{{ message }}
Repository navigation
Expand file tree
/
Copy pathcontrolnet_xs.py
More file actions
1945 lines (1679 loc) · 84.8 KB
/
Copy pathcontrolnet_xs.py
File metadata and controls
1945 lines (1679 loc) · 84.8 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
# Copyright 2024 The HuggingFace Team. All rights reserved.
#
# 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.
from dataclasses import dataclass
from math import gcd
from typing import Any, Dict, List, Optional, Tuple, Union
import torch
import torch.utils.checkpoint
from torch import Tensor, nn
from ..configuration_utils import ConfigMixin, register_to_config
from ..utils import BaseOutput, is_torch_version, logging
from ..utils.torch_utils import apply_freeu
from .attention_processor import (
ADDED_KV_ATTENTION_PROCESSORS,
CROSS_ATTENTION_PROCESSORS,
Attention,
AttentionProcessor,
AttnAddedKVProcessor,
AttnProcessor,
FusedAttnProcessor2_0,
)
from .controlnet import ControlNetConditioningEmbedding
from .embeddings import TimestepEmbedding, Timesteps
from .modeling_utils import ModelMixin
from .unets.unet_2d_blocks import (
CrossAttnDownBlock2D,
CrossAttnUpBlock2D,
Downsample2D,
ResnetBlock2D,
Transformer2DModel,
UNetMidBlock2DCrossAttn,
Upsample2D,
)
from .unets.unet_2d_condition import UNet2DConditionModel
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
@dataclass
class ControlNetXSOutput(BaseOutput):
"""
The output of [`UNetControlNetXSModel`].
Args:
sample (`Tensor` of shape `(batch_size, num_channels, height, width)`):
The output of the `UNetControlNetXSModel`. Unlike `ControlNetOutput` this is NOT to be added to the base
model output, but is already the final output.
"""
sample: Tensor = None
class DownBlockControlNetXSAdapter(nn.Module):
"""Components that together with corresponding components from the base model will form a
`ControlNetXSCrossAttnDownBlock2D`"""
def __init__(
self,
resnets: nn.ModuleList,
base_to_ctrl: nn.ModuleList,
ctrl_to_base: nn.ModuleList,
attentions: Optional[nn.ModuleList] = None,
downsampler: Optional[nn.Conv2d] = None,
):
super().__init__()
self.resnets = resnets
self.base_to_ctrl = base_to_ctrl
self.ctrl_to_base = ctrl_to_base
self.attentions = attentions
self.downsamplers = downsampler
class MidBlockControlNetXSAdapter(nn.Module):
"""Components that together with corresponding components from the base model will form a
`ControlNetXSCrossAttnMidBlock2D`"""
def __init__(self, midblock: UNetMidBlock2DCrossAttn, base_to_ctrl: nn.ModuleList, ctrl_to_base: nn.ModuleList):
super().__init__()
self.midblock = midblock
self.base_to_ctrl = base_to_ctrl
self.ctrl_to_base = ctrl_to_base
class UpBlockControlNetXSAdapter(nn.Module):
"""Components that together with corresponding components from the base model will form a `ControlNetXSCrossAttnUpBlock2D`"""
def __init__(self, ctrl_to_base: nn.ModuleList):
super().__init__()
self.ctrl_to_base = ctrl_to_base
def get_down_block_adapter(
base_in_channels: int,
base_out_channels: int,
ctrl_in_channels: int,
ctrl_out_channels: int,
temb_channels: int,
max_norm_num_groups: Optional[int] = 32,
has_crossattn=True,
transformer_layers_per_block: Optional[Union[int, Tuple[int]]] = 1,
num_attention_heads: Optional[int] = 1,
cross_attention_dim: Optional[int] = 1024,
add_downsample: bool = True,
upcast_attention: Optional[bool] = False,
use_linear_projection: Optional[bool] = True,
):
num_layers = 2 # only support sd + sdxl
resnets = []
attentions = []
ctrl_to_base = []
base_to_ctrl = []
if isinstance(transformer_layers_per_block, int):
transformer_layers_per_block = [transformer_layers_per_block] * num_layers
for i in range(num_layers):
base_in_channels = base_in_channels if i == 0 else base_out_channels
ctrl_in_channels = ctrl_in_channels if i == 0 else ctrl_out_channels
# Before the resnet/attention application, information is concatted from base to control.
# Concat doesn't require change in number of channels
base_to_ctrl.append(make_zero_conv(base_in_channels, base_in_channels))
resnets.append(
ResnetBlock2D(
in_channels=ctrl_in_channels + base_in_channels, # information from base is concatted to ctrl
out_channels=ctrl_out_channels,
temb_channels=temb_channels,
groups=find_largest_factor(ctrl_in_channels + base_in_channels, max_factor=max_norm_num_groups),
groups_out=find_largest_factor(ctrl_out_channels, max_factor=max_norm_num_groups),
eps=1e-5,
)
)
if has_crossattn:
attentions.append(
Transformer2DModel(
num_attention_heads,
ctrl_out_channels // num_attention_heads,
in_channels=ctrl_out_channels,
num_layers=transformer_layers_per_block[i],
cross_attention_dim=cross_attention_dim,
use_linear_projection=use_linear_projection,
upcast_attention=upcast_attention,
norm_num_groups=find_largest_factor(ctrl_out_channels, max_factor=max_norm_num_groups),
)
)
# After the resnet/attention application, information is added from control to base
# Addition requires change in number of channels
ctrl_to_base.append(make_zero_conv(ctrl_out_channels, base_out_channels))
if add_downsample:
# Before the downsampler application, information is concatted from base to control
# Concat doesn't require change in number of channels
base_to_ctrl.append(make_zero_conv(base_out_channels, base_out_channels))
downsamplers = Downsample2D(
ctrl_out_channels + base_out_channels, use_conv=True, out_channels=ctrl_out_channels, name="op"
)
# After the downsampler application, information is added from control to base
# Addition requires change in number of channels
ctrl_to_base.append(make_zero_conv(ctrl_out_channels, base_out_channels))
else:
downsamplers = None
down_block_components = DownBlockControlNetXSAdapter(
resnets=nn.ModuleList(resnets),
base_to_ctrl=nn.ModuleList(base_to_ctrl),
ctrl_to_base=nn.ModuleList(ctrl_to_base),
)
if has_crossattn:
down_block_components.attentions = nn.ModuleList(attentions)
if downsamplers is not None:
down_block_components.downsamplers = downsamplers
return down_block_components
def get_mid_block_adapter(
base_channels: int,
ctrl_channels: int,
temb_channels: Optional[int] = None,
max_norm_num_groups: Optional[int] = 32,
transformer_layers_per_block: int = 1,
num_attention_heads: Optional[int] = 1,
cross_attention_dim: Optional[int] = 1024,
upcast_attention: bool = False,
use_linear_projection: bool = True,
):
# Before the midblock application, information is concatted from base to control.
# Concat doesn't require change in number of channels
base_to_ctrl = make_zero_conv(base_channels, base_channels)
midblock = UNetMidBlock2DCrossAttn(
transformer_layers_per_block=transformer_layers_per_block,
in_channels=ctrl_channels + base_channels,
out_channels=ctrl_channels,
temb_channels=temb_channels,
# number or norm groups must divide both in_channels and out_channels
resnet_groups=find_largest_factor(gcd(ctrl_channels, ctrl_channels + base_channels), max_norm_num_groups),
cross_attention_dim=cross_attention_dim,
num_attention_heads=num_attention_heads,
use_linear_projection=use_linear_projection,
upcast_attention=upcast_attention,
)
# After the midblock application, information is added from control to base
# Addition requires change in number of channels
ctrl_to_base = make_zero_conv(ctrl_channels, base_channels)
return MidBlockControlNetXSAdapter(base_to_ctrl=base_to_ctrl, midblock=midblock, ctrl_to_base=ctrl_to_base)
def get_up_block_adapter(
out_channels: int,
prev_output_channel: int,
ctrl_skip_channels: List[int],
):
ctrl_to_base = []
num_layers = 3 # only support sd + sdxl
for i in range(num_layers):
resnet_in_channels = prev_output_channel if i == 0 else out_channels
ctrl_to_base.append(make_zero_conv(ctrl_skip_channels[i], resnet_in_channels))
return UpBlockControlNetXSAdapter(ctrl_to_base=nn.ModuleList(ctrl_to_base))
class ControlNetXSAdapter(ModelMixin, ConfigMixin):
r"""
A `ControlNetXSAdapter` model. To use it, pass it into a `UNetControlNetXSModel` (together with a
`UNet2DConditionModel` base model).
This model inherits from [`ModelMixin`] and [`ConfigMixin`]. Check the superclass documentation for it's generic
methods implemented for all models (such as downloading or saving).
Like `UNetControlNetXSModel`, `ControlNetXSAdapter` is compatible with StableDiffusion and StableDiffusion-XL. It's
default parameters are compatible with StableDiffusion.
Parameters:
conditioning_channels (`int`, defaults to 3):
Number of channels of conditioning input (e.g. an image)
conditioning_channel_order (`str`, defaults to `"rgb"`):
The channel order of conditional image. Will convert to `rgb` if it's `bgr`.
conditioning_embedding_out_channels (`tuple[int]`, defaults to `(16, 32, 96, 256)`):
The tuple of output channels for each block in the `controlnet_cond_embedding` layer.
time_embedding_mix (`float`, defaults to 1.0):
If 0, then only the control adapters's time embedding is used. If 1, then only the base unet's time
embedding is used. Otherwise, both are combined.
learn_time_embedding (`bool`, defaults to `False`):
Whether a time embedding should be learned. If yes, `UNetControlNetXSModel` will combine the time
embeddings of the base model and the control adapter. If no, `UNetControlNetXSModel` will use the base
model's time embedding.
num_attention_heads (`list[int]`, defaults to `[4]`):
The number of attention heads.
block_out_channels (`list[int]`, defaults to `[4, 8, 16, 16]`):
The tuple of output channels for each block.
base_block_out_channels (`list[int]`, defaults to `[320, 640, 1280, 1280]`):
The tuple of output channels for each block in the base unet.
cross_attention_dim (`int`, defaults to 1024):
The dimension of the cross attention features.
down_block_types (`list[str]`, defaults to `["CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D"]`):
The tuple of downsample blocks to use.
sample_size (`int`, defaults to 96):
Height and width of input/output sample.
transformer_layers_per_block (`Union[int, Tuple[int]]`, defaults to 1):
The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
[`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
upcast_attention (`bool`, defaults to `True`):
Whether the attention computation should always be upcasted.
max_norm_num_groups (`int`, defaults to 32):
Maximum number of groups in group normal. The actual number will be the largest divisor of the respective
channels, that is <= max_norm_num_groups.
"""
@register_to_config
def __init__(
self,
conditioning_channels: int = 3,
conditioning_channel_order: str = "rgb",
conditioning_embedding_out_channels: Tuple[int] = (16, 32, 96, 256),
time_embedding_mix: float = 1.0,
learn_time_embedding: bool = False,
num_attention_heads: Union[int, Tuple[int]] = 4,
block_out_channels: Tuple[int] = (4, 8, 16, 16),
base_block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
cross_attention_dim: int = 1024,
down_block_types: Tuple[str] = (
"CrossAttnDownBlock2D",
"CrossAttnDownBlock2D",
"CrossAttnDownBlock2D",
"DownBlock2D",
),
sample_size: Optional[int] = 96,
transformer_layers_per_block: Union[int, Tuple[int]] = 1,
upcast_attention: bool = True,
max_norm_num_groups: int = 32,
use_linear_projection: bool = True,
):
super().__init__()
time_embedding_input_dim = base_block_out_channels[0]
time_embedding_dim = base_block_out_channels[0] * 4
# Check inputs
if conditioning_channel_order not in ["rgb", "bgr"]:
raise ValueError(f"unknown `conditioning_channel_order`: {conditioning_channel_order}")
if len(block_out_channels) != len(down_block_types):
raise ValueError(
f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
)
if not isinstance(transformer_layers_per_block, (list, tuple)):
transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
if not isinstance(cross_attention_dim, (list, tuple)):
cross_attention_dim = [cross_attention_dim] * len(down_block_types)
# see https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131 for why `ControlNetXSAdapter` takes `num_attention_heads` instead of `attention_head_dim`
if not isinstance(num_attention_heads, (list, tuple)):
num_attention_heads = [num_attention_heads] * len(down_block_types)
if len(num_attention_heads) != len(down_block_types):
raise ValueError(
f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
)
# 5 - Create conditioning hint embedding
self.controlnet_cond_embedding = ControlNetConditioningEmbedding(
conditioning_embedding_channels=block_out_channels[0],
block_out_channels=conditioning_embedding_out_channels,
conditioning_channels=conditioning_channels,
)
# time
if learn_time_embedding:
self.time_embedding = TimestepEmbedding(time_embedding_input_dim, time_embedding_dim)
else:
self.time_embedding = None
self.down_blocks = nn.ModuleList([])
self.up_connections = nn.ModuleList([])
# input
self.conv_in = nn.Conv2d(4, block_out_channels[0], kernel_size=3, padding=1)
self.control_to_base_for_conv_in = make_zero_conv(block_out_channels[0], base_block_out_channels[0])
# down
base_out_channels = base_block_out_channels[0]
ctrl_out_channels = block_out_channels[0]
for i, down_block_type in enumerate(down_block_types):
base_in_channels = base_out_channels
base_out_channels = base_block_out_channels[i]
ctrl_in_channels = ctrl_out_channels
ctrl_out_channels = block_out_channels[i]
has_crossattn = "CrossAttn" in down_block_type
is_final_block = i == len(down_block_types) - 1
self.down_blocks.append(
get_down_block_adapter(
base_in_channels=base_in_channels,
base_out_channels=base_out_channels,
ctrl_in_channels=ctrl_in_channels,
ctrl_out_channels=ctrl_out_channels,
temb_channels=time_embedding_dim,
max_norm_num_groups=max_norm_num_groups,
has_crossattn=has_crossattn,
transformer_layers_per_block=transformer_layers_per_block[i],
num_attention_heads=num_attention_heads[i],
cross_attention_dim=cross_attention_dim[i],
add_downsample=not is_final_block,
upcast_attention=upcast_attention,
use_linear_projection=use_linear_projection,
)
)
# mid
self.mid_block = get_mid_block_adapter(
base_channels=base_block_out_channels[-1],
ctrl_channels=block_out_channels[-1],
temb_channels=time_embedding_dim,
transformer_layers_per_block=transformer_layers_per_block[-1],
num_attention_heads=num_attention_heads[-1],
cross_attention_dim=cross_attention_dim[-1],
upcast_attention=upcast_attention,
use_linear_projection=use_linear_projection,
)
# up
# The skip connection channels are the output of the conv_in and of all the down subblocks
ctrl_skip_channels = [block_out_channels[0]]
for i, out_channels in enumerate(block_out_channels):
number_of_subblocks = (
3 if i < len(block_out_channels) - 1 else 2
) # every block has 3 subblocks, except last one, which has 2 as it has no downsampler
ctrl_skip_channels.extend([out_channels] * number_of_subblocks)
reversed_base_block_out_channels = list(reversed(base_block_out_channels))
base_out_channels = reversed_base_block_out_channels[0]
for i in range(len(down_block_types)):
prev_base_output_channel = base_out_channels
base_out_channels = reversed_base_block_out_channels[i]
ctrl_skip_channels_ = [ctrl_skip_channels.pop() for _ in range(3)]
self.up_connections.append(
get_up_block_adapter(
out_channels=base_out_channels,
prev_output_channel=prev_base_output_channel,
ctrl_skip_channels=ctrl_skip_channels_,
)
)
@classmethod
def from_unet(
cls,
unet: UNet2DConditionModel,
size_ratio: Optional[float] = None,
block_out_channels: Optional[List[int]] = None,
num_attention_heads: Optional[List[int]] = None,
learn_time_embedding: bool = False,
time_embedding_mix: int = 1.0,
conditioning_channels: int = 3,
conditioning_channel_order: str = "rgb",
conditioning_embedding_out_channels: Tuple[int] = (16, 32, 96, 256),
):
r"""
Instantiate a [`ControlNetXSAdapter`] from a [`UNet2DConditionModel`].
Parameters:
unet (`UNet2DConditionModel`):
The UNet model we want to control. The dimensions of the ControlNetXSAdapter will be adapted to it.
size_ratio (float, *optional*, defaults to `None`):
When given, block_out_channels is set to a fraction of the base model's block_out_channels. Either this
or `block_out_channels` must be given.
block_out_channels (`List[int]`, *optional*, defaults to `None`):
Down blocks output channels in control model. Either this or `size_ratio` must be given.
num_attention_heads (`List[int]`, *optional*, defaults to `None`):
The dimension of the attention heads. The naming seems a bit confusing and it is, see
https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131 for why.
learn_time_embedding (`bool`, defaults to `False`):
Whether the `ControlNetXSAdapter` should learn a time embedding.
time_embedding_mix (`float`, defaults to 1.0):
If 0, then only the control adapter's time embedding is used. If 1, then only the base unet's time
embedding is used. Otherwise, both are combined.
conditioning_channels (`int`, defaults to 3):
Number of channels of conditioning input (e.g. an image)
conditioning_channel_order (`str`, defaults to `"rgb"`):
The channel order of conditional image. Will convert to `rgb` if it's `bgr`.
conditioning_embedding_out_channels (`Tuple[int]`, defaults to `(16, 32, 96, 256)`):
The tuple of output channel for each block in the `controlnet_cond_embedding` layer.
"""
# Check input
fixed_size = block_out_channels is not None
relative_size = size_ratio is not None
if not (fixed_size ^ relative_size):
raise ValueError(
"Pass exactly one of `block_out_channels` (for absolute sizing) or `size_ratio` (for relative sizing)."
)
# Create model
block_out_channels = block_out_channels or [int(b * size_ratio) for b in unet.config.block_out_channels]
if num_attention_heads is None:
# The naming seems a bit confusing and it is, see https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131 for why.
num_attention_heads = unet.config.attention_head_dim
model = cls(
conditioning_channels=conditioning_channels,
conditioning_channel_order=conditioning_channel_order,
conditioning_embedding_out_channels=conditioning_embedding_out_channels,
time_embedding_mix=time_embedding_mix,
learn_time_embedding=learn_time_embedding,
num_attention_heads=num_attention_heads,
block_out_channels=block_out_channels,
base_block_out_channels=unet.config.block_out_channels,
cross_attention_dim=unet.config.cross_attention_dim,
down_block_types=unet.config.down_block_types,
sample_size=unet.config.sample_size,
transformer_layers_per_block=unet.config.transformer_layers_per_block,
upcast_attention=unet.config.upcast_attention,
max_norm_num_groups=unet.config.norm_num_groups,
use_linear_projection=unet.config.use_linear_projection,
)
# ensure that the ControlNetXSAdapter is the same dtype as the UNet2DConditionModel
model.to(unet.dtype)
return model
def forward(self, *args, **kwargs):
raise ValueError(
"A ControlNetXSAdapter cannot be run by itself. Use it together with a UNet2DConditionModel to instantiate a UNetControlNetXSModel."
)
class UNetControlNetXSModel(ModelMixin, ConfigMixin):
r"""
A UNet fused with a ControlNet-XS adapter model
This model inherits from [`ModelMixin`] and [`ConfigMixin`]. Check the superclass documentation for it's generic
methods implemented for all models (such as downloading or saving).
`UNetControlNetXSModel` is compatible with StableDiffusion and StableDiffusion-XL. It's default parameters are
compatible with StableDiffusion.
It's parameters are either passed to the underlying `UNet2DConditionModel` or used exactly like in
`ControlNetXSAdapter` . See their documentation for details.
"""
_supports_gradient_checkpointing = True
@register_to_config
def __init__(
self,
# unet configs
sample_size: Optional[int] = 96,
down_block_types: Tuple[str] = (
"CrossAttnDownBlock2D",
"CrossAttnDownBlock2D",
"CrossAttnDownBlock2D",
"DownBlock2D",
),
up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"),
block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
norm_num_groups: Optional[int] = 32,
cross_attention_dim: Union[int, Tuple[int]] = 1024,
transformer_layers_per_block: Union[int, Tuple[int]] = 1,
num_attention_heads: Union[int, Tuple[int]] = 8,
addition_embed_type: Optional[str] = None,
addition_time_embed_dim: Optional[int] = None,
upcast_attention: bool = True,
use_linear_projection: bool = True,
time_cond_proj_dim: Optional[int] = None,
projection_class_embeddings_input_dim: Optional[int] = None,
# additional controlnet configs
time_embedding_mix: float = 1.0,
ctrl_conditioning_channels: int = 3,
ctrl_conditioning_embedding_out_channels: Tuple[int] = (16, 32, 96, 256),
ctrl_conditioning_channel_order: str = "rgb",
ctrl_learn_time_embedding: bool = False,
ctrl_block_out_channels: Tuple[int] = (4, 8, 16, 16),
ctrl_num_attention_heads: Union[int, Tuple[int]] = 4,
ctrl_max_norm_num_groups: int = 32,
):
super().__init__()
if time_embedding_mix < 0 or time_embedding_mix > 1:
raise ValueError("`time_embedding_mix` needs to be between 0 and 1.")
if time_embedding_mix < 1 and not ctrl_learn_time_embedding:
raise ValueError("To use `time_embedding_mix` < 1, `ctrl_learn_time_embedding` must be `True`")
if addition_embed_type is not None and addition_embed_type != "text_time":
raise ValueError(
"As `UNetControlNetXSModel` currently only supports StableDiffusion and StableDiffusion-XL, `addition_embed_type` must be `None` or `'text_time'`."
)
if not isinstance(transformer_layers_per_block, (list, tuple)):
transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
if not isinstance(cross_attention_dim, (list, tuple)):
cross_attention_dim = [cross_attention_dim] * len(down_block_types)
if not isinstance(num_attention_heads, (list, tuple)):
num_attention_heads = [num_attention_heads] * len(down_block_types)
if not isinstance(ctrl_num_attention_heads, (list, tuple)):
ctrl_num_attention_heads = [ctrl_num_attention_heads] * len(down_block_types)
base_num_attention_heads = num_attention_heads
self.in_channels = 4
# # Input
self.base_conv_in = nn.Conv2d(4, block_out_channels[0], kernel_size=3, padding=1)
self.controlnet_cond_embedding = ControlNetConditioningEmbedding(
conditioning_embedding_channels=ctrl_block_out_channels[0],
block_out_channels=ctrl_conditioning_embedding_out_channels,
conditioning_channels=ctrl_conditioning_channels,
)
self.ctrl_conv_in = nn.Conv2d(4, ctrl_block_out_channels[0], kernel_size=3, padding=1)
self.control_to_base_for_conv_in = make_zero_conv(ctrl_block_out_channels[0], block_out_channels[0])
# # Time
time_embed_input_dim = block_out_channels[0]
time_embed_dim = block_out_channels[0] * 4
self.base_time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos=True, downscale_freq_shift=0)
self.base_time_embedding = TimestepEmbedding(
time_embed_input_dim,
time_embed_dim,
cond_proj_dim=time_cond_proj_dim,
)
if ctrl_learn_time_embedding:
self.ctrl_time_embedding = TimestepEmbedding(
in_channels=time_embed_input_dim, time_embed_dim=time_embed_dim
)
else:
self.ctrl_time_embedding = None
if addition_embed_type is None:
self.base_add_time_proj = None
self.base_add_embedding = None
else:
self.base_add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos=True, downscale_freq_shift=0)
self.base_add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
# # Create down blocks
down_blocks = []
base_out_channels = block_out_channels[0]
ctrl_out_channels = ctrl_block_out_channels[0]
for i, down_block_type in enumerate(down_block_types):
base_in_channels = base_out_channels
base_out_channels = block_out_channels[i]
ctrl_in_channels = ctrl_out_channels
ctrl_out_channels = ctrl_block_out_channels[i]
has_crossattn = "CrossAttn" in down_block_type
is_final_block = i == len(down_block_types) - 1
down_blocks.append(
ControlNetXSCrossAttnDownBlock2D(
base_in_channels=base_in_channels,
base_out_channels=base_out_channels,
ctrl_in_channels=ctrl_in_channels,
ctrl_out_channels=ctrl_out_channels,
temb_channels=time_embed_dim,
norm_num_groups=norm_num_groups,
ctrl_max_norm_num_groups=ctrl_max_norm_num_groups,
has_crossattn=has_crossattn,
transformer_layers_per_block=transformer_layers_per_block[i],
base_num_attention_heads=base_num_attention_heads[i],
ctrl_num_attention_heads=ctrl_num_attention_heads[i],
cross_attention_dim=cross_attention_dim[i],
add_downsample=not is_final_block,
upcast_attention=upcast_attention,
use_linear_projection=use_linear_projection,
)
)
# # Create mid block
self.mid_block = ControlNetXSCrossAttnMidBlock2D(
base_channels=block_out_channels[-1],
ctrl_channels=ctrl_block_out_channels[-1],
temb_channels=time_embed_dim,
norm_num_groups=norm_num_groups,
ctrl_max_norm_num_groups=ctrl_max_norm_num_groups,
transformer_layers_per_block=transformer_layers_per_block[-1],
base_num_attention_heads=base_num_attention_heads[-1],
ctrl_num_attention_heads=ctrl_num_attention_heads[-1],
cross_attention_dim=cross_attention_dim[-1],
upcast_attention=upcast_attention,
use_linear_projection=use_linear_projection,
)
# # Create up blocks
up_blocks = []
rev_transformer_layers_per_block = list(reversed(transformer_layers_per_block))
rev_num_attention_heads = list(reversed(base_num_attention_heads))
rev_cross_attention_dim = list(reversed(cross_attention_dim))
# The skip connection channels are the output of the conv_in and of all the down subblocks
ctrl_skip_channels = [ctrl_block_out_channels[0]]
for i, out_channels in enumerate(ctrl_block_out_channels):
number_of_subblocks = (
3 if i < len(ctrl_block_out_channels) - 1 else 2
) # every block has 3 subblocks, except last one, which has 2 as it has no downsampler
ctrl_skip_channels.extend([out_channels] * number_of_subblocks)
reversed_block_out_channels = list(reversed(block_out_channels))
out_channels = reversed_block_out_channels[0]
for i, up_block_type in enumerate(up_block_types):
prev_output_channel = out_channels
out_channels = reversed_block_out_channels[i]
in_channels = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
ctrl_skip_channels_ = [ctrl_skip_channels.pop() for _ in range(3)]
has_crossattn = "CrossAttn" in up_block_type
is_final_block = i == len(block_out_channels) - 1
up_blocks.append(
ControlNetXSCrossAttnUpBlock2D(
in_channels=in_channels,
out_channels=out_channels,
prev_output_channel=prev_output_channel,
ctrl_skip_channels=ctrl_skip_channels_,
temb_channels=time_embed_dim,
resolution_idx=i,
has_crossattn=has_crossattn,
transformer_layers_per_block=rev_transformer_layers_per_block[i],
num_attention_heads=rev_num_attention_heads[i],
cross_attention_dim=rev_cross_attention_dim[i],
add_upsample=not is_final_block,
upcast_attention=upcast_attention,
norm_num_groups=norm_num_groups,
use_linear_projection=use_linear_projection,
)
)
self.down_blocks = nn.ModuleList(down_blocks)
self.up_blocks = nn.ModuleList(up_blocks)
self.base_conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=norm_num_groups)
self.base_conv_act = nn.SiLU()
self.base_conv_out = nn.Conv2d(block_out_channels[0], 4, kernel_size=3, padding=1)
@classmethod
def from_unet(
cls,
unet: UNet2DConditionModel,
controlnet: Optional[ControlNetXSAdapter] = None,
size_ratio: Optional[float] = None,
ctrl_block_out_channels: Optional[List[float]] = None,
time_embedding_mix: Optional[float] = None,
ctrl_optional_kwargs: Optional[Dict] = None,
):
r"""
Instantiate a [`UNetControlNetXSModel`] from a [`UNet2DConditionModel`] and an optional [`ControlNetXSAdapter`]
.
Parameters:
unet (`UNet2DConditionModel`):
The UNet model we want to control.
controlnet (`ControlNetXSAdapter`):
The ConntrolNet-XS adapter with which the UNet will be fused. If none is given, a new ConntrolNet-XS
adapter will be created.
size_ratio (float, *optional*, defaults to `None`):
Used to contruct the controlnet if none is given. See [`ControlNetXSAdapter.from_unet`] for details.
ctrl_block_out_channels (`List[int]`, *optional*, defaults to `None`):
Used to contruct the controlnet if none is given. See [`ControlNetXSAdapter.from_unet`] for details,
where this parameter is called `block_out_channels`.
time_embedding_mix (`float`, *optional*, defaults to None):
Used to contruct the controlnet if none is given. See [`ControlNetXSAdapter.from_unet`] for details.
ctrl_optional_kwargs (`Dict`, *optional*, defaults to `None`):
Passed to the `init` of the new controlent if no controlent was given.
"""
if controlnet is None:
controlnet = ControlNetXSAdapter.from_unet(
unet, size_ratio, ctrl_block_out_channels, **ctrl_optional_kwargs
)
else:
if any(
o is not None for o in (size_ratio, ctrl_block_out_channels, time_embedding_mix, ctrl_optional_kwargs)
):
raise ValueError(
"When a controlnet is passed, none of these parameters should be passed: size_ratio, ctrl_block_out_channels, time_embedding_mix, ctrl_optional_kwargs."
)
# # get params
params_for_unet = [
"sample_size",
"down_block_types",
"up_block_types",
"block_out_channels",
"norm_num_groups",
"cross_attention_dim",
"transformer_layers_per_block",
"addition_embed_type",
"addition_time_embed_dim",
"upcast_attention",
"use_linear_projection",
"time_cond_proj_dim",
"projection_class_embeddings_input_dim",
]
params_for_unet = {k: v for k, v in unet.config.items() if k in params_for_unet}
# The naming seems a bit confusing and it is, see https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131 for why.
params_for_unet["num_attention_heads"] = unet.config.attention_head_dim
params_for_controlnet = [
"conditioning_channels",
"conditioning_embedding_out_channels",
"conditioning_channel_order",
"learn_time_embedding",
"block_out_channels",
"num_attention_heads",
"max_norm_num_groups",
]
params_for_controlnet = {"ctrl_" + k: v for k, v in controlnet.config.items() if k in params_for_controlnet}
params_for_controlnet["time_embedding_mix"] = controlnet.config.time_embedding_mix
# # create model
model = cls.from_config({**params_for_unet, **params_for_controlnet})
# # load weights
# from unet
modules_from_unet = [
"time_embedding",
"conv_in",
"conv_norm_out",
"conv_out",
]
for m in modules_from_unet:
getattr(model, "base_" + m).load_state_dict(getattr(unet, m).state_dict())
optional_modules_from_unet = [
"add_time_proj",
"add_embedding",
]
for m in optional_modules_from_unet:
if hasattr(unet, m) and getattr(unet, m) is not None:
getattr(model, "base_" + m).load_state_dict(getattr(unet, m).state_dict())
# from controlnet
model.controlnet_cond_embedding.load_state_dict(controlnet.controlnet_cond_embedding.state_dict())
model.ctrl_conv_in.load_state_dict(controlnet.conv_in.state_dict())
if controlnet.time_embedding is not None:
model.ctrl_time_embedding.load_state_dict(controlnet.time_embedding.state_dict())
model.control_to_base_for_conv_in.load_state_dict(controlnet.control_to_base_for_conv_in.state_dict())
# from both
model.down_blocks = nn.ModuleList(
ControlNetXSCrossAttnDownBlock2D.from_modules(b, c)
for b, c in zip(unet.down_blocks, controlnet.down_blocks)
)
model.mid_block = ControlNetXSCrossAttnMidBlock2D.from_modules(unet.mid_block, controlnet.mid_block)
model.up_blocks = nn.ModuleList(
ControlNetXSCrossAttnUpBlock2D.from_modules(b, c)
for b, c in zip(unet.up_blocks, controlnet.up_connections)
)
# ensure that the UNetControlNetXSModel is the same dtype as the UNet2DConditionModel
model.to(unet.dtype)
return model
def freeze_unet_params(self) -> None:
"""Freeze the weights of the parts belonging to the base UNet2DConditionModel, and leave everything else unfrozen for fine
tuning."""
# Freeze everything
for param in self.parameters():
param.requires_grad = True
# Unfreeze ControlNetXSAdapter
base_parts = [
"base_time_proj",
"base_time_embedding",
"base_add_time_proj",
"base_add_embedding",
"base_conv_in",
"base_conv_norm_out",
"base_conv_act",
"base_conv_out",
]
base_parts = [getattr(self, part) for part in base_parts if getattr(self, part) is not None]
for part in base_parts:
for param in part.parameters():
param.requires_grad = False
for d in self.down_blocks:
d.freeze_base_params()
self.mid_block.freeze_base_params()
for u in self.up_blocks:
u.freeze_base_params()
def _set_gradient_checkpointing(self, module, value=False):
if hasattr(module, "gradient_checkpointing"):
module.gradient_checkpointing = value
@property
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
def attn_processors(self) -> Dict[str, AttentionProcessor]:
r"""
Returns:
`dict` of attention processors: A dictionary containing all attention processors used in the model with
indexed by its weight name.
"""
# set recursively
processors = {}
def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
if hasattr(module, "get_processor"):
processors[f"{name}.processor"] = module.get_processor()
for sub_name, child in module.named_children():
fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
return processors
for name, module in self.named_children():
fn_recursive_add_processors(name, module, processors)
return processors
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
r"""
Sets the attention processor to use to compute attention.
Parameters:
processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
The instantiated processor class or a dictionary of processor classes that will be set as the processor
for **all** `Attention` layers.
If `processor` is a dict, the key needs to define the path to the corresponding cross attention
processor. This is strongly recommended when setting trainable attention processors.
"""
count = len(self.attn_processors.keys())
if isinstance(processor, dict) and len(processor) != count:
raise ValueError(
f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
)
def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
if hasattr(module, "set_processor"):
if not isinstance(processor, dict):
module.set_processor(processor)
else:
module.set_processor(processor.pop(f"{name}.processor"))
for sub_name, child in module.named_children():
fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
for name, module in self.named_children():
fn_recursive_attn_processor(name, module, processor)
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
def set_default_attn_processor(self):
"""
Disables custom attention processors and sets the default attention implementation.
"""
if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
processor = AttnAddedKVProcessor()
elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
processor = AttnProcessor()
else:
raise ValueError(
f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
)
self.set_attn_processor(processor)
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.enable_freeu
def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
The suffixes after the scaling factors represent the stage blocks where they are being applied.
Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
Args:
s1 (`float`):
Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
mitigate the "oversmoothing effect" in the enhanced denoising process.
s2 (`float`):
Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
mitigate the "oversmoothing effect" in the enhanced denoising process.
b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
"""
for i, upsample_block in enumerate(self.up_blocks):
setattr(upsample_block, "s1", s1)
setattr(upsample_block, "s2", s2)
setattr(upsample_block, "b1", b1)
setattr(upsample_block, "b2", b2)
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.disable_freeu
def disable_freeu(self):
"""Disables the FreeU mechanism."""
freeu_keys = {"s1", "s2", "b1", "b2"}
for i, upsample_block in enumerate(self.up_blocks):
for k in freeu_keys:
if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None:
setattr(upsample_block, k, None)
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections
def fuse_qkv_projections(self):
"""
Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
are fused. For cross-attention modules, key and value projection matrices are fused.
<Tip warning={true}>
This API is 🧪 experimental.
</Tip>
"""
self.original_attn_processors = None
for _, attn_processor in self.attn_processors.items():
if "Added" in str(attn_processor.__class__.__name__):
raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
self.original_attn_processors = self.attn_processors
You can’t perform that action at this time.