File size: 66,332 Bytes
68f9d7d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
249a7a1
 
 
 
 
 
 
 
c71b83b
 
 
 
 
 
 
 
 
 
249a7a1
 
 
c71b83b
 
249a7a1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c71b83b
249a7a1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c71b83b
249a7a1
c71b83b
 
249a7a1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c71b83b
249a7a1
 
 
 
 
 
 
 
 
 
 
 
 
c71b83b
 
 
 
 
249a7a1
c71b83b
 
 
 
249a7a1
 
 
 
 
 
 
 
c71b83b
249a7a1
 
 
c71b83b
 
249a7a1
 
 
 
 
 
 
 
 
 
c71b83b
249a7a1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c71b83b
249a7a1
 
 
 
 
 
 
 
 
 
 
 
 
 
c71b83b
 
 
 
 
 
 
 
 
 
 
 
249a7a1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c71b83b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
249a7a1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c71b83b
249a7a1
 
c71b83b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
249a7a1
c71b83b
249a7a1
c71b83b
 
 
 
249a7a1
c71b83b
 
 
 
 
 
 
 
 
 
249a7a1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c71b83b
 
249a7a1
 
 
 
 
 
 
 
c71b83b
 
 
249a7a1
c71b83b
249a7a1
 
 
 
 
c71b83b
 
 
249a7a1
 
c71b83b
 
249a7a1
 
 
 
 
 
 
 
 
 
c71b83b
 
249a7a1
c71b83b
 
 
249a7a1
c71b83b
 
249a7a1
c71b83b
 
 
249a7a1
c71b83b
 
 
249a7a1
c71b83b
 
249a7a1
c71b83b
 
 
 
249a7a1
c71b83b
 
 
 
 
 
 
 
249a7a1
 
c71b83b
 
249a7a1
c71b83b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
249a7a1
c71b83b
 
249a7a1
c71b83b
 
 
54787fd
c71b83b
 
 
 
 
 
 
 
 
 
 
 
 
 
54787fd
c71b83b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
249a7a1
c71b83b
 
249a7a1
c71b83b
 
 
 
 
 
 
 
249a7a1
c71b83b
249a7a1
c71b83b
 
 
 
249a7a1
c71b83b
 
 
249a7a1
c71b83b
249a7a1
c71b83b
 
 
 
 
 
 
 
 
 
 
249a7a1
 
 
 
 
 
 
c71b83b
249a7a1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c71b83b
249a7a1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c71b83b
249a7a1
 
 
 
 
 
 
 
c71b83b
 
 
8f22003
c71b83b
249a7a1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c71b83b
249a7a1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
import inspect
import warnings
from itertools import repeat
from typing import Callable, List, Optional, Union

import torch
from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer

from diffusers.image_processor import VaeImageProcessor
from diffusers.models import AutoencoderKL, UNet2DConditionModel
from diffusers.models.attention_processor import AttnProcessor, Attention
from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
from diffusers.schedulers import DDIMScheduler
from scheduling_dpmsolver_multistep_inject import DPMSolverMultistepSchedulerInject
# from diffusers.utils import logging, randn_tensor
from diffusers.utils import logging
from diffusers.utils.torch_utils import randn_tensor
from diffusers.pipelines.pipeline_utils import DiffusionPipeline
from diffusers.pipelines.semantic_stable_diffusion import SemanticStableDiffusionPipelineOutput

import numpy as np
from PIL import Image
from tqdm import tqdm
import torch.nn.functional as F
import math
from collections.abc import Iterable

logger = logging.get_logger(__name__)  # pylint: disable=invalid-name


class AttentionStore():
    @staticmethod
    def get_empty_store():
        return {"down_cross": [], "mid_cross": [], "up_cross": [],
                "down_self": [], "mid_self": [], "up_self": []}

    def __call__(self, attn, is_cross: bool, place_in_unet: str, editing_prompts, PnP=False):
        # attn.shape = batch_size * head_size, seq_len query, seq_len_key
        bs = 2 + int(PnP) + editing_prompts
        skip = 2 if PnP else 1  # skip PnP & unconditional

        head_size = int(attn.shape[0] / self.batch_size)
        attn = torch.stack(attn.split(self.batch_size)).permute(1, 0, 2, 3)
        source_batch_size = int(attn.shape[1] // bs)
        self.forward(
            attn[:, skip * source_batch_size:],
            is_cross,
            place_in_unet)

    def forward(self, attn, is_cross: bool, place_in_unet: str):
        key = f"{place_in_unet}_{'cross' if is_cross else 'self'}"
        if attn.shape[1] <= 32 ** 2:  # avoid memory overhead
            self.step_store[key].append(attn)

    def between_steps(self, store_step=True):
        if store_step:
            if self.average:
                if len(self.attention_store) == 0:
                    self.attention_store = self.step_store
                else:
                    for key in self.attention_store:
                        for i in range(len(self.attention_store[key])):
                            self.attention_store[key][i] += self.step_store[key][i]
            else:
                if len(self.attention_store) == 0:
                    self.attention_store = [self.step_store]
                else:
                    self.attention_store.append(self.step_store)

            self.cur_step += 1
        self.step_store = self.get_empty_store()

    def get_attention(self, step: int):
        if self.average:
            attention = {key: [item / self.cur_step for item in self.attention_store[key]] for key in
                         self.attention_store}
        else:
            assert (step is not None)
            attention = self.attention_store[step]
        return attention

    def aggregate_attention(self, attention_maps, prompts, res: int,
                            from_where: List[str], is_cross: bool, select: int
                            ):
        out = [[] for x in range(self.batch_size)]
        num_pixels = res ** 2
        for location in from_where:
            for bs_item in attention_maps[f"{location}_{'cross' if is_cross else 'self'}"]:
                for batch, item in enumerate(bs_item):
                    if item.shape[1] == num_pixels:
                        cross_maps = item.reshape(len(prompts), -1, res, res, item.shape[-1])[select]
                        out[batch].append(cross_maps)

        out = torch.stack([torch.cat(x, dim=0) for x in out])
        # average over heads
        out = out.sum(1) / out.shape[1]
        return out

    def __init__(self, average: bool, batch_size=1):
        self.step_store = self.get_empty_store()
        self.attention_store = []
        self.cur_step = 0
        self.average = average
        self.batch_size = batch_size


class CrossAttnProcessor:

    def __init__(self, attention_store, place_in_unet, PnP, editing_prompts):
        self.attnstore = attention_store
        self.place_in_unet = place_in_unet
        self.editing_prompts = editing_prompts
        self.PnP = PnP

    def __call__(
            self,
            attn: Attention,
            hidden_states,
            encoder_hidden_states=None,
            attention_mask=None,
            temb=None,
    ):
        assert (not attn.residual_connection)
        assert (attn.spatial_norm is None)
        assert (attn.group_norm is None)
        assert (hidden_states.ndim != 4)
        assert (encoder_hidden_states is not None)  # is cross

        batch_size, sequence_length, _ = (
            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
        )
        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)

        query = attn.to_q(hidden_states)

        if encoder_hidden_states is None:
            encoder_hidden_states = hidden_states
        elif attn.norm_cross:
            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)

        key = attn.to_k(encoder_hidden_states)
        value = attn.to_v(encoder_hidden_states)

        query = attn.head_to_batch_dim(query)
        key = attn.head_to_batch_dim(key)
        value = attn.head_to_batch_dim(value)

        attention_probs = attn.get_attention_scores(query, key, attention_mask)
        self.attnstore(attention_probs,
                       is_cross=True,
                       place_in_unet=self.place_in_unet,
                       editing_prompts=self.editing_prompts,
                       PnP=self.PnP)

        hidden_states = torch.bmm(attention_probs, value)
        hidden_states = attn.batch_to_head_dim(hidden_states)

        # linear proj
        hidden_states = attn.to_out[0](hidden_states)
        # dropout
        hidden_states = attn.to_out[1](hidden_states)

        hidden_states = hidden_states / attn.rescale_output_factor
        return hidden_states


# Modified from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionAttendAndExcitePipeline.GaussianSmoothing
class GaussianSmoothing():

    def __init__(self, device):
        kernel_size = [3, 3]
        sigma = [0.5, 0.5]

        # The gaussian kernel is the product of the gaussian function of each dimension.
        kernel = 1
        meshgrids = torch.meshgrid([torch.arange(size, dtype=torch.float32) for size in kernel_size])
        for size, std, mgrid in zip(kernel_size, sigma, meshgrids):
            mean = (size - 1) / 2
            kernel *= 1 / (std * math.sqrt(2 * math.pi)) * torch.exp(-(((mgrid - mean) / (2 * std)) ** 2))

        # Make sure sum of values in gaussian kernel equals 1.
        kernel = kernel / torch.sum(kernel)

        # Reshape to depthwise convolutional weight
        kernel = kernel.view(1, 1, *kernel.size())
        kernel = kernel.repeat(1, *[1] * (kernel.dim() - 1))

        self.weight = kernel.to(device)

    def __call__(self, input):
        """
        Arguments:
        Apply gaussian filter to input.
            input (torch.Tensor): Input to apply gaussian filter on.
        Returns:
            filtered (torch.Tensor): Filtered output.
        """
        return F.conv2d(input, weight=self.weight.to(input.dtype))


def load_512(image_path, size, left=0, right=0, top=0, bottom=0, device=None, dtype=None):
    def pre_process(im, size, left=0, right=0, top=0, bottom=0):
        if type(im) is str:
            image = np.array(Image.open(im).convert('RGB'))[:, :, :3]
        elif isinstance(im, Image.Image):
            image = np.array((im).convert('RGB'))[:, :, :3]
        else:
            image = im
        h, w, c = image.shape
        left = min(left, w - 1)
        right = min(right, w - left - 1)
        top = min(top, h - left - 1)
        bottom = min(bottom, h - top - 1)
        image = image[top:h - bottom, left:w - right]
        h, w, c = image.shape
        if h < w:
            offset = (w - h) // 2
            image = image[:, offset:offset + h]
        elif w < h:
            offset = (h - w) // 2
            image = image[offset:offset + w]
        image = np.array(Image.fromarray(image).resize((size, size)))
        image = torch.from_numpy(image).float().permute(2, 0, 1)
        return image

    tmps = []
    if isinstance(image_path, list):
        for item in image_path:
            tmps.append(pre_process(item, size, left, right, top, bottom))
    else:
        tmps.append(pre_process(image_path, size, left, right, top, bottom))
    image = torch.stack(tmps) / 127.5 - 1

    image = image.to(device=device, dtype=dtype)
    return image


# Modified from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionAttendAndExcitePipeline.GaussianSmoothing

def reset_dpm(scheduler):
    if isinstance(scheduler, DPMSolverMultistepSchedulerInject):
        scheduler.model_outputs = [
                                           None,
                                       ] * scheduler.config.solver_order
        scheduler.lower_order_nums = 0


class SemanticStableDiffusionImg2ImgPipeline_DPMSolver(DiffusionPipeline):
    r"""
    Pipeline for text-to-image generation with latent editing.

    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)

    This model builds on the implementation of ['StableDiffusionPipeline']

    Args:
        vae ([`AutoencoderKL`]):
            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
        text_encoder ([`CLIPTextModel`]):
            Frozen text-encoder. Stable Diffusion uses the text portion of
            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
        tokenizer (`CLIPTokenizer`):
            Tokenizer of class
            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
        scheduler ([`SchedulerMixin`]):
            A scheduler to be used in combination with `unet` to denoise the encoded image latens. Can be one of
            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
        safety_checker ([`Q16SafetyChecker`]):
            Classification module that estimates whether generated images could be considered offensive or harmful.
            Please, refer to the [model card](https://huggingface.co/CompVis/stable-diffusion-v1-4) for details.
        feature_extractor ([`CLIPImageProcessor`]):
            Model that extracts features from generated images to be used as inputs for the `safety_checker`.
    """

    _optional_components = ["safety_checker", "feature_extractor"]

    def __init__(
            self,
            vae: AutoencoderKL,
            text_encoder: CLIPTextModel,
            tokenizer: CLIPTokenizer,
            unet: UNet2DConditionModel,
            scheduler: Union[DDIMScheduler,DPMSolverMultistepSchedulerInject],
            safety_checker: StableDiffusionSafetyChecker,
            feature_extractor: CLIPImageProcessor,
            requires_safety_checker: bool = True,
    ):
        super().__init__()

        if not isinstance(scheduler, DDIMScheduler) or not isinstance(scheduler, DPMSolverMultistepSchedulerInject):
            scheduler = DPMSolverMultistepSchedulerInject.from_config(scheduler.config, algorithm_type="sde-dpmsolver++", solver_order=2)
            logger.warning("This pipeline only supports DDIMScheduler and DPMSolverMultistepSchedulerInject. "
                           "The scheduler has been changed to DPMSolverMultistepSchedulerInject.")

        if safety_checker is None and requires_safety_checker:
            logger.warning(
                f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
                " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
                " results in services or applications open to the public. Both the diffusers team and Hugging Face"
                " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
                " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
                " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
            )

        if safety_checker is not None and feature_extractor is None:
            raise ValueError(
                "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
                " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
            )

        self.register_modules(
            vae=vae,
            text_encoder=text_encoder,
            tokenizer=tokenizer,
            unet=unet,
            scheduler=scheduler,
            safety_checker=safety_checker,
            feature_extractor=feature_extractor,
        )
        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
        self.register_to_config(requires_safety_checker=requires_safety_checker)

    def progress_bar(self, iterable=None, total=None, verbose=True):
        if not hasattr(self, "_progress_bar_config"):
            self._progress_bar_config = {}
        elif not isinstance(self._progress_bar_config, dict):
            raise ValueError(
                f"`self._progress_bar_config` should be of type `dict`, but is {type(self._progress_bar_config)}."
            )
        if not verbose:
            return iterable
        elif iterable is not None:
            return tqdm(iterable, **self._progress_bar_config)
        elif total is not None:
            return tqdm(total=total, **self._progress_bar_config)
        else:
            raise ValueError("Either `total` or `iterable` has to be defined.")

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
    def run_safety_checker(self, image, device, dtype):
        if self.safety_checker is None:
            has_nsfw_concept = None
        else:
            if torch.is_tensor(image):
                feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
            else:
                feature_extractor_input = self.image_processor.numpy_to_pil(image)
            safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
            image, has_nsfw_concept = self.safety_checker(
                images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
            )
        return image, has_nsfw_concept

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
    def decode_latents(self, latents):
        warnings.warn(
            "The decode_latents method is deprecated and will be removed in a future version. Please"
            " use VaeImageProcessor instead",
            FutureWarning,
        )
        latents = 1 / self.vae.config.scaling_factor * latents
        image = self.vae.decode(latents, return_dict=False)[0]
        image = (image / 2 + 0.5).clamp(0, 1)
        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
        return image

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
    def prepare_extra_step_kwargs(self, eta):
        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
        # and should be between [0, 1]

        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
        extra_step_kwargs = {}
        if accepts_eta:
            extra_step_kwargs["eta"] = eta

        return extra_step_kwargs

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.check_inputs
    def check_inputs(
            self,
            prompt,
            height,
            width,
            callback_steps,
            negative_prompt=None,
            prompt_embeds=None,
            negative_prompt_embeds=None,
    ):
        if height % 8 != 0 or width % 8 != 0:
            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")

        if (callback_steps is None) or (
                callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
        ):
            raise ValueError(
                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
                f" {type(callback_steps)}."
            )

        if prompt is not None and prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
                " only forward one of the two."
            )
        elif prompt is None and prompt_embeds is None:
            raise ValueError(
                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
            )
        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")

        if negative_prompt is not None and negative_prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
            )

        if prompt_embeds is not None and negative_prompt_embeds is not None:
            if prompt_embeds.shape != negative_prompt_embeds.shape:
                raise ValueError(
                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
                    f" {negative_prompt_embeds.shape}."
                )

    # Modified from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
    def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, latents):
        shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)

        if latents.shape != shape:
            raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")

        latents = latents.to(device)

        # scale the initial noise by the standard deviation required by the scheduler
        latents = latents * self.scheduler.init_noise_sigma
        return latents

    def prepare_unet(self, attention_store, PnP: bool = False):
        attn_procs = {}
        for name in self.unet.attn_processors.keys():
            if name.startswith("mid_block"):
                place_in_unet = "mid"
            elif name.startswith("up_blocks"):
                place_in_unet = "up"
            elif name.startswith("down_blocks"):
                place_in_unet = "down"
            else:
                continue

            if "attn2" in name:
                attn_procs[name] = CrossAttnProcessor(
                    attention_store=attention_store,
                    place_in_unet=place_in_unet,
                    PnP=PnP,
                    editing_prompts=self.enabled_editing_prompts)
            else:
                attn_procs[name] = AttnProcessor()

        self.unet.set_attn_processor(attn_procs)

    @torch.no_grad()
    def __call__(
            self,
            prompt: Union[str, List[str]] = "",
            height: Optional[int] = None,
            width: Optional[int] = None,
            # num_inference_steps: int = 50,
            guidance_scale: float = 7.5,
            negative_prompt: Optional[Union[str, List[str]]] = None,
            # num_images_per_prompt: int = 1,
            eta: float = 1.0,
            # generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
            # latents: Optional[torch.FloatTensor] = None,
            output_type: Optional[str] = "pil",
            return_dict: bool = True,
            callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
            callback_steps: int = 1,
            editing_prompt: Optional[Union[str, List[str]]] = None,
            editing_prompt_embeddings: Optional[torch.Tensor] = None,
            reverse_editing_direction: Optional[Union[bool, List[bool]]] = False,
            edit_guidance_scale: Optional[Union[float, List[float]]] = 5,
            edit_warmup_steps: Optional[Union[int, List[int]]] = 10,
            edit_cooldown_steps: Optional[Union[int, List[int]]] = None,
            edit_threshold: Optional[Union[float, List[float]]] = 0.9,
            user_mask: Optional[torch.FloatTensor] = None,
            edit_momentum_scale: Optional[float] = 0.1,
            edit_mom_beta: Optional[float] = 0.4,
            edit_weights: Optional[List[float]] = None,
            sem_guidance: Optional[List[torch.Tensor]] = None,
            verbose=True,
            use_cross_attn_mask: bool = False,
            # Attention store (just for visualization purposes)
            attn_store_steps: Optional[List[int]] = [],
            store_averaged_over_steps: bool = True,
            use_intersect_mask: bool = False,
            init_latents = None,
            zs = None,
            
    ):
        r"""
        Function invoked when calling the pipeline for generation.

        Args:
            prompt (`str` or `List[str]`):
                The prompt or prompts to guide the image generation.
            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                The height in pixels of the generated image.
            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                The width in pixels of the generated image.
            num_inference_steps (`int`, *optional*, defaults to 50):
                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                expense of slower inference.
            guidance_scale (`float`, *optional*, defaults to 7.5):
                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
                `guidance_scale` is defined as `w` of equation 2. of [Imagen
                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
                usually at the expense of lower image quality.
            negative_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
                if `guidance_scale` is less than `1`).
            num_images_per_prompt (`int`, *optional*, defaults to 1):
                The number of images to generate per prompt.
            eta (`float`, *optional*, defaults to 0.0):
                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
                [`schedulers.DDIMScheduler`], will be ignored for others.
            generator (`torch.Generator`, *optional*):
                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
                to make generation deterministic.
            latents (`torch.FloatTensor`, *optional*):
                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
                tensor will ge generated by sampling using the supplied random `generator`.
            output_type (`str`, *optional*, defaults to `"pil"`):
                The output format of the generate image. Choose between
                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
                plain tuple.
            callback (`Callable`, *optional*):
                A function that will be called every `callback_steps` steps during inference. The function will be
                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
            callback_steps (`int`, *optional*, defaults to 1):
                The frequency at which the `callback` function will be called. If not specified, the callback will be
                called at every step.
            editing_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts to use for Semantic guidance. Semantic guidance is disabled by setting
                `editing_prompt = None`. Guidance direction of prompt should be specified via
                `reverse_editing_direction`.
            editing_prompt_embeddings (`torch.Tensor>`, *optional*):
                Pre-computed embeddings to use for semantic guidance. Guidance direction of embedding should be
                specified via `reverse_editing_direction`.
            reverse_editing_direction (`bool` or `List[bool]`, *optional*, defaults to `False`):
                Whether the corresponding prompt in `editing_prompt` should be increased or decreased.
            edit_guidance_scale (`float` or `List[float]`, *optional*, defaults to 5):
                Guidance scale for semantic guidance. If provided as list values should correspond to `editing_prompt`.
                `edit_guidance_scale` is defined as `s_e` of equation 6 of [SEGA
                Paper](https://arxiv.org/pdf/2301.12247.pdf).
            edit_warmup_steps (`float` or `List[float]`, *optional*, defaults to 10):
                Number of diffusion steps (for each prompt) for which semantic guidance will not be applied. Momentum
                will still be calculated for those steps and applied once all warmup periods are over.
                `edit_warmup_steps` is defined as `delta` (δ) of [SEGA Paper](https://arxiv.org/pdf/2301.12247.pdf).
            edit_cooldown_steps (`float` or `List[float]`, *optional*, defaults to `None`):
                Number of diffusion steps (for each prompt) after which semantic guidance will no longer be applied.
            edit_threshold (`float` or `List[float]`, *optional*, defaults to 0.9):
                Threshold of semantic guidance.
            edit_momentum_scale (`float`, *optional*, defaults to 0.1):
                Scale of the momentum to be added to the semantic guidance at each diffusion step. If set to 0.0
                momentum will be disabled. Momentum is already built up during warmup, i.e. for diffusion steps smaller
                than `sld_warmup_steps`. Momentum will only be added to latent guidance once all warmup periods are
                finished. `edit_momentum_scale` is defined as `s_m` of equation 7 of [SEGA
                Paper](https://arxiv.org/pdf/2301.12247.pdf).
            edit_mom_beta (`float`, *optional*, defaults to 0.4):
                Defines how semantic guidance momentum builds up. `edit_mom_beta` indicates how much of the previous
                momentum will be kept. Momentum is already built up during warmup, i.e. for diffusion steps smaller
                than `edit_warmup_steps`. `edit_mom_beta` is defined as `beta_m` (β) of equation 8 of [SEGA
                Paper](https://arxiv.org/pdf/2301.12247.pdf).
            edit_weights (`List[float]`, *optional*, defaults to `None`):
                Indicates how much each individual concept should influence the overall guidance. If no weights are
                provided all concepts are applied equally. `edit_mom_beta` is defined as `g_i` of equation 9 of [SEGA
                Paper](https://arxiv.org/pdf/2301.12247.pdf).
            sem_guidance (`List[torch.Tensor]`, *optional*):
                List of pre-generated guidance vectors to be applied at generation. Length of the list has to
                correspond to `num_inference_steps`.

        Returns:
            [`~pipelines.semantic_stable_diffusion.SemanticStableDiffusionPipelineOutput`] or `tuple`:
            [`~pipelines.semantic_stable_diffusion.SemanticStableDiffusionPipelineOutput`] if `return_dict` is True,
            otherwise a `tuple. When returning a tuple, the first element is a list with the generated images, and the
            second element is a list of `bool`s denoting whether the corresponding generated image likely represents
            "not-safe-for-work" (nsfw) content, according to the `safety_checker`.
        """
        # eta = self.eta
        num_images_per_prompt = 1
        # latents = self.init_latents
        latents = init_latents

        use_ddpm = True
        # zs = self.zs
        reset_dpm(self.scheduler)

        if use_intersect_mask:
            use_cross_attn_mask = True

        if use_cross_attn_mask:
            self.smoothing = GaussianSmoothing(self.device)

        # 0. Default height and width to unet
        height = height or self.unet.config.sample_size * self.vae_scale_factor
        width = width or self.unet.config.sample_size * self.vae_scale_factor

        # 1. Check inputs. Raise error if not correct
        self.check_inputs(prompt, height, width, callback_steps)

        org_prompt = prompt
        if isinstance(prompt, list):
            assert len(prompt) == self.batch_size
        elif isinstance(prompt, str):
            prompt = list(repeat(prompt, self.batch_size))

        # 2. Define call parameters
        batch_size = self.batch_size

        if editing_prompt:
            enable_edit_guidance = True
            if isinstance(editing_prompt, str):
                editing_prompt = [editing_prompt]
            self.enabled_editing_prompts = len(editing_prompt)
        elif editing_prompt_embeddings is not None:
            enable_edit_guidance = True
            self.enabled_editing_prompts = editing_prompt_embeddings.shape[0]
        else:
            self.enabled_editing_prompts = 0
            enable_edit_guidance = False

        # get prompt text embeddings
        text_inputs = self.tokenizer(
            prompt,
            padding="max_length",
            max_length=self.tokenizer.model_max_length,
            truncation=True,
            return_tensors="pt",
        )
        text_input_ids = text_inputs.input_ids
        untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids

        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
                text_input_ids, untruncated_ids
        ):
            removed_text = self.tokenizer.batch_decode(
                untruncated_ids[:, self.tokenizer.model_max_length - 1: -1]
            )
            logger.warning(
                "The following part of your input was truncated because CLIP can only handle sequences up to"
                f" {self.tokenizer.model_max_length} tokens: {removed_text}"
            )

        text_embeddings = self.text_encoder(text_input_ids.to(self.device))[0]

        # duplicate text embeddings for each generation per prompt, using mps friendly method
        bs_embed, seq_len, _ = text_embeddings.shape
        text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1)
        text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)

        if enable_edit_guidance:
            # get safety text embeddings
            if editing_prompt_embeddings is None:
                edit_concepts_input = self.tokenizer(
                    [x for item in editing_prompt for x in repeat(item, batch_size)],
                    padding="max_length",
                    max_length=self.tokenizer.model_max_length,
                    truncation=True,
                    return_tensors="pt",
                    return_length=True
                )

                num_edit_tokens = edit_concepts_input.length - 2  # not counting startoftext and endoftext
                edit_concepts_input_ids = edit_concepts_input.input_ids
                untruncated_ids = self.tokenizer(
                    [x for item in editing_prompt for x in repeat(item, batch_size)],
                    padding="longest",
                    return_tensors="pt").input_ids

                if untruncated_ids.shape[-1] >= edit_concepts_input_ids.shape[-1] and not torch.equal(
                        edit_concepts_input_ids, untruncated_ids
                ):
                    removed_text = self.tokenizer.batch_decode(
                        untruncated_ids[:, self.tokenizer.model_max_length - 1: -1]
                    )
                    logger.warning(
                        "The following part of your input was truncated because CLIP can only handle sequences up to"
                        f" {self.tokenizer.model_max_length} tokens: {removed_text}"
                    )

                edit_concepts = self.text_encoder(edit_concepts_input_ids.to(self.device))[0]
            else:
                edit_concepts = editing_prompt_embeddings.to(self.device).repeat(batch_size, 1, 1)

            # duplicate text embeddings for each generation per prompt, using mps friendly method
            bs_embed_edit, seq_len_edit, _ = edit_concepts.shape
            edit_concepts = edit_concepts.repeat(1, num_images_per_prompt, 1)
            edit_concepts = edit_concepts.view(bs_embed_edit * num_images_per_prompt, seq_len_edit, -1)

        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
        # corresponds to doing no classifier free guidance.
        do_classifier_free_guidance = guidance_scale > 1.0
        # get unconditional embeddings for classifier free guidance

        if do_classifier_free_guidance:
            uncond_tokens: List[str]
            if negative_prompt is None:
                uncond_tokens = [""]
            elif type(prompt) is not type(negative_prompt):
                raise TypeError(
                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
                    f" {type(prompt)}."
                )
            elif isinstance(negative_prompt, str):
                uncond_tokens = [negative_prompt]
            elif batch_size != len(negative_prompt):
                raise ValueError(
                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
                    " the batch size of `prompt`."
                )
            else:
                uncond_tokens = negative_prompt

            max_length = text_input_ids.shape[-1]
            uncond_input = self.tokenizer(
                uncond_tokens,
                padding="max_length",
                max_length=max_length,
                truncation=True,
                return_tensors="pt",
            )
            uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]

            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
            seq_len = uncond_embeddings.shape[1]
            uncond_embeddings = uncond_embeddings.repeat(batch_size, num_images_per_prompt, 1)
            uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1)

            # For classifier free guidance, we need to do two forward passes.
            # Here we concatenate the unconditional and text embeddings into a single batch
            # to avoid doing two forward passes
            self.text_cross_attention_maps = [org_prompt] if isinstance(org_prompt, str) else org_prompt
            if enable_edit_guidance:
                text_embeddings = torch.cat([uncond_embeddings, text_embeddings, edit_concepts])
                self.text_cross_attention_maps += \
                    ([editing_prompt] if isinstance(editing_prompt, str) else editing_prompt)
            else:
                text_embeddings = torch.cat([uncond_embeddings, text_embeddings])

        # 4. Prepare timesteps
        #self.scheduler.set_timesteps(num_inference_steps, device=self.device)
        timesteps = self.scheduler.timesteps

        if use_ddpm:
            t_to_idx = {int(v): k for k, v in enumerate(timesteps[-zs.shape[0]:])}
            timesteps = timesteps[-zs.shape[0]:]

        if use_cross_attn_mask:
            self.attention_store = AttentionStore(average=store_averaged_over_steps, batch_size=batch_size)
            self.prepare_unet(self.attention_store, PnP=False)
        # 5. Prepare latent variables
        num_channels_latents = self.unet.config.in_channels
        latents = self.prepare_latents(
            batch_size * num_images_per_prompt,
            num_channels_latents,
            height,
            width,
            text_embeddings.dtype,
            self.device,
            latents,
        )

        # 6. Prepare extra step kwargs.
        extra_step_kwargs = self.prepare_extra_step_kwargs(eta)

        # Initialize edit_momentum to None
        edit_momentum = None

        self.uncond_estimates = None
        self.text_estimates = None
        self.edit_estimates = None
        self.sem_guidance = None
        self.activation_mask = None

        for i, t in enumerate(self.progress_bar(timesteps, verbose=verbose)):
            idx = t_to_idx[int(t)]


            # expand the latents if we are doing classifier free guidance

            if do_classifier_free_guidance:
                latent_model_input = torch.cat([latents] * (2 + self.enabled_editing_prompts))
            else:
                latent_model_input = latents

            latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)

            text_embed_input = text_embeddings

            # predict the noise residual
            noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embed_input).sample

            # perform guidance
            if do_classifier_free_guidance:

                noise_pred_out = noise_pred.chunk(2 + self.enabled_editing_prompts)  # [b,4, 64, 64]
                noise_pred_uncond, noise_pred_text = noise_pred_out[0], noise_pred_out[1]
                noise_pred_edit_concepts = noise_pred_out[2:]

                # default text guidance
                noise_guidance = guidance_scale * (noise_pred_text - noise_pred_uncond)

                if self.uncond_estimates is None:
                    self.uncond_estimates = torch.zeros((len(timesteps), *noise_pred_uncond.shape))
                self.uncond_estimates[i] = noise_pred_uncond.detach().cpu()

                if self.text_estimates is None:
                    self.text_estimates = torch.zeros((len(timesteps), *noise_pred_text.shape))
                self.text_estimates[i] = noise_pred_text.detach().cpu()

                if edit_momentum is None:
                    edit_momentum = torch.zeros_like(noise_guidance)

                if sem_guidance is not None and len(sem_guidance) > i:
                    edit_guidance = sem_guidance[i].to(self.device)
                    edit_momentum = edit_mom_beta * edit_momentum + (1 - edit_mom_beta) * edit_guidance
                    noise_guidance = noise_guidance + edit_guidance

                elif enable_edit_guidance:
                    if self.activation_mask is None:
                        self.activation_mask = torch.zeros(
                            (len(timesteps), len(noise_pred_edit_concepts), *noise_pred_edit_concepts[0].shape)
                        )
                    if self.edit_estimates is None and enable_edit_guidance:
                        self.edit_estimates = torch.zeros(
                            (len(timesteps), len(noise_pred_edit_concepts), *noise_pred_edit_concepts[0].shape)
                        )

                    if self.sem_guidance is None:
                        self.sem_guidance = torch.zeros((len(timesteps), *noise_pred_text.shape))

                    concept_weights = torch.zeros(
                        (len(noise_pred_edit_concepts), noise_guidance.shape[0]),
                        device=self.device,
                        dtype=noise_guidance.dtype,
                    )
                    noise_guidance_edit = torch.zeros(
                        (len(noise_pred_edit_concepts), *noise_guidance.shape),
                        device=self.device,
                        dtype=noise_guidance.dtype,
                    )
                    # noise_guidance_edit = torch.zeros_like(noise_guidance)
                    warmup_inds = []
                    for c, noise_pred_edit_concept in enumerate(noise_pred_edit_concepts):
                        self.edit_estimates[i, c] = noise_pred_edit_concept
                        if isinstance(edit_guidance_scale, list):
                            edit_guidance_scale_c = edit_guidance_scale[c]
                        else:
                            edit_guidance_scale_c = edit_guidance_scale

                        if isinstance(edit_threshold, list):
                            edit_threshold_c = edit_threshold[c]
                        else:
                            edit_threshold_c = edit_threshold
                        if isinstance(reverse_editing_direction, list):
                            reverse_editing_direction_c = reverse_editing_direction[c]
                        else:
                            reverse_editing_direction_c = reverse_editing_direction
                        if edit_weights:
                            edit_weight_c = edit_weights[c]
                        else:
                            edit_weight_c = 1.0
                        if isinstance(edit_warmup_steps, list):
                            edit_warmup_steps_c = edit_warmup_steps[c]
                        else:
                            edit_warmup_steps_c = edit_warmup_steps

                        if isinstance(edit_cooldown_steps, list):
                            edit_cooldown_steps_c = edit_cooldown_steps[c]
                        elif edit_cooldown_steps is None:
                            edit_cooldown_steps_c = i + 1
                        else:
                            edit_cooldown_steps_c = edit_cooldown_steps
                        if i >= edit_warmup_steps_c:
                            warmup_inds.append(c)
                        if i >= edit_cooldown_steps_c:
                            noise_guidance_edit[c, :, :, :, :] = torch.zeros_like(noise_pred_edit_concept)
                            continue

                        noise_guidance_edit_tmp = noise_pred_edit_concept - noise_pred_uncond
                        # tmp_weights = (noise_pred_text - noise_pred_edit_concept).sum(dim=(1, 2, 3))
                        tmp_weights = (noise_guidance - noise_pred_edit_concept).sum(dim=(1, 2, 3))

                        tmp_weights = torch.full_like(tmp_weights, edit_weight_c)  # * (1 / enabled_editing_prompts)
                        if reverse_editing_direction_c:
                            noise_guidance_edit_tmp = noise_guidance_edit_tmp * -1
                        concept_weights[c, :] = tmp_weights

                        noise_guidance_edit_tmp = noise_guidance_edit_tmp * edit_guidance_scale_c

                        if user_mask is not None:
                            noise_guidance_edit_tmp = noise_guidance_edit_tmp * user_mask

                        if use_cross_attn_mask:
                            out = self.attention_store.aggregate_attention(
                                attention_maps=self.attention_store.step_store,
                                prompts=self.text_cross_attention_maps,
                                res=16,
                                from_where=["up", "down"],
                                is_cross=True,
                                select=self.text_cross_attention_maps.index(editing_prompt[c]),
                            )
                            attn_map = out[:, :, :, 1:1 + num_edit_tokens[c]]  # 0 -> startoftext

                            # average over all tokens
                            assert (attn_map.shape[3] == num_edit_tokens[c])
                            attn_map = torch.sum(attn_map, dim=3)

                            # gaussian_smoothing
                            attn_map = F.pad(attn_map.unsqueeze(1), (1, 1, 1, 1), mode="reflect")
                            attn_map = self.smoothing(attn_map).squeeze(1)

                            # create binary mask
                            if attn_map.dtype == torch.float32:
                                tmp = torch.quantile(attn_map.flatten(start_dim=1), edit_threshold_c, dim=1)
                            else:
                                tmp = torch.quantile(attn_map.flatten(start_dim=1).to(torch.float32), edit_threshold_c, dim=1).to(attn_map.dtype)
                            attn_mask = torch.where(attn_map >= tmp.unsqueeze(1).unsqueeze(1).repeat(1,16,16), 1.0, 0.0)

                            # resolution must match latent space dimension
                            attn_mask = F.interpolate(
                                attn_mask.unsqueeze(1),
                                noise_guidance_edit_tmp.shape[-2:]  # 64,64
                            ).repeat(1, 4, 1, 1)
                            self.activation_mask[i, c] = attn_mask.detach().cpu()
                            if not use_intersect_mask:
                                noise_guidance_edit_tmp = noise_guidance_edit_tmp * attn_mask

                        if use_intersect_mask:
                            noise_guidance_edit_tmp_quantile = torch.abs(noise_guidance_edit_tmp)
                            noise_guidance_edit_tmp_quantile = torch.sum(noise_guidance_edit_tmp_quantile, dim=1,
                                                                         keepdim=True)
                            noise_guidance_edit_tmp_quantile = noise_guidance_edit_tmp_quantile.repeat(1, 4, 1, 1)

                            # torch.quantile function expects float32
                            if noise_guidance_edit_tmp_quantile.dtype == torch.float32:
                                tmp = torch.quantile(
                                    noise_guidance_edit_tmp_quantile.flatten(start_dim=2),
                                    edit_threshold_c,
                                    dim=2,
                                    keepdim=False,
                                )
                            else:
                                tmp = torch.quantile(
                                    noise_guidance_edit_tmp_quantile.flatten(start_dim=2).to(torch.float32),
                                    edit_threshold_c,
                                    dim=2,
                                    keepdim=False,
                                ).to(noise_guidance_edit_tmp_quantile.dtype)

                            intersect_mask = torch.where(
                                noise_guidance_edit_tmp_quantile >= tmp[:, :, None, None],
                                torch.ones_like(noise_guidance_edit_tmp),
                                torch.zeros_like(noise_guidance_edit_tmp),
                            ) * attn_mask

                            self.activation_mask[i, c] = intersect_mask.detach().cpu()

                            noise_guidance_edit_tmp = noise_guidance_edit_tmp * intersect_mask

                        elif not use_cross_attn_mask:
                            # calculate quantile
                            noise_guidance_edit_tmp_quantile = torch.abs(noise_guidance_edit_tmp)
                            noise_guidance_edit_tmp_quantile = torch.sum(noise_guidance_edit_tmp_quantile, dim=1,
                                                                         keepdim=True)
                            noise_guidance_edit_tmp_quantile = noise_guidance_edit_tmp_quantile.repeat(1, 4, 1, 1)

                            # torch.quantile function expects float32
                            if noise_guidance_edit_tmp_quantile.dtype == torch.float32:
                                tmp = torch.quantile(
                                    noise_guidance_edit_tmp_quantile.flatten(start_dim=2),
                                    edit_threshold_c,
                                    dim=2,
                                    keepdim=False,
                                )
                            else:
                                tmp = torch.quantile(
                                    noise_guidance_edit_tmp_quantile.flatten(start_dim=2).to(torch.float32),
                                    edit_threshold_c,
                                    dim=2,
                                    keepdim=False,
                                ).to(noise_guidance_edit_tmp_quantile.dtype)

                            self.activation_mask[i, c] = torch.where(
                                noise_guidance_edit_tmp_quantile >= tmp[:, :, None, None],
                                torch.ones_like(noise_guidance_edit_tmp),
                                torch.zeros_like(noise_guidance_edit_tmp),
                            ).detach().cpu()

                            noise_guidance_edit_tmp = torch.where(
                                noise_guidance_edit_tmp_quantile >= tmp[:, :, None, None],
                                noise_guidance_edit_tmp,
                                torch.zeros_like(noise_guidance_edit_tmp),
                            )

                        noise_guidance_edit[c, :, :, :, :] = noise_guidance_edit_tmp

                    warmup_inds = torch.tensor(warmup_inds).to(self.device)
                    if len(noise_pred_edit_concepts) > warmup_inds.shape[0] > 0:
                        concept_weights = concept_weights.to("cpu")  # Offload to cpu
                        noise_guidance_edit = noise_guidance_edit.to("cpu")

                        concept_weights_tmp = torch.index_select(concept_weights.to(self.device), 0, warmup_inds)
                        concept_weights_tmp = torch.where(
                            concept_weights_tmp < 0, torch.zeros_like(concept_weights_tmp), concept_weights_tmp
                        )
                        concept_weights_tmp = concept_weights_tmp / concept_weights_tmp.sum(dim=0)
                        # concept_weights_tmp = torch.nan_to_num(concept_weights_tmp)

                        noise_guidance_edit_tmp = torch.index_select(
                            noise_guidance_edit.to(self.device), 0, warmup_inds
                        )
                        noise_guidance_edit_tmp = torch.einsum(
                            "cb,cbijk->bijk", concept_weights_tmp, noise_guidance_edit_tmp
                        )
                        noise_guidance_edit_tmp = noise_guidance_edit_tmp
                        noise_guidance = noise_guidance + noise_guidance_edit_tmp

                        self.sem_guidance[i] = noise_guidance_edit_tmp.detach().cpu()

                        del noise_guidance_edit_tmp
                        del concept_weights_tmp
                        concept_weights = concept_weights.to(self.device)
                        noise_guidance_edit = noise_guidance_edit.to(self.device)

                    concept_weights = torch.where(
                        concept_weights < 0, torch.zeros_like(concept_weights), concept_weights
                    )

                    concept_weights = torch.nan_to_num(concept_weights)

                    noise_guidance_edit = torch.einsum("cb,cbijk->bijk", concept_weights, noise_guidance_edit)

                    noise_guidance_edit = noise_guidance_edit + edit_momentum_scale * edit_momentum

                    edit_momentum = edit_mom_beta * edit_momentum + (1 - edit_mom_beta) * noise_guidance_edit

                    if warmup_inds.shape[0] == len(noise_pred_edit_concepts):
                        noise_guidance = noise_guidance + noise_guidance_edit
                        self.sem_guidance[i] = noise_guidance_edit.detach().cpu()

                noise_pred = noise_pred_uncond + noise_guidance

            # compute the previous noisy sample x_t -> x_t-1
            if use_ddpm:
                idx = t_to_idx[int(t)]
                latents = self.scheduler.step(noise_pred, t, latents, variance_noise=zs[idx],
                                              **extra_step_kwargs).prev_sample

            else: #if not use_ddpm:
                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample

            # step callback
            if use_cross_attn_mask:
                store_step = i in attn_store_steps
                if store_step:
                    print(f"storing attention for step {i}")
                self.attention_store.between_steps(store_step)

            # call the callback, if provided
            if callback is not None and i % callback_steps == 0:
                callback(i, t, latents)

        # 8. Post-processing
        if not output_type == "latent":
            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
            image, has_nsfw_concept = self.run_safety_checker(image, self.device, text_embeddings.dtype)
        else:
            image = latents
            has_nsfw_concept = None

        if has_nsfw_concept is None:
            do_denormalize = [True] * image.shape[0]
        else:
            do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]

        image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)

        if not return_dict:
            return (image, has_nsfw_concept)

        return SemanticStableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)

    def encode_text(self, prompts):
        text_inputs = self.tokenizer(
            prompts,
            padding="max_length",
            max_length=self.tokenizer.model_max_length,
            return_tensors="pt",
        )
        text_input_ids = text_inputs.input_ids

        if text_input_ids.shape[-1] > self.tokenizer.model_max_length:
            removed_text = self.tokenizer.batch_decode(text_input_ids[:, self.tokenizer.model_max_length:])
            logger.warning(
                "The following part of your input was truncated because CLIP can only handle sequences up to"
                f" {self.tokenizer.model_max_length} tokens: {removed_text}"
            )
            text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length]
        text_embeddings = self.text_encoder(text_input_ids.to(self.device))[0]

        return text_embeddings

    @torch.no_grad()
    def invert(self,
               image_path: str,
               source_prompt: str = "",
               source_guidance_scale=3.5,
               num_inversion_steps: int = 30,
               skip: float = 0.15,
               eta: float = 1.0,
               generator: Optional[torch.Generator] = None,
               verbose=True,
               ):
        """
        Inverts a real image according to Algorihm 1 in https://arxiv.org/pdf/2304.06140.pdf,
        based on the code in https://github.com/inbarhub/DDPM_inversion

        returns:
        zs - noise maps
        xts - intermediate inverted latents
        """

        # self.eta = eta
        # assert (self.eta > 0)
        skip = skip/100
        print("YOOOOOOOOOOOOOOOOO ", skip, num_inversion_steps)
        train_steps = self.scheduler.config.num_train_timesteps
        timesteps = torch.from_numpy(
            np.linspace(train_steps - skip * train_steps - 1, 1, num_inversion_steps).astype(np.int64)).to(self.device)


        self.num_inversion_steps = timesteps.shape[0]
        self.scheduler.num_inference_steps = timesteps.shape[0]
        self.scheduler.timesteps = timesteps
        
        # Reset attn processor, we do not want to store attn maps during inversion
        # self.unet.set_default_attn_processor()
        self.unet.set_attn_processor(AttnProcessor())

        # 1. get embeddings

        uncond_embedding = self.encode_text("")

        # 2. encode image
        x0 = self.encode_image(image_path, dtype=uncond_embedding.dtype)
        self.batch_size = x0.shape[0]

        if not source_prompt == "":
            text_embeddings = self.encode_text(source_prompt).repeat((self.batch_size, 1, 1))
        uncond_embedding = uncond_embedding.repeat((self.batch_size, 1, 1))
        # autoencoder reconstruction
        # image_rec = self.vae.decode(x0 / self.vae.config.scaling_factor, return_dict=False)[0]
        # image_rec = self.image_processor.postprocess(image_rec, output_type="pil")
        # 3. find zs and xts
        variance_noise_shape = (
            self.num_inversion_steps,
            self.batch_size,
            self.unet.config.in_channels,
            self.unet.sample_size,
            self.unet.sample_size)

        # intermediate latents
        t_to_idx = {int(v): k for k, v in enumerate(timesteps)}
        xts = torch.zeros(size=variance_noise_shape, device=self.device, dtype=uncond_embedding.dtype)

        for t in reversed(timesteps):
            idx = self.num_inversion_steps-t_to_idx[int(t)] - 1
            noise = randn_tensor(shape=x0.shape, generator=generator, device=self.device, dtype=x0.dtype)
            xts[idx] = self.scheduler.add_noise(x0, noise, t)
        xts = torch.cat([x0.unsqueeze(0), xts], dim=0)

        reset_dpm(self.scheduler)
        # noise maps
        zs = torch.zeros(size=variance_noise_shape, device=self.device, dtype=uncond_embedding.dtype)

        for t in self.progress_bar(timesteps, verbose=verbose):

            idx = self.num_inversion_steps-t_to_idx[int(t)]-1
            # 1. predict noise residual
            xt = xts[idx+1]

            noise_pred = self.unet(xt, timestep=t, encoder_hidden_states=uncond_embedding).sample

            if not source_prompt == "":
                noise_pred_cond = self.unet(xt, timestep=t, encoder_hidden_states=text_embeddings).sample
                noise_pred = noise_pred + source_guidance_scale * (noise_pred_cond - noise_pred)

            xtm1 = xts[idx]
            z, xtm1_corrected = compute_noise(self.scheduler, xtm1, xt, t, noise_pred, eta)
            zs[idx] = z

            # correction to avoid error accumulation
            xts[idx] = xtm1_corrected

        # TODO: I don't think that the noise map for the last step should be discarded ?!
        # if not zs is None:
        #     zs[-1] = torch.zeros_like(zs[-1])
        # self.init_latents = xts[-1].expand(self.batch_size, -1, -1, -1)
        zs = zs.flip(0)
        # self.zs = zs


        return zs, xts
        # return zs, xts, image_rec

    @torch.no_grad()
    def encode_image(self, image_path, dtype=None):
        image = load_512(image_path,
                         size=self.unet.sample_size * self.vae_scale_factor,
                         device=self.device,
                         dtype=dtype)
        x0 = self.vae.encode(image).latent_dist.mode()
        x0 = self.vae.config.scaling_factor * x0
        return x0

def compute_noise_ddim(scheduler, prev_latents, latents, timestep, noise_pred, eta):
    # 1. get previous step value (=t-1)
    prev_timestep = timestep - scheduler.config.num_train_timesteps // scheduler.num_inference_steps

    # 2. compute alphas, betas
    alpha_prod_t = scheduler.alphas_cumprod[timestep]
    alpha_prod_t_prev = (
        scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else scheduler.final_alpha_cumprod
    )

    beta_prod_t = 1 - alpha_prod_t

    # 3. compute predicted original sample from predicted noise also called
    # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
    pred_original_sample = (latents - beta_prod_t ** (0.5) * noise_pred) / alpha_prod_t ** (0.5)

    # 4. Clip "predicted x_0"
    if scheduler.config.clip_sample:
        pred_original_sample = torch.clamp(pred_original_sample, -1, 1)

    # 5. compute variance: "sigma_t(η)" -> see formula (16)
    # σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1)
    variance = scheduler._get_variance(timestep, prev_timestep)
    std_dev_t = eta * variance ** (0.5)

    # 6. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
    pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t ** 2) ** (0.5) * noise_pred

    # modifed so that updated xtm1 is returned as well (to avoid error accumulation)
    mu_xt = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
    noise = (prev_latents - mu_xt) / (variance ** (0.5) * eta)

    return noise, mu_xt + (eta * variance ** 0.5) * noise

# Copied from pipelines.StableDiffusion.CycleDiffusionPipeline.compute_noise
def compute_noise_sde_dpm_pp_2nd(scheduler, prev_latents, latents, timestep, noise_pred, eta):

    def first_order_update(model_output, timestep, prev_timestep, sample):
        lambda_t, lambda_s = scheduler.lambda_t[prev_timestep], scheduler.lambda_t[timestep]
        alpha_t, alpha_s = scheduler.alpha_t[prev_timestep], scheduler.alpha_t[timestep]
        sigma_t, sigma_s = scheduler.sigma_t[prev_timestep], scheduler.sigma_t[timestep]
        h = lambda_t - lambda_s

        mu_xt = (
                (sigma_t / sigma_s * torch.exp(-h)) * sample
                + (alpha_t * (1 - torch.exp(-2.0 * h))) * model_output
        )
        sigma = sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h))

        noise = (prev_latents - mu_xt) / sigma

        prev_sample = mu_xt + sigma * noise

        return noise, prev_sample
    def second_order_update(model_output_list, timestep_list, prev_timestep, sample):
        t, s0, s1 = prev_timestep, timestep_list[-1], timestep_list[-2]
        m0, m1 = model_output_list[-1], model_output_list[-2]
        lambda_t, lambda_s0, lambda_s1 = scheduler.lambda_t[t], scheduler.lambda_t[s0], scheduler.lambda_t[s1]
        alpha_t, alpha_s0 = scheduler.alpha_t[t], scheduler.alpha_t[s0]
        sigma_t, sigma_s0 = scheduler.sigma_t[t], scheduler.sigma_t[s0]
        h, h_0 = lambda_t - lambda_s0, lambda_s0 - lambda_s1
        r0 = h_0 / h
        D0, D1 = m0, (1.0 / r0) * (m0 - m1)

        mu_xt = (
            (sigma_t / sigma_s0 * torch.exp(-h)) * sample
                    + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0
                    + 0.5 * (alpha_t * (1 - torch.exp(-2.0 * h))) * D1
        )
        sigma = sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h))

        noise = (prev_latents - mu_xt) / sigma

        prev_sample = mu_xt + sigma * noise

        return noise, prev_sample

    step_index = (scheduler.timesteps == timestep).nonzero()
    if len(step_index) == 0:
        step_index = len(scheduler.timesteps) - 1
    else:
        step_index = step_index.item()

    prev_timestep = 0 if step_index == len(scheduler.timesteps) - 1 else scheduler.timesteps[step_index + 1]

    model_output = scheduler.convert_model_output(noise_pred, timestep, latents)

    for i in range(scheduler.config.solver_order - 1):
        scheduler.model_outputs[i] = scheduler.model_outputs[i + 1]
    scheduler.model_outputs[-1] = model_output

    if scheduler.lower_order_nums < 1:
        noise, prev_sample = first_order_update(model_output, timestep, prev_timestep, latents)
    else:
        timestep_list = [scheduler.timesteps[step_index - 1], timestep]
        noise, prev_sample = second_order_update(scheduler.model_outputs, timestep_list, prev_timestep, latents)

    if scheduler.lower_order_nums < scheduler.config.solver_order:
        scheduler.lower_order_nums += 1

    return noise, prev_sample

def compute_noise(scheduler, *args):
    if isinstance(scheduler, DDIMScheduler):
        return compute_noise_ddim(scheduler, *args)
    elif isinstance(scheduler, DPMSolverMultistepSchedulerInject) and scheduler.config.algorithm_type == 'sde-dpmsolver++'\
            and scheduler.config.solver_order == 2:
        return compute_noise_sde_dpm_pp_2nd(scheduler, *args)
    else:
        raise NotImplementedError