File size: 43,305 Bytes
4db9215
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
UNIStainNet: Pixel-Space UNI-Guided Virtual Staining Network.

Architecture:
    Generator: SPADE-UNet conditioned on UNI pathology features + stain/class embedding
    Discriminator: Multi-scale PatchGAN (512 + 256)
    Losses: LPIPS@128 + adversarial + DAB intensity + DAB contrast

References:
    - Park et al., "Semantic Image Synthesis with SPADE" (CVPR 2019)
    - Chen et al., "A general-purpose self-supervised model for pathology" (Nature Medicine 2024)
    - Isola et al., "Image-to-Image Translation with pix2pix" (CVPR 2017)
"""

import copy
from collections import OrderedDict

import torch
import torch.nn as nn
import torch.nn.functional as F
import lpips
import pytorch_lightning as pl
import torchvision
import wandb

from src.models.discriminator import (
    PatchDiscriminator, MultiScaleDiscriminator,
    hinge_loss_d, hinge_loss_g, r1_gradient_penalty, feature_matching_loss,
)
from src.models.generator import SPADEUNetGenerator
from src.models.losses import VGGFeatureExtractor, gram_matrix, PatchNCELoss
from src.utils.dab import DABExtractor


# ======================================================================
# Training Module
# ======================================================================

class UNIStainNetTrainer(pl.LightningModule):
    """PyTorch Lightning training module for UNIStainNet.

    Handles GAN training with manual optimization, CFG dropout, EMA, and
    all loss computations.
    """

    def __init__(
        self,
        # Architecture
        num_classes=5,
        null_class=4,
        class_dim=64,
        uni_dim=1024,
        ndf=64,
        disc_n_layers=3,
        input_skip=False,
        # Optimizer
        gen_lr=1e-4,
        disc_lr=4e-4,
        warmup_steps=1000,
        # Loss weights
        lpips_weight=1.0,
        lpips_256_weight=0.5,
        lpips_512_weight=0.0,
        adversarial_weight=1.0,
        dab_intensity_weight=0.1,
        dab_contrast_weight=0.05,
        dab_sharpness_weight=0.0,
        gram_style_weight=0.0,
        edge_weight=0.0,
        he_edge_weight=0.0,
        bg_white_weight=0.0,
        bg_threshold=0.85,
        l1_lowres_weight=0.0,
        edge_encoder=False,
        edge_base_ch=32,
        uni_spatial_size=4,
        uncond_disc_weight=0.0,
        crop_disc_weight=0.0,
        crop_size=128,
        feat_match_weight=0.0,
        patchnce_weight=0.0,
        patchnce_layers=(2, 3, 4),
        patchnce_n_patches=256,
        patchnce_temperature=0.07,
        # Ablation
        disable_uni=False,
        disable_class=False,
        # GAN training
        r1_weight=10.0,
        r1_every=16,
        adversarial_start_step=2000,
        # CFG
        cfg_drop_class_prob=0.10,
        cfg_drop_uni_prob=0.10,
        cfg_drop_both_prob=0.05,
        # EMA
        ema_decay=0.999,
        # On-the-fly UNI extraction (for crop-based training)
        extract_uni_on_the_fly=False,
        uni_spatial_pool_size=32,
        # Resolution
        image_size=512,
        # 1024 architecture: extend UNI SPADE to 512 level
        uni_spade_at_512=False,
        # Per-label names for multi-stain logging
        label_names=None,
    ):
        super().__init__()
        self.save_hyperparameters()
        self.automatic_optimization = False

        self.null_class = null_class

        # On-the-fly UNI feature extraction (loaded lazily on first use)
        self._uni_model = None
        self._uni_extract_on_the_fly = extract_uni_on_the_fly

        # Generator
        self.generator = SPADEUNetGenerator(
            num_classes=num_classes,
            class_dim=class_dim,
            uni_dim=uni_dim,
            input_skip=input_skip,
            edge_encoder=edge_encoder,
            edge_base_ch=edge_base_ch,
            uni_spatial_size=uni_spatial_size,
            image_size=image_size,
            uni_spade_at_512=uni_spade_at_512,
        )

        # Discriminator (global multi-scale)
        self.discriminator = MultiScaleDiscriminator(
            in_channels=6, ndf=ndf, n_layers=disc_n_layers,
        )

        # Crop discriminator (local full-res detail)
        if crop_disc_weight > 0:
            self.crop_discriminator = PatchDiscriminator(
                in_channels=6, ndf=ndf, n_layers=disc_n_layers,
            )
        else:
            self.crop_discriminator = None

        # Unconditional discriminator (HER2-only, alignment-free texture judge)
        # Also needed for feature matching loss (FM uses uncond disc features)
        if uncond_disc_weight > 0 or feat_match_weight > 0:
            self.uncond_discriminator = PatchDiscriminator(
                in_channels=3, ndf=ndf, n_layers=disc_n_layers,
            )
        else:
            self.uncond_discriminator = None

        # PatchNCE loss (contrastive, alignment-free: H&E input vs generated)
        if patchnce_weight > 0:
            # Encoder channel dims: {1: 64, 2: 128, 3: 256, 4: 512}
            enc_channels = {1: 64, 2: 128, 3: 256, 4: 512}
            layer_channels = {l: enc_channels[l] for l in patchnce_layers}
            self.patchnce_loss = PatchNCELoss(
                layer_channels=layer_channels,
                num_patches=patchnce_n_patches,
                temperature=patchnce_temperature,
            )
        else:
            self.patchnce_loss = None

        # EMA generator
        self.generator_ema = copy.deepcopy(self.generator)
        self.generator_ema.requires_grad_(False)

        # Losses
        self.lpips_fn = lpips.LPIPS(net='alex')
        self.lpips_fn.requires_grad_(False)
        self.lpips_fn.eval()

        self.dab_extractor = DABExtractor(device='cpu')

        # VGG feature extractor for Gram-matrix style loss
        if gram_style_weight > 0:
            self.vgg_extractor = VGGFeatureExtractor()
            self.vgg_extractor.requires_grad_(False)
            self.vgg_extractor.eval()
        else:
            self.vgg_extractor = None

        # Param counts
        n_gen = sum(p.numel() for p in self.generator.parameters())
        n_disc = sum(p.numel() for p in self.discriminator.parameters())
        n_crop = sum(p.numel() for p in self.crop_discriminator.parameters()) if self.crop_discriminator else 0
        n_uncond = sum(p.numel() for p in self.uncond_discriminator.parameters()) if self.uncond_discriminator else 0
        print(f"Generator: {n_gen:,} params")
        print(f"Discriminator: {n_disc:,} params (global) + {n_crop:,} (crop) + {n_uncond:,} (uncond)")

    def configure_optimizers(self):
        gen_params = list(self.generator.parameters())
        if self.patchnce_loss is not None:
            gen_params += list(self.patchnce_loss.parameters())
        opt_g = torch.optim.Adam(
            gen_params,
            lr=self.hparams.gen_lr,
            betas=(0.0, 0.999),
        )
        # All discriminator params in one optimizer
        disc_params = list(self.discriminator.parameters())
        if self.crop_discriminator is not None:
            disc_params += list(self.crop_discriminator.parameters())
        if self.uncond_discriminator is not None:
            disc_params += list(self.uncond_discriminator.parameters())
        opt_d = torch.optim.Adam(
            disc_params,
            lr=self.hparams.disc_lr,
            betas=(0.0, 0.999),
        )
        return [opt_g, opt_d]

    def _get_lr_scale(self):
        """Linear warmup."""
        if self.global_step < self.hparams.warmup_steps:
            return self.global_step / max(1, self.hparams.warmup_steps)
        return 1.0

    @torch.no_grad()
    def _update_ema(self):
        """Update EMA generator weights."""
        decay = self.hparams.ema_decay
        for p_ema, p in zip(self.generator_ema.parameters(), self.generator.parameters()):
            p_ema.data.mul_(decay).add_(p.data, alpha=1 - decay)

    def on_save_checkpoint(self, checkpoint):
        """Exclude frozen UNI model from checkpoint (it's reloaded on-the-fly)."""
        state_dict = checkpoint.get('state_dict', {})
        keys_to_remove = [k for k in state_dict if k.startswith('_uni_model.')]
        for k in keys_to_remove:
            del state_dict[k]

    def on_load_checkpoint(self, checkpoint):
        """Filter out UNI model keys from old checkpoints that included them."""
        state_dict = checkpoint.get('state_dict', {})
        keys_to_remove = [k for k in state_dict if k.startswith('_uni_model.')]
        for k in keys_to_remove:
            del state_dict[k]

    def _load_uni_model(self):
        """Lazily load UNI ViT-L/16 for on-the-fly feature extraction."""
        if self._uni_model is None:
            import timm
            self._uni_model = timm.create_model(
                "hf-hub:MahmoodLab/uni",
                pretrained=True,
                init_values=1e-5,
                dynamic_img_size=True,
            )
            self._uni_model.eval()
            self._uni_model.requires_grad_(False)
            self._uni_model = self._uni_model.to(self.device)
            n_params = sum(p.numel() for p in self._uni_model.parameters())
            print(f"UNI model loaded for on-the-fly extraction: {n_params:,} params")
        return self._uni_model

    @torch.no_grad()
    def _extract_uni_from_sub_crops(self, uni_sub_crops):
        """Extract UNI features from pre-prepared sub-crops on GPU.

        Args:
            uni_sub_crops: [B, 16, 3, 224, 224] — batch of 4x4 sub-crop grids,
                           already normalized with ImageNet stats.

        Returns:
            uni_features: [B, S*S, 1024] where S = uni_spatial_pool_size (default 32)
        """
        uni_model = self._load_uni_model()
        B = uni_sub_crops.shape[0]
        spatial_size = self.hparams.uni_spatial_pool_size
        num_crops = 4  # 4x4 grid
        patches_per_side = 14  # 224/16

        # Batched UNI forward: [B, 16, 3, 224, 224] -> [B*16, 3, 224, 224]
        all_crops = uni_sub_crops.reshape(B * 16, 3, 224, 224).to(self.device)
        all_feats = uni_model.forward_features(all_crops)  # [B*16, 197, 1024]
        patch_tokens = all_feats[:, 1:, :]  # [B*16, 196, 1024]

        # Reshape back to per-sample grids: [B, 4, 4, 14, 14, 1024]
        patch_tokens = patch_tokens.reshape(
            B, num_crops, num_crops,
            patches_per_side, patches_per_side, 1024
        )
        # Interleave to spatial grid: [B, 56, 56, 1024]
        full_size = num_crops * patches_per_side  # 56
        full_grid = patch_tokens.permute(0, 1, 3, 2, 4, 5)
        full_grid = full_grid.reshape(B, full_size, full_size, 1024)

        # Pool to target spatial size (batched)
        if spatial_size < full_size:
            grid_bchw = full_grid.permute(0, 3, 1, 2)  # [B, 1024, 56, 56]
            pooled = F.adaptive_avg_pool2d(grid_bchw, spatial_size)  # [B, 1024, S, S]
            result = pooled.permute(0, 2, 3, 1)  # [B, S, S, 1024]
        else:
            result = full_grid

        S = result.shape[1]
        return result.reshape(B, S * S, 1024)  # [B, S*S, 1024]

    def _apply_cfg_dropout(self, labels, uni_features):
        """Apply classifier-free guidance dropout during training (vectorized)."""
        B = labels.shape[0]
        device = labels.device

        new_labels = labels.clone()
        new_uni = uni_features.clone()

        r = torch.rand(B, device=device)
        p_both = self.hparams.cfg_drop_both_prob
        p_class = p_both + self.hparams.cfg_drop_class_prob
        p_uni = p_class + self.hparams.cfg_drop_uni_prob

        drop_both = r < p_both
        drop_class = (r >= p_both) & (r < p_class)
        drop_uni = (r >= p_class) & (r < p_uni)

        new_labels[drop_both | drop_class] = self.null_class
        new_uni[drop_both | drop_uni] = 0.0

        return new_labels, new_uni

    def compute_dab_intensity_loss(self, generated, target):
        """Top-10% percentile matching for DAB intensity."""
        with torch.amp.autocast('cuda', enabled=False):
            gen = generated.float()
            tgt = target.float()

            dab_gen = self.dab_extractor.extract_dab_intensity(gen, normalize="none")
            dab_tgt = self.dab_extractor.extract_dab_intensity(tgt, normalize="none")

            def _batched_top10_mean(dab):
                """Compute mean of top-10% DAB intensity per sample (vectorized)."""
                B = dab.shape[0]
                flat = dab.reshape(B, -1)  # [B, H*W]
                p99 = torch.quantile(flat, 0.99, dim=1, keepdim=True)
                flat = flat.clamp(max=p99)
                p90 = torch.quantile(flat, 0.9, dim=1, keepdim=True)
                mask = flat >= p90  # [B, H*W]
                # Use masked mean: sum(vals * mask) / sum(mask), fallback to flat mean
                masked_sum = (flat * mask).sum(dim=1)
                mask_count = mask.sum(dim=1).clamp(min=1)
                return masked_sum / mask_count  # [B]

            gen_scores = _batched_top10_mean(dab_gen)
            tgt_scores = _batched_top10_mean(dab_tgt)
            return F.l1_loss(gen_scores, tgt_scores)

    def compute_dab_contrast_loss(self, generated, labels):
        """Class-ordering hinge loss: DAB(3+) > DAB(2+) > DAB(1+) > DAB(0)."""
        with torch.amp.autocast('cuda', enabled=False):
            gen = generated.float()
            # Only use non-null labels
            valid = labels < self.null_class
            if valid.sum() < 2:
                return torch.tensor(0.0, device=self.device, requires_grad=True)

            gen_valid = gen[valid]
            labels_valid = labels[valid]

            dab_gen = self.dab_extractor.extract_dab_intensity(gen_valid, normalize="none")

            B = dab_gen.shape[0]
            flat = dab_gen.reshape(B, -1)
            p99 = torch.quantile(flat, 0.99, dim=1, keepdim=True)
            flat = flat.clamp(max=p99)
            p90 = torch.quantile(flat, 0.9, dim=1, keepdim=True)
            mask = flat >= p90
            masked_sum = (flat * mask).sum(dim=1)
            mask_count = mask.sum(dim=1).clamp(min=1)
            dab_scores = masked_sum / mask_count

            class_pairs = [
                (3, 0, 0.20), (3, 1, 0.15),
                (2, 0, 0.08), (3, 2, 0.10),
            ]

            losses = []
            for high_cls, low_cls, margin in class_pairs:
                high_mask = labels_valid == high_cls
                low_mask = labels_valid == low_cls
                if high_mask.sum() > 0 and low_mask.sum() > 0:
                    high_score = dab_scores[high_mask].mean()
                    low_score = dab_scores[low_mask].mean()
                    losses.append(F.relu(margin - (high_score - low_score)))

            if losses:
                return torch.stack(losses).mean()
            return torch.tensor(0.0, device=self.device, requires_grad=True)

    def compute_edge_loss(self, generated, target):
        """Fourier spectral loss at 256x256 for boundary sharpness.

        Compares power spectrum magnitudes between generated and target.
        The Fourier magnitude is inherently translation-invariant — shifting
        an image doesn't change its frequency content — so this is robust to
        the ~30px misalignment in consecutive-cut BCI pairs.

        Focuses on high-frequency bands (outer 75% of spectrum) where
        blurriness manifests as reduced power.
        """
        with torch.amp.autocast('cuda', enabled=False):
            gen = F.interpolate(generated.float(), size=256, mode='bilinear', align_corners=False)
            tgt = F.interpolate(target.float(), size=256, mode='bilinear', align_corners=False)

            # Grayscale
            gen_gray = gen.mean(dim=1, keepdim=True)
            tgt_gray = tgt.mean(dim=1, keepdim=True)

            # 2D FFT -> power spectrum (log-scale for stability)
            gen_fft = torch.fft.fft2(gen_gray)
            tgt_fft = torch.fft.fft2(tgt_gray)
            gen_mag = torch.log1p(gen_fft.abs())
            tgt_mag = torch.log1p(tgt_fft.abs())

            # High-frequency mask: keep outer 75% of spectrum
            H, W = gen_mag.shape[-2], gen_mag.shape[-1]
            cy, cx = H // 2, W // 2
            y = torch.arange(H, device=gen.device).float() - cy
            x = torch.arange(W, device=gen.device).float() - cx
            dist = (y[:, None] ** 2 + x[None, :] ** 2).sqrt()
            max_dist = (cy ** 2 + cx ** 2) ** 0.5
            hf_mask = (dist > 0.25 * max_dist).float()

            # L1 on high-frequency magnitudes
            return F.l1_loss(gen_mag * hf_mask, tgt_mag * hf_mask)

    def compute_dab_sharpness_loss(self, generated, target):
        """DAB spatial sharpness loss: penalizes diffuse brown, rewards membrane-localized DAB.

        Two components:
        1. DAB gradient magnitude: mean Sobel gradient magnitude per image.
        2. DAB local variance distribution: sorted-L1 (Wasserstein-1) on
           patch variance vectors.
        """
        with torch.amp.autocast('cuda', enabled=False):
            gen = generated.float()
            tgt = target.float()

            dab_gen = self.dab_extractor.extract_dab_intensity(gen, normalize="none")
            dab_tgt = self.dab_extractor.extract_dab_intensity(tgt, normalize="none")

            # Ensure [B, 1, H, W]
            if dab_gen.dim() == 3:
                dab_gen = dab_gen.unsqueeze(1)
            if dab_tgt.dim() == 3:
                dab_tgt = dab_tgt.unsqueeze(1)

            B = dab_gen.shape[0]

            # --- Component 1: Gradient magnitude (batched) ---
            sobel_x = torch.tensor([[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]],
                                   dtype=torch.float32, device=gen.device).view(1, 1, 3, 3)
            sobel_y = sobel_x.transpose(-1, -2)

            gx_gen = F.conv2d(dab_gen, sobel_x, padding=1)
            gy_gen = F.conv2d(dab_gen, sobel_y, padding=1)
            grad_gen = (gx_gen**2 + gy_gen**2 + 1e-8).sqrt()

            gx_tgt = F.conv2d(dab_tgt, sobel_x, padding=1)
            gy_tgt = F.conv2d(dab_tgt, sobel_y, padding=1)
            grad_tgt = (gx_tgt**2 + gy_tgt**2 + 1e-8).sqrt()

            # Match mean gradient magnitude per image
            grad_loss = F.l1_loss(grad_gen.mean(dim=[1, 2, 3]), grad_tgt.mean(dim=[1, 2, 3]))

            # --- Component 2: Local variance distribution (sorted-L1) ---
            ps = 16  # patch size
            var_losses = []
            for i in range(B):
                g = dab_gen[i, 0]  # [H, W]
                t = dab_tgt[i, 0]

                H, W = g.shape
                nH, nW = H // ps, W // ps
                g_patches = g[:nH*ps, :nW*ps].reshape(nH, ps, nW, ps).permute(0, 2, 1, 3).reshape(-1, ps*ps)
                t_patches = t[:nH*ps, :nW*ps].reshape(nH, ps, nW, ps).permute(0, 2, 1, 3).reshape(-1, ps*ps)

                g_var = g_patches.var(dim=1)
                t_var = t_patches.var(dim=1)

                g_sorted, _ = g_var.sort()
                t_sorted, _ = t_var.sort()
                var_losses.append(F.l1_loss(g_sorted, t_sorted.detach()))

            var_loss = torch.stack(var_losses).mean()

            return grad_loss + var_loss

    def compute_he_edge_loss(self, generated, he_input):
        """H&E edge structure preservation loss.

        Extracts Sobel edges from H&E input and generated output, then
        computes L1 loss between edge maps at multiple scales.
        """
        with torch.amp.autocast('cuda', enabled=False):
            gen = generated.float()
            he = he_input.float()

            gen_gray = ((gen + 1) / 2).mean(dim=1, keepdim=True)
            he_gray = ((he + 1) / 2).mean(dim=1, keepdim=True)

            sobel_x = torch.tensor([[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]],
                                   dtype=torch.float32, device=gen.device).view(1, 1, 3, 3)
            sobel_y = sobel_x.transpose(-1, -2)

            loss = 0.0
            full_size = gen_gray.shape[-1]
            scales = [full_size, full_size // 2]
            for size in scales:
                if size < full_size:
                    g = F.interpolate(gen_gray, size=size, mode='bilinear', align_corners=False)
                    h = F.interpolate(he_gray, size=size, mode='bilinear', align_corners=False)
                else:
                    g, h = gen_gray, he_gray

                gx_gen = F.conv2d(g, sobel_x, padding=1)
                gy_gen = F.conv2d(g, sobel_y, padding=1)
                edge_gen = (gx_gen**2 + gy_gen**2 + 1e-8).sqrt()

                gx_he = F.conv2d(h, sobel_x, padding=1)
                gy_he = F.conv2d(h, sobel_y, padding=1)
                edge_he = (gx_he**2 + gy_he**2 + 1e-8).sqrt()

                loss = loss + F.l1_loss(edge_gen, edge_he.detach())

            return loss / 2.0

    def compute_background_loss(self, generated, he_input):
        """Background white loss: push background regions toward white."""
        with torch.amp.autocast('cuda', enabled=False):
            gen = generated.float()
            he = he_input.float()

            he_bright = ((he + 1) / 2).mean(dim=1, keepdim=True)

            threshold = self.hparams.bg_threshold
            mask = torch.sigmoid((he_bright - threshold) * 20.0)

            white_target = torch.ones_like(gen)
            diff = (gen - white_target).abs()

            weighted_diff = diff * mask

            mask_sum = mask.sum() * 3
            if mask_sum > 0:
                return weighted_diff.sum() / mask_sum
            return torch.tensor(0.0, device=gen.device, requires_grad=True)

    def compute_gram_style_loss(self, generated, target):
        """Gram-matrix style loss: match texture statistics via VGG feature correlations."""
        with torch.amp.autocast('cuda', enabled=False):
            gen = generated.float()
            tgt = target.float()

            gen_256 = F.interpolate(gen, size=256, mode='bilinear', align_corners=False)
            tgt_256 = F.interpolate(tgt, size=256, mode='bilinear', align_corners=False)

            gen_feats = self.vgg_extractor(gen_256)
            tgt_feats = self.vgg_extractor(tgt_256)

            loss = 0.0
            for gf, tf in zip(gen_feats, tgt_feats):
                gram_gen = gram_matrix(gf)
                gram_tgt = gram_matrix(tf)
                loss = loss + F.l1_loss(gram_gen, gram_tgt.detach())

            return loss / len(gen_feats)

    def training_step(self, batch, batch_idx):
        he, her2, uni_or_crops, labels, fnames = batch
        opt_g, opt_d = self.optimizers()

        # On-the-fly UNI extraction: dataset returns [B, 16, 3, 224, 224] sub-crops
        if self._uni_extract_on_the_fly:
            uni = self._extract_uni_from_sub_crops(uni_or_crops)
        else:
            uni = uni_or_crops

        # Apply CFG dropout
        labels_dropped, uni_dropped = self._apply_cfg_dropout(labels, uni)

        # Ablation: zero out UNI features
        if self.hparams.disable_uni:
            uni_dropped = torch.zeros_like(uni_dropped)

        # Ablation: force all labels to null class
        if self.hparams.disable_class:
            labels_dropped = torch.full_like(labels_dropped, self.hparams.null_class)

        # ----------------------------------------------------------------
        # Generator step
        # ----------------------------------------------------------------
        generated = self.generator(he, uni_dropped, labels_dropped)

        # LPIPS main: 4x downsample (128 for 512 input, 256 for 1024)
        lpips_main_size = self.hparams.image_size // 4
        gen_lpips = F.interpolate(generated, size=lpips_main_size, mode='bilinear', align_corners=False)
        her2_lpips = F.interpolate(her2, size=lpips_main_size, mode='bilinear', align_corners=False)
        loss_lpips = self.lpips_fn(gen_lpips, her2_lpips).mean()

        loss_g = self.hparams.lpips_weight * loss_lpips

        # LPIPS fine: 2x downsample (256 for 512 input, 512 for 1024)
        if self.hparams.lpips_256_weight > 0:
            lpips_fine_size = self.hparams.image_size // 2
            gen_fine = F.interpolate(generated, size=lpips_fine_size, mode='bilinear', align_corners=False)
            her2_fine = F.interpolate(her2, size=lpips_fine_size, mode='bilinear', align_corners=False)
            loss_lpips_256 = self.lpips_fn(gen_fine, her2_fine).mean()
            loss_g = loss_g + self.hparams.lpips_256_weight * loss_lpips_256
            self.log('train/lpips_fine', loss_lpips_256, prog_bar=False)

        # LPIPS at full resolution (expensive)
        if self.hparams.lpips_512_weight > 0:
            loss_lpips_512 = self.lpips_fn(generated, her2).mean()
            loss_g = loss_g + self.hparams.lpips_512_weight * loss_lpips_512
            self.log('train/lpips_fullres', loss_lpips_512, prog_bar=False)

        # Low-resolution L1 (color fidelity, misalignment-robust at 64x64)
        if self.hparams.l1_lowres_weight > 0:
            gen_64 = F.interpolate(generated, size=64, mode='bilinear', align_corners=False)
            her2_64 = F.interpolate(her2, size=64, mode='bilinear', align_corners=False)
            loss_l1_lowres = F.l1_loss(gen_64, her2_64)
            loss_g = loss_g + self.hparams.l1_lowres_weight * loss_l1_lowres
            self.log('train/l1_lowres', loss_l1_lowres, prog_bar=False)

        # DAB losses (use original labels, not dropped)
        if self.hparams.dab_intensity_weight > 0:
            loss_dab = self.compute_dab_intensity_loss(generated, her2)
            loss_g = loss_g + self.hparams.dab_intensity_weight * loss_dab
            self.log('train/dab_intensity', loss_dab, prog_bar=False)

        if self.hparams.dab_contrast_weight > 0:
            loss_dab_contrast = self.compute_dab_contrast_loss(generated, labels)
            loss_g = loss_g + self.hparams.dab_contrast_weight * loss_dab_contrast
            self.log('train/dab_contrast', loss_dab_contrast, prog_bar=False)

        # Edge loss (boundary sharpness)
        if self.hparams.edge_weight > 0:
            loss_edge = self.compute_edge_loss(generated, her2)
            loss_g = loss_g + self.hparams.edge_weight * loss_edge
            self.log('train/edge_loss', loss_edge, prog_bar=False)

        # DAB sharpness loss (membrane-localized vs diffuse brown)
        if self.hparams.dab_sharpness_weight > 0:
            loss_dab_sharp = self.compute_dab_sharpness_loss(generated, her2)
            loss_g = loss_g + self.hparams.dab_sharpness_weight * loss_dab_sharp
            self.log('train/dab_sharpness', loss_dab_sharp, prog_bar=False)

        # Gram-matrix style loss
        if self.hparams.gram_style_weight > 0 and self.vgg_extractor is not None:
            loss_gram = self.compute_gram_style_loss(generated, her2)
            loss_g = loss_g + self.hparams.gram_style_weight * loss_gram
            self.log('train/gram_style', loss_gram, prog_bar=False)

        # H&E edge structure preservation (pixel-aligned)
        if self.hparams.he_edge_weight > 0:
            loss_he_edge = self.compute_he_edge_loss(generated, he)
            loss_g = loss_g + self.hparams.he_edge_weight * loss_he_edge
            self.log('train/he_edge', loss_he_edge, prog_bar=False)

        # Background white loss
        if self.hparams.bg_white_weight > 0:
            loss_bg = self.compute_background_loss(generated, he)
            loss_g = loss_g + self.hparams.bg_white_weight * loss_bg
            self.log('train/bg_white', loss_bg, prog_bar=False)

        # PatchNCE loss (contrastive: H&E input vs generated, never sees GT)
        if self.hparams.patchnce_weight > 0 and self.patchnce_loss is not None:
            feats_he = self.generator.encode(he)
            feats_gen = self.generator.encode(generated)
            loss_nce = self.patchnce_loss(feats_he, feats_gen)
            loss_g = loss_g + self.hparams.patchnce_weight * loss_nce
            self.log('train/patchnce', loss_nce, prog_bar=False)

        # Adversarial losses (after warmup)
        loss_adv = torch.tensor(0.0, device=self.device)
        loss_feat_match = torch.tensor(0.0, device=self.device)
        loss_crop_adv = torch.tensor(0.0, device=self.device)
        loss_uncond_adv = torch.tensor(0.0, device=self.device)
        any_adv = (self.hparams.adversarial_weight > 0 or
                   self.hparams.uncond_disc_weight > 0 or
                   self.hparams.crop_disc_weight > 0 or
                   self.hparams.feat_match_weight > 0)
        img_sz = self.hparams.image_size
        # Pre-compute disc-resolution tensors (512 for 1024 input, identity for 512)
        if img_sz == 1024:
            he_for_disc = F.interpolate(he, size=512, mode='bilinear', align_corners=False)
            her2_for_disc = F.interpolate(her2, size=512, mode='bilinear', align_corners=False)
        else:
            he_for_disc = he
            her2_for_disc = her2
        if self.global_step >= self.hparams.adversarial_start_step and any_adv:
            if img_sz == 1024:
                gen_for_disc = F.interpolate(generated, size=512, mode='bilinear', align_corners=False)
            else:
                gen_for_disc = generated

            # Conditional discriminator (paired: generated+HE vs real_HER2+HE)
            if self.hparams.adversarial_weight > 0:
                fake_input = torch.cat([gen_for_disc, he_for_disc], dim=1)
                disc_outputs = self.discriminator(fake_input)
                loss_adv = sum(hinge_loss_g(out) for out in disc_outputs) / len(disc_outputs)
                loss_g = loss_g + self.hparams.adversarial_weight * loss_adv

            # Feature matching from unconditional disc
            if (self.hparams.feat_match_weight > 0 and
                    self.uncond_discriminator is not None):
                _, fake_feats = self.uncond_discriminator(gen_for_disc, return_features=True)
                with torch.no_grad():
                    _, real_feats = self.uncond_discriminator(her2_for_disc, return_features=True)
                loss_feat_match = feature_matching_loss(fake_feats, real_feats)
                loss_g = loss_g + self.hparams.feat_match_weight * loss_feat_match

            # Crop discriminator: random crops at full resolution
            if self.crop_discriminator is not None and self.hparams.crop_disc_weight > 0:
                fake_input_crop = torch.cat([generated, he], dim=1)
                cs = self.hparams.crop_size
                top = torch.randint(0, img_sz - cs, (1,)).item()
                left = torch.randint(0, img_sz - cs, (1,)).item()
                fake_crop = fake_input_crop[:, :, top:top+cs, left:left+cs]
                loss_crop_adv = hinge_loss_g(self.crop_discriminator(fake_crop))
                loss_g = loss_g + self.hparams.crop_disc_weight * loss_crop_adv

            # Unconditional discriminator: HER2-only adversarial
            if self.uncond_discriminator is not None and self.hparams.uncond_disc_weight > 0:
                loss_uncond_adv = hinge_loss_g(self.uncond_discriminator(gen_for_disc))
                loss_g = loss_g + self.hparams.uncond_disc_weight * loss_uncond_adv

        # Generator backward + step
        lr_scale = self._get_lr_scale()
        for pg in opt_g.param_groups:
            pg['lr'] = self.hparams.gen_lr * lr_scale

        opt_g.zero_grad()
        self.manual_backward(loss_g)
        torch.nn.utils.clip_grad_norm_(self.generator.parameters(), 1.0)
        opt_g.step()

        # Update EMA
        self._update_ema()

        # ----------------------------------------------------------------
        # Discriminator step
        # ----------------------------------------------------------------
        loss_d = torch.tensor(0.0, device=self.device)
        loss_crop_d = torch.tensor(0.0, device=self.device)
        loss_uncond_d = torch.tensor(0.0, device=self.device)
        if self.global_step >= self.hparams.adversarial_start_step and any_adv:
            with torch.no_grad():
                fake_detached = self.generator(he, uni_dropped, labels_dropped)

            # For 1024, downsample for disc
            if img_sz == 1024:
                fake_det_disc = F.interpolate(fake_detached, size=512, mode='bilinear', align_corners=False)
            else:
                fake_det_disc = fake_detached

            # Conditional discriminator
            if self.hparams.adversarial_weight > 0:
                real_input = torch.cat([her2_for_disc, he_for_disc], dim=1)
                fake_input = torch.cat([fake_det_disc, he_for_disc], dim=1)

                disc_real = self.discriminator(real_input)
                disc_fake = self.discriminator(fake_input)

                loss_d = sum(
                    hinge_loss_d(dr, df)
                    for dr, df in zip(disc_real, disc_fake)
                ) / len(disc_real)

            # Crop discriminator
            if self.crop_discriminator is not None and self.hparams.crop_disc_weight > 0:
                real_input_c = torch.cat([her2, he], dim=1)
                fake_input_c = torch.cat([fake_detached, he], dim=1)
                cs = self.hparams.crop_size
                top = torch.randint(0, img_sz - cs, (1,)).item()
                left = torch.randint(0, img_sz - cs, (1,)).item()
                real_crop = real_input_c[:, :, top:top+cs, left:left+cs]
                fake_crop = fake_input_c[:, :, top:top+cs, left:left+cs]
                loss_crop_d = hinge_loss_d(
                    self.crop_discriminator(real_crop),
                    self.crop_discriminator(fake_crop),
                )
                loss_d = loss_d + self.hparams.crop_disc_weight * loss_crop_d

            # Unconditional discriminator
            if self.uncond_discriminator is not None and (
                    self.hparams.uncond_disc_weight > 0 or self.hparams.feat_match_weight > 0):
                uncond_real_out = self.uncond_discriminator(her2_for_disc)
                uncond_fake_out = self.uncond_discriminator(fake_det_disc)
                loss_uncond_d = hinge_loss_d(uncond_real_out, uncond_fake_out)
                loss_d = loss_d + max(self.hparams.uncond_disc_weight, 1.0) * loss_uncond_d

            # R1 gradient penalty
            loss_r1 = torch.tensor(0.0, device=self.device)
            if self.global_step % self.hparams.r1_every == 0:
                with torch.amp.autocast('cuda', enabled=False):
                    if self.hparams.adversarial_weight > 0:
                        real_input_r1 = torch.cat([her2_for_disc, he_for_disc], dim=1).float().detach().requires_grad_(True)
                        for disc in [self.discriminator.disc_512]:
                            d_real = disc(real_input_r1)
                            grad_real = torch.autograd.grad(
                                outputs=d_real.sum(), inputs=real_input_r1,
                                create_graph=True,
                            )[0]
                            loss_r1 = loss_r1 + self.hparams.r1_weight * grad_real.pow(2).mean()
                    if self.uncond_discriminator is not None and (
                            self.hparams.uncond_disc_weight > 0 or self.hparams.feat_match_weight > 0):
                        her2_r1 = her2_for_disc.float().detach().requires_grad_(True)
                        d_real_uncond = self.uncond_discriminator(her2_r1)
                        grad_uncond = torch.autograd.grad(
                            outputs=d_real_uncond.sum(), inputs=her2_r1,
                            create_graph=True,
                        )[0]
                        loss_r1 = loss_r1 + self.hparams.r1_weight * grad_uncond.pow(2).mean()
                loss_d = loss_d + loss_r1
                self.log('train/r1_penalty', loss_r1, prog_bar=False)

            opt_d.zero_grad()
            self.manual_backward(loss_d)
            if self.hparams.adversarial_weight > 0:
                torch.nn.utils.clip_grad_norm_(self.discriminator.parameters(), 1.0)
            if self.crop_discriminator is not None:
                torch.nn.utils.clip_grad_norm_(self.crop_discriminator.parameters(), 1.0)
            if self.uncond_discriminator is not None:
                torch.nn.utils.clip_grad_norm_(self.uncond_discriminator.parameters(), 1.0)
            opt_d.step()

        # Logging
        self.log('train/loss_g', loss_g, prog_bar=True)
        self.log('train/loss_d', loss_d, prog_bar=True)
        self.log('train/lpips', loss_lpips, prog_bar=True)
        self.log('train/adversarial', loss_adv, prog_bar=False)
        self.log('train/lr_scale', lr_scale, prog_bar=False)
        if self.crop_discriminator is not None:
            self.log('train/crop_adv_g', loss_crop_adv, prog_bar=False)
            self.log('train/crop_adv_d', loss_crop_d, prog_bar=False)
        if self.uncond_discriminator is not None:
            self.log('train/uncond_adv_g', loss_uncond_adv, prog_bar=False)
            self.log('train/uncond_adv_d', loss_uncond_d, prog_bar=False)
        if self.hparams.feat_match_weight > 0:
            self.log('train/feat_match', loss_feat_match, prog_bar=False)

    def on_validation_epoch_start(self):
        # Pick a random batch index for the second sample grid
        n_val_batches = max(1, len(self.trainer.val_dataloaders))
        self._random_val_batch_idx = torch.randint(1, max(2, n_val_batches), (1,)).item()
        # Per-label sample collectors (for multi-stain visual grids)
        self._val_per_label_samples = {}

    def _log_sample_grid(self, he, her2_01, gen_01, key):
        """Log H&E | Real | Gen grid to wandb."""
        n = min(4, len(he))
        he_01 = ((he[:n].cpu() + 1) / 2).clamp(0, 1)
        grid_images = []
        for i in range(n):
            grid_images.extend([
                he_01[i],
                her2_01[i].cpu(),
                gen_01[i].cpu(),
            ])
        grid = torchvision.utils.make_grid(grid_images, nrow=3, padding=2)
        if self.logger:
            self.logger.experiment.log({
                key: [wandb.Image(grid, caption='H&E | Real | Gen')],
                'global_step': self.global_step,
            })

    def validation_step(self, batch, batch_idx):
        he, her2, uni_or_crops, labels, fnames = batch

        # On-the-fly UNI extraction
        if self._uni_extract_on_the_fly:
            uni = self._extract_uni_from_sub_crops(uni_or_crops)
        else:
            uni = uni_or_crops

        if self.hparams.disable_uni:
            uni = torch.zeros_like(uni)

        if self.hparams.disable_class:
            labels = torch.full_like(labels, self.hparams.null_class)

        # Use EMA generator
        with torch.no_grad():
            generated = self.generator_ema(he, uni, labels)

        # LPIPS (4x downsample: 128 for 512, 256 for 1024)
        lpips_size = self.hparams.image_size // 4
        gen_lpips = F.interpolate(generated, size=lpips_size, mode='bilinear', align_corners=False)
        her2_lpips = F.interpolate(her2, size=lpips_size, mode='bilinear', align_corners=False)
        lpips_val = self.lpips_fn(gen_lpips, her2_lpips).mean()

        # SSIM
        gen_01 = ((generated + 1) / 2).clamp(0, 1)
        her2_01 = ((her2 + 1) / 2).clamp(0, 1)
        from torchmetrics.functional.image import structural_similarity_index_measure
        ssim_val = structural_similarity_index_measure(gen_01, her2_01, data_range=1.0)

        # DAB MAE (canonical: mean of top-10%)
        dab_gen = self.dab_extractor.extract_dab_intensity(generated.float().cpu(), normalize="none")
        dab_real = self.dab_extractor.extract_dab_intensity(her2.float().cpu(), normalize="none")

        def p90_score(dab):
            flat = dab.flatten()
            p90 = torch.quantile(flat, 0.9)
            mask = flat >= p90
            return flat[mask].mean().item() if mask.sum() > 0 else flat.mean().item()

        dab_mae = sum(
            abs(p90_score(dab_gen[i]) - p90_score(dab_real[i]))
            for i in range(len(dab_gen))
        ) / len(dab_gen)

        self.log('val/lpips', lpips_val, prog_bar=True, sync_dist=True)
        self.log('val/ssim', ssim_val, prog_bar=True, sync_dist=True)
        self.log('val/dab_mae', dab_mae, prog_bar=True, sync_dist=True)

        # Collect per-label samples for visual grids (multi-stain only)
        if hasattr(self, '_val_per_label_samples'):
            for i in range(len(labels)):
                lbl = labels[i].item()
                if lbl == self.hparams.null_class:
                    continue
                if lbl not in self._val_per_label_samples:
                    self._val_per_label_samples[lbl] = {'he': [], 'real': [], 'gen': []}
                bucket = self._val_per_label_samples[lbl]
                if len(bucket['he']) < 4:
                    bucket['he'].append(he[i].cpu())
                    bucket['real'].append(her2_01[i].cpu())
                    bucket['gen'].append(gen_01[i].cpu())

        # Log sample grids: first batch (fixed) + one random batch
        if batch_idx == 0:
            self._log_sample_grid(he, her2_01, gen_01, 'val/samples_fixed')
        elif batch_idx == self._random_val_batch_idx:
            self._log_sample_grid(he, her2_01, gen_01, 'val/samples_random')

    def on_validation_epoch_end(self):
        """Log per-label sample grids if multiple labels are present."""
        if not hasattr(self, '_val_per_label_samples') or len(self._val_per_label_samples) <= 1:
            return

        label_names = getattr(self.hparams, 'label_names', None)

        for lbl, bucket in sorted(self._val_per_label_samples.items()):
            if not bucket['he'] or not self.logger:
                continue
            name = label_names[lbl] if label_names and lbl < len(label_names) else str(lbl)
            self._log_sample_grid(
                torch.stack(bucket['he']),
                torch.stack(bucket['real']),
                torch.stack(bucket['gen']),
                f'val/samples_{name}',
            )

        self._val_per_label_samples = {}

    @torch.no_grad()
    def generate(self, he_images, uni_features, labels,
                 num_inference_steps=None, guidance_scale=1.0, seed=None):
        """Generate IHC images from H&E input.

        Args:
            he_images: [B, 3, H, H] where H=512 or H=1024
            uni_features: [B, N, 1024] where N=16 (4x4 CLS) or N=1024 (32x32 patch)
            labels: [B] class/stain labels
            num_inference_steps: ignored (single forward pass)
            guidance_scale: CFG scale (1.0 = no guidance)
            seed: random seed (for reproducibility, though model is deterministic)
        """
        if seed is not None:
            torch.manual_seed(seed)

        gen = self.generator_ema if hasattr(self, 'generator_ema') else self.generator

        if guidance_scale <= 1.0:
            return gen(he_images, uni_features, labels)

        # Classifier-free guidance
        null_labels = torch.full_like(labels, self.null_class)

        output_cond = gen(he_images, uni_features, labels)
        output_uncond = gen(he_images, uni_features, null_labels)

        output = output_uncond + guidance_scale * (output_cond - output_uncond)
        return output.clamp(-1, 1)