raw.githubusercontent.com/Stability-AI/stablediffusion/main/configs/stable-diffusion/v2-inference-v.yaml",
"xl": "https://raw.githubusercontent.com/Stability-AI/generative-models/main/configs/inference/sd_xl_base.yaml",
"xl_refiner": "https://raw.githubusercontent.com/Stability-AI/generative-models/main/configs/inference/sd_xl_refiner.yaml",
"upscale": "https://raw.githubusercontent.com/Stability-AI/stablediffusion/main/configs/stable-diffusion/x4-upscaling.yaml",
"controlnet": "https://raw.githubusercontent.com/lllyasviel/ControlNet/main/models/cldm_v15.yaml",
}
CHECKPOINT_KEY_NAMES = {
"v2": "model.diffusion_model.input_blocks.2.1.transformer_blocks.0.attn2.to_k.weight",
"xl_base": "conditioner.embedders.1.model.transformer.resblocks.9.mlp.c_proj.bias",
"xl_refiner": "conditioner.embedders.0.model.transformer.resblocks.9.mlp.c_proj.bias",
}
SCHEDULER_DEFAULT_CONFIG = {
"beta_schedule": "scaled_linear",
"beta_start": 0.00085,
"beta_end": 0.012,
"interpolation_type": "linear",
"num_train_timesteps": 1000,
"prediction_type": "epsilon",
"sample_max_value": 1.0,
"set_alpha_to_one": False,
"skip_prk_steps": True,
"steps_offset": 1,
"timestep_spacing": "leading",
}
DIFFUSERS_TO_LDM_MAPPING = {
"unet": {
"layers": {
"time_embedding.linear_1.weight": "time_embed.0.weight",
"time_embedding.linear_1.bias": "time_embed.0.bias",
"time_embedding.linear_2.weight": "time_embed.2.weight",
"time_embedding.linear_2.bias": "time_embed.2.bias",
"conv_in.weight": "input_blocks.0.0.weight",
"conv_in.bias": "input_blocks.0.0.bias",
"conv_norm_out.weight": "out.0.weight",
"conv_norm_out.bias": "out.0.bias",
"conv_out.weight": "out.2.weight",
"conv_out.bias": "out.2.bias",
},
"class_embed_type": {
"class_embedding.linear_1.weight": "label_emb.0.0.weight",
"class_embedding.linear_1.bias": "label_emb.0.0.bias",
"class_embedding.linear_2.weight": "label_emb.0.2.weight",
"class_embedding.linear_2.bias": "label_emb.0.2.bias",
},
"addition_embed_type": {
"add_embedding.linear_1.weight": "label_emb.0.0.weight",
"add_embedding.linear_1.bias": "label_emb.0.0.bias",
"add_embedding.linear_2.weight": "label_emb.0.2.weight",
"add_embedding.linear_2.bias": "label_emb.0.2.bias",
},
},
"controlnet": {
"layers": {
"time_embedding.linear_1.weight": "time_embed.0.weight",
"time_embedding.linear_1.bias": "time_embed.0.bias",
"time_embedding.linear_2.weight": "time_embed.2.weight",
"time_embedding.linear_2.bias": "time_embed.2.bias",
"conv_in.weight": "input_blocks.0.0.weight",
"conv_in.bias": "input_blocks.0.0.bias",
"controlnet_cond_embedding.conv_in.weight": "input_hint_block.0.weight",
"controlnet_cond_embedding.conv_in.bias": "input_hint_block.0.bias",
"controlnet_cond_embedding.conv_out.weight": "input_hint_block.14.weight",
"controlnet_cond_embedding.conv_out.bias": "input_hint_block.14.bias",
},
"class_embed_type": {
"class_embedding.linear_1.weight": "label_emb.0.0.weight",
"class_embedding.linear_1.bias": "label_emb.0.0.bias",
"class_embedding.linear_2.weight": "label_emb.0.2.weight",
"class_embedding.linear_2.bias": "label_emb.0.2.bias",
},
"addition_embed_type": {
"add_embedding.linear_1.weight": "label_emb.0.0.weight",
"add_embedding.linear_1.bias": "label_emb.0.0.bias",
"add_embedding.linear_2.weight": "label_emb.0.2.weight",
"add_embedding.linear_2.bias": "label_emb.0.2.bias",
},
},
"vae": {
"encoder.conv_in.weight": "encoder.conv_in.weight",
"encoder.conv_in.bias": "encoder.conv_in.bias",
"encoder.conv_out.weight": "encoder.conv_out.weight",
"encoder.conv_out.bias": "encoder.conv_out.bias",
"encoder.conv_norm_out.weight": "encoder.norm_out.weight",
"encoder.conv_norm_out.bias": "encoder.norm_out.bias",
"decoder.conv_in.weight": "decoder.conv_in.weight",
"decoder.conv_in.bias": "decoder.conv_in.bias",
"decoder.conv_out.weight": "decoder.conv_out.weight",
"decoder.conv_out.bias": "decoder.conv_out.bias",
"decoder.conv_norm_out.weight": "decoder.norm_out.weight",
"decoder.conv_norm_out.bias": "decoder.norm_out.bias",
"quant_conv.weight": "quant_conv.weight",
"quant_conv.bias": "quant_conv.bias",
"post_quant_conv.weight": "post_quant_conv.weight",
"post_quant_conv.bias": "post_quant_conv.bias",
},
"openclip": {
"layers": {
"text_model.embeddings.position_embedding.weight": "positional_embedding",
"text_model.embeddings.token_embedding.weight": "token_embedding.weight",
"text_model.final_layer_norm.weight": "ln_final.weight",
"text_model.final_layer_norm.bias": "ln_final.bias",
"text_projection.weight": "text_projection",
},
"transformer": {
"text_model.encoder.layers.": "resblocks.",
"layer_norm1": "ln_1",
"layer_norm2": "ln_2",
".fc1.": ".c_fc.",
".fc2.": ".c_proj.",
".self_attn": ".attn",
"transformer.text_model.final_layer_norm.": "ln_final.",
"transformer.text_model.embeddings.token_embedding.weight": "token_embedding.weight",
"transformer.text_model.embeddings.position_embedding.weight": "positional_embedding",
},
},
}
LDM_VAE_KEY = "first_stage_model."
LDM_UNET_KEY = "model.diffusion_model."
LDM_CONTROLNET_KEY = "control_model."
LDM_CLIP_PREFIX_TO_REMOVE = ["cond_stage_model.transformer.", "conditioner.embedders.0.transformer."]
LDM_OPEN_CLIP_TEXT_PROJECTION_DIM = 1024
SD_2_TEXT_ENCODER_KEYS_TO_IGNORE = [
"cond_stage_model.model.transformer.resblocks.23.attn.in_proj_bias",
"cond_stage_model.model.transformer.resblocks.23.attn.in_proj_weight",
"cond_stage_model.model.transformer.resblocks.23.attn.out_proj.bias",
"cond_stage_model.model.transformer.resblocks.23.attn.out_proj.weight",
"cond_stage_model.model.transformer.resblocks.23.ln_1.bias",
"cond_stage_model.model.transformer.resblocks.23.ln_1.weight",
"cond_stage_model.model.transformer.resblocks.23.ln_2.bias",
"cond_stage_model.model.transformer.resblocks.23.ln_2.weight",
"cond_stage_model.model.transformer.resblocks.23.mlp.c_fc.bias",
"cond_stage_model.model.transformer.resblocks.23.mlp.c_fc.weight",
"cond_stage_model.model.transformer.resblocks.23.mlp.c_proj.bias",
"cond_stage_model.model.transformer.resblocks.23.mlp.c_proj.weight",
"cond_stage_model.model.text_projection",
]
VALID_URL_PREFIXES = ["https://huggingface.co/", "huggingface.co/", "hf.co/", "https://hf.co/"]
def _extract_repo_id_and_weights_name(pretrained_model_name_or_path):
pattern = r"([^/]+)/([^/]+)/(?:blob/main/)?(.+)"
weights_name = None
repo_id = (None,)
for prefix in VALID_URL_PREFIXES:
pretrained_model_name_or_path = pretrained_model_name_or_path.replace(prefix, "")
match = re.match(pattern, pretrained_model_name_or_path)
if not match:
return repo_id, weights_name
repo_id = f"{match.group(1)}/{match.group(2)}"
weights_name = match.group(3)
return repo_id, weights_name
def fetch_ldm_config_and_checkpoint(
pretrained_model_link_or_path,
class_name,
original_config_file=None,
resume_download=False,
force_download=False,
proxies=None,
token=None,
cache_dir=None,
local_files_only=None,
revision=None,
use_safetensors=True,
):
file_extension = pretrained_model_link_or_path.rsplit(".", 1)[-1]
from_safetensors = file_extension == "safetensors"
if from_safetensors and use_safetensors is False:
raise ValueError("Make sure to install `safetensors` with `pip install safetensors`.")
if os.path.isfile(pretrained_model_link_or_path):
checkpoint = load_state_dict(pretrained_model_link_or_path)
else:
repo_id, weights_name = _extract_repo_id_and_weights_name(pretrained_model_link_or_path)
checkpoint_path = _get_model_file(
repo_id,
weights_name=weights_name,
force_download=force_download,
cache_dir=cache_dir,
resume_download=resume_download,
proxies=proxies,
local_files_only=local_files_only,
token=token,
revision=revision,
)
checkpoint = load_state_dict(checkpoint_path)
# some checkpoints contain the model state dict under a "state_dict" key
while "state_dict" in checkpoint:
checkpoint = checkpoint["state_dict"]
original_config = fetch_original_config(class_name, checkpoint, original_config_file)
return original_config, checkpoint
def infer_original_config_file(class_name, checkpoint):
if CHECKPOINT_KEY_NAMES["v2"] in checkpoint and checkpoint[CHECKPOINT_KEY_NAMES["v2"]].shape[-1] == 1024:
config_url = CONFIG_URLS["v2"]
elif CHECKPOINT_KEY_NAMES["xl_base"] in checkpoint:
config_url = CONFIG_URLS["xl"]
elif CHECKPOINT_KEY_NAMES["xl_refiner"] in checkpoint:
config_url = CONFIG_URLS["xl_refiner"]
elif class_name == "StableDiffusionUpscalePipeline":
config_url = CONFIG_URLS["upscale"]
elif class_name == "ControlNetModel":
config_url = CONFIG_URLS["controlnet"]
else:
config_url = CONFIG_URLS["v1"]
original_config_file = BytesIO(requests.get(config_url).content)
return original_config_file
def fetch_original_config(pipeline_class_name, checkpoint, original_config_file=None):
def is_valid_url(url):
result = urlparse(url)
if result.scheme and result.netloc:
return True
return False
if original_config_file is None:
original_config_file = infer_original_config_file(pipeline_class_name, checkpoint)
elif os.path.isfile(original_config_file):
with open(original_config_file, "r") as fp:
original_config_file = fp.read()
elif is_valid_url(original_config_file):
original_config_file = BytesIO(requests.get(original_config_file).content)
else:
raise ValueError("Invalid `original_config_file` provided. Please set it to a valid file path or URL.")
original_config = yaml.safe_load(original_config_file)
return original_config
def infer_model_type(original_config, model_type=None):
if model_type is not None:
return model_type
has_cond_stage_config = (
"cond_stage_config" in original_config["model"]["params"]
and original_config["model"]["params"]["cond_stage_config"] is not None
)
has_network_config = (
"network_config" in original_config["model"]["params"]
and original_config["model"]["params"]["network_config"] is not None
)
if has_cond_stage_config:
model_type = original_config["model"]["params"]["cond_stage_config"]["target"].split(".")[-1]
elif has_network_config:
context_dim = original_config["model"]["params"]["network_config"]["params"]["context_dim"]
if context_dim == 2048:
model_type = "SDXL"
else:
model_type = "SDXL-Refiner"
else:
raise ValueError("Unable to infer model type from config")
logger.debug(f"No `model_type` given, `model_type` inferred as: {model_type}")
return model_type
def get_default_scheduler_config():
return SCHEDULER_DEFAULT_CONFIG
def set_image_size(pipeline_class_name, original_config, checkpoint, image_size=None, model_type=None):
if image_size:
return image_size
global_step = checkpoint["global_step"] if "global_step" in checkpoint else None
model_type = infer_model_type(original_config, model_type)
if pipeline_class_name == "StableDiffusionUpscalePipeline":
image_size = original_config["model"]["params"]["unet_config"]["params"]["image_size"]
return image_size
elif model_type in ["SDXL", "SDXL-Refiner"]:
image_size = 1024
return image_size
elif (
"parameterization" in original_config["model"]["params"]
and original_config["model"]["params"]["parameterization"] == "v"
):
# NOTE: For stable diffusion 2 base one has to pass `image_size==512`
# as it relies on a brittle global step parameter here
image_size = 512 if global_step == 875000 else 768
return image_size
else:
image_size = 512
return image_size
# Copied from diffusers.pipelines.stable_diffusion.convert_from_ckpt.conv_attn_to_linear
def conv_attn_to_linear(checkpoint):
keys = list(checkpoint.keys())
attn_keys = ["query.weight", "key.weight", "value.weight"]
for key in keys:
if ".".join(key.split(".")[-2:]) in attn_keys:
if checkpoint[key].ndim > 2:
checkpoint[key] = checkpoint[key][:, :, 0, 0]
elif "proj_attn.weight" in key:
if checkpoint[key].ndim > 2:
checkpoint[key] = checkpoint[key][:, :, 0]
def create_unet_diffusers_config(original_config, image_size: int):
"""
Creates a config for the diffusers based on the config of the LDM model.
"""
if (
"unet_config" in original_config["model"]["params"]
and original_config["model"]["params"]["unet_config"] is not None
):
unet_params = original_config["model"]["params"]["unet_config"]["params"]
else:
unet_params = original_config["model"]["params"]["network_config"]["params"]
vae_params = original_config["model"]["params"]["first_stage_config"]["params"]["ddconfig"]
block_out_channels = [unet_params["model_channels"] * mult for mult in unet_params["channel_mult"]]
down_block_types = []
resolution = 1
for i in range(len(block_out_channels)):
block_type = "CrossAttnDownBlock2D" if resolution in unet_params["attention_resolutions"] else "DownBlock2D"
down_block_types.append(block_type)
if i != len(block_out_channels) - 1:
resolution *= 2
up_block_types = []
for i in range(len(block_out_channels)):
block_type = "CrossAttnUpBlock2D" if resolution in unet_params["attention_resolutions"] else "UpBlock2D"
up_block_types.append(block_type)
resolution //= 2
if unet_params["transformer_depth"] is not None:
transformer_layers_per_block = (
unet_params["transformer_depth"]
if isinstance(unet_params["transformer_depth"], int)
else list(unet_params["transformer_depth"])
)
else:
transformer_layers_per_block = 1
vae_scale_factor = 2 ** (len(vae_params["ch_mult"]) - 1)
head_dim = unet_params["num_heads"] if "num_heads" in unet_params else None
use_linear_projection = (
unet_params["use_linear_in_transformer"] if "use_linear_in_transformer" in unet_params else False
)
if use_linear_projection:
# stable diffusion 2-base-512 and 2-768
if head_dim is None:
head_dim_mult = unet_params["model_channels"] // unet_params["num_head_channels"]
head_dim = [head_dim_mult * c for c in list(unet_params["channel_mult"])]
class_embed_type = None
addition_embed_type = None
addition_time_embed_dim = None
projection_class_embeddings_input_dim = None
context_dim = None
if unet_params["context_dim"] is not None:
context_dim = (
unet_params["context_dim"]
if isinstance(unet_params["context_dim"], int)
else unet_params["context_dim"][0]
)
if "num_classes" in unet_params:
if unet_params["num_classes"] == "sequential":
if context_dim in [2048, 1280]:
# SDXL
addition_embed_type = "text_time"
addition_time_embed_dim = 256
else:
class_embed_type = "projection"
assert "adm_in_channels" in unet_params
projection_class_embeddings_input_dim = unet_params["adm_in_channels"]
config = {
"sample_size": image_size // vae_scale_factor,
"in_channels": unet_params["in_channels"],
"down_block_types": tuple(down_block_types),
"block_out_channels": tuple(block_out_channels),
"layers_per_block": unet_params["num_res_blocks"],
"cross_attention_dim": context_dim,
"attention_head_dim": head_dim,
"use_linear_projection": use_linear_projection,
"class_embed_type": class_embed_type,
"addition_embed_type": addition_embed_type,
"addition_time_embed_dim": addition_time_embed_dim,
"projection_class_embeddings_input_dim": projection_class_embeddings_input_dim,
"transformer_layers_per_block": transformer_layers_per_block,
}
if "disable_self_attentions" in unet_params:
config["only_cross_attention"] = unet_params["disable_self_attentions"]
if "num_classes" in unet_params and isinstance(unet_params["num_classes"], int):
config["num_class_embeds"] = unet_params["num_classes"]
config["out_channels"] = unet_params["out_channels"]
config["up_block_types"] = tuple(up_block_types)
return config
def create_controlnet_diffusers_config(original_config, image_size: int):
unet_params = original_config["model"]["params"]["control_stage_config"]["params"]
diffusers_unet_config = create_unet_diffusers_config(original_config, image_size=image_size)
controlnet_config = {
"conditioning_channels": unet_params["hint_channels"],
"in_channels": diffusers_unet_config["in_channels"],
"down_block_types": diffusers_unet_config["down_block_types"],
"block_out_channels": diffusers_unet_config["block_out_channels"],
"layers_per_block": diffusers_unet_config["layers_per_block"],
"cross_attention_dim": diffusers_unet_config["cross_attention_dim"],
"attention_head_dim": diffusers_unet_config["attention_head_dim"],
"use_linear_projection": diffusers_unet_config["use_linear_projection"],
"class_embed_type": diffusers_unet_config["class_embed_type"],
"addition_embed_type": diffusers_unet_config["addition_embed_type"],
"addition_time_embed_dim": diffusers_unet_config["addition_time_embed_dim"],
"projection_class_embeddings_input_dim": diffusers_unet_config["projection_class_embeddings_input_dim"],
"transformer_layers_per_block": diffusers_unet_config["transformer_layers_per_block"],
}
return controlnet_config
def create_vae_diffusers_config(original_config, image_size, scaling_factor=0.18125):
"""
Creates a config for the diffusers based on the config of the LDM model.
"""
vae_params = original_config["model"]["params"]["first_stage_config"]["params"]["ddconfig"]
block_out_channels = [vae_params["ch"] * mult for mult in vae_params["ch_mult"]]
down_block_types = ["DownEncoderBlock2D"] * len(block_out_channels)
up_block_types = ["UpDecoderBlock2D"] * len(block_out_channels)
config = {
"sample_size": image_size,
"in_channels": vae_params["in_channels"],
"out_channels": vae_params["out_ch"],
"down_block_types": tuple(down_block_types),
"up_block_types": tuple(up_block_types),
"block_out_channels": tuple(block_out_channels),
"latent_channels": vae_params["z_channels"],
"layers_per_block": vae_params["num_res_blocks"],
"scaling_factor": scaling_factor,
}
return config
def update_unet_resnet_ldm_to_diffusers(ldm_keys, new_checkpoint, checkpoint, mapping=None):
for ldm_key in ldm_keys:
diffusers_key = (
ldm_key.replace("in_layers.0", "norm1")
.replace("in_layers.2", "conv1")
.replace("out_layers.0", "norm2")
.replace("out_layers.3", "conv2")
.replace("emb_layers.1", "time_emb_proj")
.replace("skip_connection", "conv_shortcut")
)
if mapping:
diffusers_key = diffusers_key.replace(mapping["old"], mapping["new"])
new_checkpoint[diffusers_key] = checkpoint.pop(ldm_key)
def update_unet_attention_ldm_to_diffusers(ldm_keys, new_checkpoint, checkpoint, mapping):
for ldm_key in ldm_keys:
diffusers_key = ldm_key.replace(mapping["old"], mapping["new"])
new_checkpoint[diffusers_key] = checkpoint.pop(ldm_key)
def convert_ldm_unet_checkpoint(checkpoint, config, extract_ema=False):
"""
Takes a state dict and a config, and returns a converted checkpoint.
"""
# extract state_dict for UNet
unet_state_dict = {}
keys = list(checkpoint.keys())
unet_key = LDM_UNET_KEY
# at least a 100 parameters have to start with `model_ema` in order for the checkpoint to be EMA
if sum(k.startswith("model_ema") for k in keys) > 100 and extract_ema:
logger.warning("Checkpoint has both EMA and non-EMA weights.")
logger.warning(
"In this conversion only the EMA weights are extracted. If you want to instead extract the non-EMA"
" weights (useful to continue fine-tuning), please make sure to remove the `--extract_ema` flag."
)
for key in keys:
if key.startswith("model.diffusion_model"):
flat_ema_key = "model_ema." + "".join(key.split(".")[1:])
unet_state_dict[key.replace(unet_key, "")] = checkpoint.pop(flat_ema_key)
else:
if sum(k.startswith("model_ema") for k in keys) > 100:
logger.warning(
"In this conversion only the non-EMA weights are extracted. If you want to instead extract the EMA"
" weights (usually better for inference), please make sure to add the `--extract_ema` flag."
)
for key in keys:
if key.startswith(unet_key):
unet_state_dict[key.replace(unet_key, "")] = checkpoint.pop(key)
new_checkpoint = {}
ldm_unet_keys = DIFFUSERS_TO_LDM_MAPPING["unet"]["layers"]
for diffusers_key, ldm_key in ldm_unet_keys.items():
if ldm_key not in unet_state_dict:
continue
new_checkpoint[diffusers_key] = unet_state_dict[ldm_key]
if ("class_embed_type" in config) and (config["class_embed_type"] in ["timestep", "projection"]):
class_embed_keys = DIFFUSERS_TO_LDM_MAPPING["unet"]["class_embed_type"]
for diffusers_key, ldm_key in class_embed_keys.items():
new_checkpoint[diffusers_key] = unet_state_dict[ldm_key]
if ("addition_embed_type" in config) and (config["addition_embed_type"] == "text_time"):
addition_embed_keys = DIFFUSERS_TO_LDM_MAPPING["unet"]["addition_embed_type"]
for diffusers_key, ldm_key in addition_embed_keys.items():
new_checkpoint[diffusers_key] = unet_state_dict[ldm_key]
# Relevant to StableDiffusionUpscalePipeline
if "num_class_embeds" in config:
if (config["num_class_embeds"] is not None) and ("label_emb.weight" in unet_state_dict):
new_checkpoint["class_embedding.weight"] = unet_state_dict["label_emb.weight"]
# Retrieves the keys for the input blocks only
num_input_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "input_blocks" in layer})
input_blocks = {
layer_id: [key for key in unet_state_dict if f"input_blocks.{layer_id}" in key]
for layer_id in range(num_input_blocks)
}
# Retrieves the keys for the middle blocks only
num_middle_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "middle_block" in layer})
middle_blocks = {
layer_id: [key for key in unet_state_dict if f"middle_block.{layer_id}" in key]
for layer_id in range(num_middle_blocks)
}
# Retrieves the keys for the output blocks only
num_output_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "output_blocks" in layer})
output_blocks = {
layer_id: [key for key in unet_state_dict if f"output_blocks.{layer_id}" in key]
for layer_id in range(num_output_blocks)
}
# Down blocks
for i in range(1, num_input_blocks):
block_id = (i - 1) // (config["layers_per_block"] + 1)
layer_in_block_id = (i - 1) % (config["layers_per_block"] + 1)
resnets = [
key for key in input_blocks[i] if f"input_blocks.{i}.0" in key and f"input_blocks.{i}.0.op" not in key
]
update_unet_resnet_ldm_to_diffusers(
resnets,
new_checkpoint,
unet_state_dict,
{"old": f"input_blocks.{i}.0", "new": f"down_blocks.{block_id}.resnets.{layer_in_block_id}"},
)
if f"input_blocks.{i}.0.op.weight" in unet_state_dict:
new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.weight"] = unet_state_dict.pop(
f"input_blocks.{i}.0.op.weight"
)
new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.bias"] = unet_state_dict.pop(
f"input_blocks.{i}.0.op.bias"
)
attentions = [key for key in input_blocks[i] if f"input_blocks.{i}.1" in key]
if attentions:
update_unet_attention_ldm_to_diffusers(
attentions,
new_checkpoint,
unet_state_dict,
{"old": f"input_blocks.{i}.1", "new": f"down_blocks.{block_id}.attentions.{layer_in_block_id}"},
)
# Mid blocks
resnet_0 = middle_blocks[0]
attentions = middle_blocks[1]
resnet_1 = middle_blocks[2]
update_unet_resnet_ldm_to_diffusers(
resnet_0, new_checkpoint, unet_state_dict, mapping={"old": "middle_block.0", "new": "mid_block.resnets.0"}
)
update_unet_resnet_ldm_to_diffusers(
resnet_1, new_checkpoint, unet_state_dict, mapping={"old": "middle_block.2", "new": "mid_block.resnets.1"}
)
update_unet_attention_ldm_to_diffusers(
attentions, new_checkpoint, unet_state_dict, mapping={"old": "middle_block.1", "new": "mid_block.attentions.0"}
)
# Up Blocks
for i in range(num_output_blocks):
block_id = i // (config["layers_per_block"] + 1)
layer_in_block_id = i % (config["layers_per_block"] + 1)
resnets = [
key for key in output_blocks[i] if f"output_blocks.{i}.0" in key and f"output_blocks.{i}.0.op" not in key
]
update_unet_resnet_ldm_to_diffusers(
resnets,
new_checkpoint,
unet_state_dict,
{"old": f"output_blocks.{i}.0", "new": f"up_blocks.{block_id}.resnets.{layer_in_block_id}"},
)
attentions = [
key for key in output_blocks[i] if f"output_blocks.{i}.1" in key and f"output_blocks.{i}.1.conv" not in key
]
if attentions:
update_unet_attention_ldm_to_diffusers(
attentions,
new_checkpoint,
unet_state_dict,
{"old": f"output_blocks.{i}.1", "new": f"up_blocks.{block_id}.attentions.{layer_in_block_id}"},
)
if f"output_blocks.{i}.1.conv.weight" in unet_state_dict:
new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.weight"] = unet_state_dict[
f"output_blocks.{i}.1.conv.weight"
]
new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.bias"] = unet_state_dict[
f"output_blocks.{i}.1.conv.bias"
]
if f"output_blocks.{i}.2.conv.weight" in unet_state_dict:
new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.weight"] = unet_state_dict[
f"output_blocks.{i}.2.conv.weight"
]
new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.bias"] = unet_state_dict[
f"output_blocks.{i}.2.conv.bias"
]
return new_checkpoint
def convert_controlnet_checkpoint(
checkpoint,
config,
):
# Some controlnet ckpt files are distributed independently from the rest of the
# model components i.e. https://huggingface.co/thibaud/controlnet-sd21/
if "time_embed.0.weight" in checkpoint:
controlnet_state_dict = checkpoint
else:
controlnet_state_dict = {}
keys = list(checkpoint.keys())
controlnet_key = LDM_CONTROLNET_KEY
for key in keys:
if key.startswith(controlnet_key):
controlnet_state_dict[key.replace(controlnet_key, "")] = checkpoint.pop(key)
new_checkpoint = {}
ldm_controlnet_keys = DIFFUSERS_TO_LDM_MAPPING["controlnet"]["layers"]
for diffusers_key, ldm_key in ldm_controlnet_keys.items():
if ldm_key not in controlnet_state_dict:
continue
new_checkpoint[diffusers_key] = controlnet_state_dict[ldm_key]
# Retrieves the keys for the input blocks only
num_input_blocks = len(
{".".join(layer.split(".")[:2]) for layer in controlnet_state_dict if "input_blocks" in layer}
)
input_blocks = {
layer_id: [key for key in controlnet_state_dict if f"input_blocks.{layer_id}" in key]
for layer_id in range(num_input_blocks)
}
# Down blocks
for i in range(1, num_input_blocks):
block_id = (i - 1) // (config["layers_per_block"] + 1)
layer_in_block_id = (i - 1) % (config["layers_per_block"] + 1)
resnets = [
key for key in input_blocks[i] if f"input_blocks.{i}.0" in key and f"input_blocks.{i}.0.op" not in key
]
update_unet_resnet_ldm_to_diffusers(
resnets,
new_checkpoint,
controlnet_state_dict,
{"old": f"input_blocks.{i}.0", "new": f"down_blocks.{block_id}.resnets.{layer_in_block_id}"},
)
if f"input_blocks.{i}.0.op.weight" in controlnet_state_dict:
new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.weight"] = controlnet_state_dict.pop(
f"input_blocks.{i}.0.op.weight"
)
new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.bias"] = controlnet_state_dict.pop(
f"input_blocks.{i}.0.op.bias"
)
attentions = [key for key in input_blocks[i] if f"input_blocks.{i}.1" in key]
if attentions:
update_unet_attention_ldm_to_diffusers(
attentions,
new_checkpoint,
controlnet_state_dict,
{"old": f"input_blocks.{i}.1", "new": f"down_blocks.{block_id}.attentions.{layer_in_block_id}"},
)
# controlnet down blocks
for i in range(num_input_blocks):
new_checkpoint[f"controlnet_down_blocks.{i}.weight"] = controlnet_state_dict.pop(f"zero_convs.{i}.0.weight")
new_checkpoint[f"controlnet_down_blocks.{i}.bias"] = controlnet_state_dict.pop(f"zero_convs.{i}.0.bias")
# Retrieves the keys for the middle blocks only
num_middle_blocks = len(
{".".join(layer.split(".")[:2]) for layer in controlnet_state_dict if "middle_block" in layer}
)
middle_blocks = {
layer_id: [key for key in controlnet_state_dict if f"middle_block.{layer_id}" in key]
for layer_id in range(num_middle_blocks)
}
if middle_blocks:
resnet_0 = middle_blocks[0]
attentions = middle_blocks[1]
resnet_1 = middle_blocks[2]
update_unet_resnet_ldm_to_diffusers(
resnet_0,
new_checkpoint,
controlnet_state_dict,
mapping={"old": "middle_block.0", "new": "mid_block.resnets.0"},
)
update_unet_resnet_ldm_to_diffusers(
resnet_1,
new_checkpoint,
controlnet_state_dict,
mapping={"old": "middle_block.2", "new": "mid_block.resnets.1"},
)
update_unet_attention_ldm_to_diffusers(
attentions,
new_checkpoint,
controlnet_state_dict,
mapping={"old": "middle_block.1", "new": "mid_block.attentions.0"},
)
# mid block
new_checkpoint["controlnet_mid_block.weight"] = controlnet_state_dict.pop("middle_block_out.0.weight")
new_checkpoint["controlnet_mid_block.bias"] = controlnet_state_dict.pop("middle_block_out.0.bias")
# controlnet cond embedding blocks
cond_embedding_blocks = {
".".join(layer.split(".")[:2])
for layer in controlnet_state_dict
if "input_hint_block" in layer and ("input_hint_block.0" not in layer) and ("input_hint_block.14" not in layer)
}
num_cond_embedding_blocks = len(cond_embedding_blocks)
for idx in range(1, num_cond_embedding_blocks + 1):
diffusers_idx = idx - 1
cond_block_id = 2 * idx
new_checkpoint[f"controlnet_cond_embedding.blocks.{diffusers_idx}.weight"] = controlnet_state_dict.pop(
f"input_hint_block.{cond_block_id}.weight"
)
new_checkpoint[f"controlnet_cond_embedding.blocks.{diffusers_idx}.bias"] = controlnet_state_dict.pop(
f"input_hint_block.{cond_block_id}.bias"
)
return new_checkpoint
def create_diffusers_controlnet_model_from_ldm(
pipeline_class_name, original_config, checkpoint, upcast_attention=False, image_size=None
):
# import here to avoid circular imports
from ..models import ControlNetModel
image_size = set_image_size(pipeline_class_name, original_config, checkpoint, image_size=image_size)
diffusers_config = create_controlnet_diffusers_config(original_config, image_size=image_size)
diffusers_config["upcast_attention"] = upcast_attention
diffusers_format_controlnet_checkpoint = convert_controlnet_checkpoint(checkpoint, diffusers_config)
ctx = init_empty_weights if is_accelerate_available() else nullcontext
with ctx():
controlnet = ControlNetModel(**diffusers_config)
if is_accelerate_available():
for param_name, param in diffusers_format_controlnet_checkpoint.items():
set_module_tensor_to_device(controlnet, param_name, "cpu", value=param)
else:
controlnet.load_state_dict(diffusers_format_controlnet_checkpoint)
return {"controlnet": controlnet}
def update_vae_resnet_ldm_to_diffusers(keys, new_checkpoint, checkpoint, mapping):
for ldm_key in keys:
diffusers_key = ldm_key.replace(mapping["old"], mapping["new"]).replace("nin_shortcut", "conv_shortcut")
new_checkpoint[diffusers_key] = checkpoint.pop(ldm_key)
def update_vae_attentions_ldm_to_diffusers(keys, new_checkpoint, checkpoint, mapping):
for ldm_key in keys:
diffusers_key = (
ldm_key.replace(mapping["old"], mapping["new"])
.replace("norm.weight", "group_norm.weight")
.replace("norm.bias", "group_norm.bias")
.replace("q.weight", "to_q.weight")
.replace("q.bias", "to_q.bias")
.replace("k.weight", "to_k.weight")
.replace("k.bias", "to_k.bias")
.replace("v.weight", "to_v.weight")
.replace("v.bias", "to_v.bias")
.replace("proj_out.weight", "to_out.0.weight")
.replace("proj_out.bias", "to_out.0.bias")
)
new_checkpoint[diffusers_key] = checkpoint.pop(ldm_key)
# proj_attn.weight has to be converted from conv 1D to linear
shape = new_checkpoint[diffusers_key].shape
if len(shape) == 3:
new_checkpoint[diffusers_key] = new_checkpoint[diffusers_key][:, :, 0]
elif len(shape) == 4:
new_checkpoint[diffusers_key] = new_checkpoint[diffusers_key][:, :, 0, 0]
def convert_ldm_vae_checkpoint(checkpoint, config):
# extract state dict for VAE
# remove the LDM_VAE_KEY prefix from the ldm checkpoint keys so that it is easier to map them to diffusers keys
vae_state_dict = {}
keys = list(checkpoint.keys())
vae_key = LDM_VAE_KEY if any(k.startswith(LDM_VAE_KEY) for k in keys) else ""
for key in keys:
if key.startswith(vae_key):
vae_state_dict[key.replace(vae_key, "")] = checkpoint.get(key)
new_checkpoint = {}
vae_diffusers_ldm_map = DIFFUSERS_TO_LDM_MAPPING["vae"]
for diffusers_key, ldm_key in vae_diffusers_ldm_map.items():
if ldm_key not in vae_state_dict:
continue
new_checkpoint[diffusers_key] = vae_state_dict[ldm_key]
# Retrieves the keys for the encoder down blocks only
num_down_blocks = len(config["down_block_types"])
down_blocks = {
layer_id: [key for key in vae_state_dict if f"down.{layer_id}" in key] for layer_id in range(num_down_blocks)
}
for i in range(num_down_blocks):
resnets = [key for key in down_blocks[i] if f"down.{i}" in key and f"down.{i}.downsample" not in key]
update_vae_resnet_ldm_to_diffusers(
resnets,
new_checkpoint,
vae_state_dict,
mapping={"old": f"down.{i}.block", "new": f"down_blocks.{i}.resnets"},
)
if f"encoder.down.{i}.downsample.conv.weight" in vae_state_dict:
new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.weight"] = vae_state_dict.pop(
f"encoder.down.{i}.downsample.conv.weight"
)
new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.bias"] = vae_state_dict.pop(
f"encoder.down.{i}.downsample.conv.bias"
)
mid_resnets = [key for key in vae_state_dict if "encoder.mid.block" in key]
num_mid_res_blocks = 2
for i in range(1, num_mid_res_blocks + 1):
resnets = [key for key in mid_resnets if f"encoder.mid.block_{i}" in key]
update_vae_resnet_ldm_to_diffusers(
resnets,
new_checkpoint,
vae_state_dict,
mapping={"old": f"mid.block_{i}", "new": f"mid_block.resnets.{i - 1}"},
)
mid_attentions = [key for key in vae_state_dict if "encoder.mid.attn" in key]
update_vae_attentions_ldm_to_diffusers(
mid_attentions, new_checkpoint, vae_state_dict, mapping={"old": "mid.attn_1", "new": "mid_block.attentions.0"}
)
# Retrieves the keys for the decoder up blocks only
num_up_blocks = len(config["up_block_types"])
up_blocks = {
layer_id: [key for key in vae_state_dict if f"up.{layer_id}" in key] for layer_id in range(num_up_blocks)
}
for i in range(num_up_blocks):
block_id = num_up_blocks - 1 - i
resnets = [
key for key in up_blocks[block_id] if f"up.{block_id}" in key and f"up.{block_id}.upsample" not in key
]
update_vae_resnet_ldm_to_diffusers(
resnets,
new_checkpoint,
vae_state_dict,
mapping={"old": f"up.{block_id}.block", "new": f"up_blocks.{i}.resnets"},
)
if f"decoder.up.{block_id}.upsample.conv.weight" in vae_state_dict:
new_checkpoint[f"decoder.up_blocks.{i}.upsamplers.0.conv.weight"] = vae_state_dict[
f"decoder.up.{block_id}.upsample.conv.weight"
]
new_checkpoint[f"decoder.up_blocks.{i}.upsamplers.0.conv.bias"] = vae_state_dict[
f"decoder.up.{block_id}.upsample.conv.bias"
]
mid_resnets = [key for key in vae_state_dict if "decoder.mid.block" in key]
num_mid_res_blocks = 2
for i in range(1, num_mid_res_blocks + 1):
resnets = [key for key in mid_resnets if f"decoder.mid.block_{i}" in key]
update_vae_resnet_ldm_to_diffusers(
resnets,
new_checkpoint,
vae_state_dict,
mapping={"old": f"mid.block_{i}", "new": f"mid_block.resnets.{i - 1}"},