Update: fix mask passing in validation, add MedicalVisualizationCallback, optimize for 2xH800

code/configs_medical/PixelGen_Medical_CVC_B16_CrossAttn.yaml

@@ -1,5 +1,6 @@
 # PixelGen Medical - CVC-ClinicDB Polyp Segmentation - Binary mask-conditional colonoscopy image generation
 # Model: JiTMedical-B/16, mask_mode=cross_attention
+# Optimized for 2x H800 80GB
 seed_everything: 1234
 tags:
   exp: &exp PixelGen_Medical_CVC_B16_CrossAttn
@@ -34,6 +35,13 @@ trainer:
       init_args:
         save_dir: val_samples
         save_compressed: true
+    - class_path: src.callbacks.medical_visualization.MedicalVisualizationCallback
+      init_args:
+        every_n_steps: 5000
+        num_samples: 8
+        num_sampling_steps: 50
+        cfg_scale: 2.0
+        save_dir: training_vis
   plugins:
     - src.plugins.bd_env.BDEnvironment
 
@@ -123,7 +131,7 @@ data:
       resolution: 256
       max_num_instances: 612
       noise_scale: 1.0
-  train_batch_size: 16
+  train_batch_size: 96
   train_num_workers: 4
-  pred_batch_size: 16
+  pred_batch_size: 32
   pred_num_workers: 1

code/configs_medical/PixelGen_Medical_CVC_B8_Spatial.yaml

@@ -1,5 +1,6 @@
 # PixelGen Medical - CVC-ClinicDB Polyp Segmentation - Binary mask-conditional colonoscopy image generation
 # Model: JiTMedical-B/8, mask_mode=spatial
+# Optimized for 2x H800 80GB
 seed_everything: 1234
 tags:
   exp: &exp PixelGen_Medical_CVC_B8_Spatial
@@ -34,6 +35,13 @@ trainer:
       init_args:
         save_dir: val_samples
         save_compressed: true
+    - class_path: src.callbacks.medical_visualization.MedicalVisualizationCallback
+      init_args:
+        every_n_steps: 5000
+        num_samples: 8
+        num_sampling_steps: 50
+        cfg_scale: 2.0
+        save_dir: training_vis
   plugins:
     - src.plugins.bd_env.BDEnvironment
 
@@ -123,7 +131,7 @@ data:
       resolution: 256
       max_num_instances: 612
       noise_scale: 1.0
-  train_batch_size: 16
+  train_batch_size: 32
   train_num_workers: 4
-  pred_batch_size: 16
+  pred_batch_size: 32
   pred_num_workers: 1

code/configs_medical/PixelGen_Medical_CVC_S8_Spatial.yaml

@@ -1,5 +1,6 @@
 # PixelGen Medical - CVC-ClinicDB Polyp Segmentation - Binary mask-conditional colonoscopy image generation
 # Model: JiTMedical-S/8, mask_mode=spatial
+# Optimized for 2x H800 80GB
 seed_everything: 1234
 tags:
   exp: &exp PixelGen_Medical_CVC_S8_Spatial
@@ -34,6 +35,13 @@ trainer:
       init_args:
         save_dir: val_samples
         save_compressed: true
+    - class_path: src.callbacks.medical_visualization.MedicalVisualizationCallback
+      init_args:
+        every_n_steps: 5000
+        num_samples: 8
+        num_sampling_steps: 50
+        cfg_scale: 2.0
+        save_dir: training_vis
   plugins:
     - src.plugins.bd_env.BDEnvironment
 
@@ -123,7 +131,7 @@ data:
       resolution: 256
       max_num_instances: 612
       noise_scale: 1.0
-  train_batch_size: 16
+  train_batch_size: 64
   train_num_workers: 4
-  pred_batch_size: 16
+  pred_batch_size: 32
   pred_num_workers: 1

code/configs_medical/PixelGen_Medical_Kvasir_B16_CrossAttn.yaml

@@ -1,5 +1,6 @@
 # PixelGen Medical - Kvasir-SEG Polyp Segmentation - Binary mask-conditional colonoscopy image generation
 # Model: JiTMedical-B/16, mask_mode=cross_attention
+# Optimized for 2x H800 80GB
 seed_everything: 1234
 tags:
   exp: &exp PixelGen_Medical_Kvasir_B16_CrossAttn
@@ -34,6 +35,13 @@ trainer:
       init_args:
         save_dir: val_samples
         save_compressed: true
+    - class_path: src.callbacks.medical_visualization.MedicalVisualizationCallback
+      init_args:
+        every_n_steps: 5000
+        num_samples: 8
+        num_sampling_steps: 50
+        cfg_scale: 2.0
+        save_dir: training_vis
   plugins:
     - src.plugins.bd_env.BDEnvironment
 
@@ -123,7 +131,7 @@ data:
       resolution: 256
       max_num_instances: 1000
       noise_scale: 1.0
-  train_batch_size: 16
+  train_batch_size: 96
   train_num_workers: 4
-  pred_batch_size: 16
+  pred_batch_size: 32
   pred_num_workers: 1

code/configs_medical/PixelGen_Medical_Kvasir_B8_Spatial.yaml

@@ -1,5 +1,6 @@
 # PixelGen Medical - Kvasir-SEG Polyp Segmentation - Binary mask-conditional colonoscopy image generation
 # Model: JiTMedical-B/8, mask_mode=spatial
+# Optimized for 2x H800 80GB
 seed_everything: 1234
 tags:
   exp: &exp PixelGen_Medical_Kvasir_B8_Spatial
@@ -34,6 +35,13 @@ trainer:
       init_args:
         save_dir: val_samples
         save_compressed: true
+    - class_path: src.callbacks.medical_visualization.MedicalVisualizationCallback
+      init_args:
+        every_n_steps: 5000
+        num_samples: 8
+        num_sampling_steps: 50
+        cfg_scale: 2.0
+        save_dir: training_vis
   plugins:
     - src.plugins.bd_env.BDEnvironment
 
@@ -123,7 +131,7 @@ data:
       resolution: 256
       max_num_instances: 1000
       noise_scale: 1.0
-  train_batch_size: 16
+  train_batch_size: 32
   train_num_workers: 4
-  pred_batch_size: 16
+  pred_batch_size: 32
   pred_num_workers: 1

code/configs_medical/PixelGen_Medical_Kvasir_S8_Spatial.yaml

@@ -1,5 +1,6 @@
 # PixelGen Medical - Kvasir-SEG Polyp Segmentation - Binary mask-conditional colonoscopy image generation
 # Model: JiTMedical-S/8, mask_mode=spatial
+# Optimized for 2x H800 80GB
 seed_everything: 1234
 tags:
   exp: &exp PixelGen_Medical_Kvasir_S8_Spatial
@@ -34,6 +35,13 @@ trainer:
       init_args:
         save_dir: val_samples
         save_compressed: true
+    - class_path: src.callbacks.medical_visualization.MedicalVisualizationCallback
+      init_args:
+        every_n_steps: 5000
+        num_samples: 8
+        num_sampling_steps: 50
+        cfg_scale: 2.0
+        save_dir: training_vis
   plugins:
     - src.plugins.bd_env.BDEnvironment
 
@@ -123,7 +131,7 @@ data:
       resolution: 256
       max_num_instances: 1000
       noise_scale: 1.0
-  train_batch_size: 16
+  train_batch_size: 64
   train_num_workers: 4
-  pred_batch_size: 16
+  pred_batch_size: 32
   pred_num_workers: 1

code/configs_medical/PixelGen_Medical_OCTA500_B16_CrossAttn.yaml

@@ -1,5 +1,6 @@
 # PixelGen Medical - OCTA500 OCT Retinal Layer Segmentation - 6-class mask-conditional image generation
 # Model: JiTMedical-B/16, mask_mode=cross_attention
+# Optimized for 2x H800 80GB
 seed_everything: 1234
 tags:
   exp: &exp PixelGen_Medical_OCTA500_B16_CrossAttn
@@ -34,6 +35,13 @@ trainer:
       init_args:
         save_dir: val_samples
         save_compressed: true
+    - class_path: src.callbacks.medical_visualization.MedicalVisualizationCallback
+      init_args:
+        every_n_steps: 5000
+        num_samples: 8
+        num_sampling_steps: 50
+        cfg_scale: 2.0
+        save_dir: training_vis
   plugins:
     - src.plugins.bd_env.BDEnvironment
 
@@ -123,7 +131,7 @@ data:
       resolution: 256
       max_num_instances: 5000
       noise_scale: 1.0
-  train_batch_size: 32
+  train_batch_size: 96
   train_num_workers: 4
   pred_batch_size: 32
   pred_num_workers: 1

code/configs_medical/PixelGen_Medical_OCTA500_L16_Global.yaml

@@ -1,5 +1,6 @@
 # PixelGen Medical - OCTA500 OCT Retinal Layer Segmentation - 6-class mask-conditional image generation
 # Model: JiTMedical-L/16, mask_mode=global
+# Optimized for 2x H800 80GB
 seed_everything: 1234
 tags:
   exp: &exp PixelGen_Medical_OCTA500_L16_Global
@@ -34,6 +35,13 @@ trainer:
       init_args:
         save_dir: val_samples
         save_compressed: true
+    - class_path: src.callbacks.medical_visualization.MedicalVisualizationCallback
+      init_args:
+        every_n_steps: 5000
+        num_samples: 8
+        num_sampling_steps: 50
+        cfg_scale: 2.0
+        save_dir: training_vis
   plugins:
     - src.plugins.bd_env.BDEnvironment
 
@@ -123,7 +131,7 @@ data:
       resolution: 256
       max_num_instances: 5000
       noise_scale: 1.0
-  train_batch_size: 8
+  train_batch_size: 48
   train_num_workers: 4
-  pred_batch_size: 8
+  pred_batch_size: 32
   pred_num_workers: 1

code/configs_medical/PixelGen_Medical_OCTA500_XL16_Global.yaml

@@ -1,5 +1,6 @@
 # PixelGen Medical - OCTA500 OCT Retinal Layer Segmentation - 6-class mask-conditional image generation
 # Model: JiTMedical-XL/16, mask_mode=global
+# Optimized for 2x H800 80GB
 seed_everything: 1234
 tags:
   exp: &exp PixelGen_Medical_OCTA500_XL16_Global
@@ -34,6 +35,13 @@ trainer:
       init_args:
         save_dir: val_samples
         save_compressed: true
+    - class_path: src.callbacks.medical_visualization.MedicalVisualizationCallback
+      init_args:
+        every_n_steps: 5000
+        num_samples: 8
+        num_sampling_steps: 50
+        cfg_scale: 2.0
+        save_dir: training_vis
   plugins:
     - src.plugins.bd_env.BDEnvironment
 
@@ -123,7 +131,7 @@ data:
       resolution: 256
       max_num_instances: 5000
       noise_scale: 1.0
-  train_batch_size: 8
+  train_batch_size: 32
   train_num_workers: 4
-  pred_batch_size: 8
+  pred_batch_size: 32
   pred_num_workers: 1

code/configs_medical/PixelGen_Medical_REFUGE2_B16_CrossAttn.yaml

@@ -1,5 +1,6 @@
 # PixelGen Medical - REFUGE2 Optic Disc/Cup Segmentation - 3-class mask-conditional fundus image generation
 # Model: JiTMedical-B/16, mask_mode=cross_attention
+# Optimized for 2x H800 80GB
 seed_everything: 1234
 tags:
   exp: &exp PixelGen_Medical_REFUGE2_B16_CrossAttn
@@ -34,6 +35,13 @@ trainer:
       init_args:
         save_dir: val_samples
         save_compressed: true
+    - class_path: src.callbacks.medical_visualization.MedicalVisualizationCallback
+      init_args:
+        every_n_steps: 5000
+        num_samples: 8
+        num_sampling_steps: 50
+        cfg_scale: 2.0
+        save_dir: training_vis
   plugins:
     - src.plugins.bd_env.BDEnvironment
 
@@ -125,7 +133,7 @@ data:
       resolution: 256
       max_num_instances: 1000
       noise_scale: 1.0
-  train_batch_size: 16
+  train_batch_size: 96
   train_num_workers: 4
-  pred_batch_size: 16
+  pred_batch_size: 32
   pred_num_workers: 1

code/configs_medical/PixelGen_Medical_REFUGE2_B8_Spatial.yaml

@@ -1,5 +1,6 @@
 # PixelGen Medical - REFUGE2 Optic Disc/Cup Segmentation - 3-class mask-conditional fundus image generation
 # Model: JiTMedical-B/8, mask_mode=spatial
+# Optimized for 2x H800 80GB
 seed_everything: 1234
 tags:
   exp: &exp PixelGen_Medical_REFUGE2_B8_Spatial
@@ -34,6 +35,13 @@ trainer:
       init_args:
         save_dir: val_samples
         save_compressed: true
+    - class_path: src.callbacks.medical_visualization.MedicalVisualizationCallback
+      init_args:
+        every_n_steps: 5000
+        num_samples: 8
+        num_sampling_steps: 50
+        cfg_scale: 2.0
+        save_dir: training_vis
   plugins:
     - src.plugins.bd_env.BDEnvironment
 
@@ -125,7 +133,7 @@ data:
       resolution: 256
       max_num_instances: 1000
       noise_scale: 1.0
-  train_batch_size: 16
+  train_batch_size: 32
   train_num_workers: 4
-  pred_batch_size: 16
+  pred_batch_size: 32
   pred_num_workers: 1

code/configs_medical/PixelGen_Medical_REFUGE2_S8_Spatial.yaml

@@ -1,5 +1,6 @@
 # PixelGen Medical - REFUGE2 Optic Disc/Cup Segmentation - 3-class mask-conditional fundus image generation
 # Model: JiTMedical-S/8, mask_mode=spatial
+# Optimized for 2x H800 80GB
 seed_everything: 1234
 tags:
   exp: &exp PixelGen_Medical_REFUGE2_S8_Spatial
@@ -34,6 +35,13 @@ trainer:
       init_args:
         save_dir: val_samples
         save_compressed: true
+    - class_path: src.callbacks.medical_visualization.MedicalVisualizationCallback
+      init_args:
+        every_n_steps: 5000
+        num_samples: 8
+        num_sampling_steps: 50
+        cfg_scale: 2.0
+        save_dir: training_vis
   plugins:
     - src.plugins.bd_env.BDEnvironment
 
@@ -125,7 +133,7 @@ data:
       resolution: 256
       max_num_instances: 1000
       noise_scale: 1.0
-  train_batch_size: 16
+  train_batch_size: 64
   train_num_workers: 4
-  pred_batch_size: 16
+  pred_batch_size: 32
   pred_num_workers: 1

code/run_ablation.sh

@@ -1,59 +1,60 @@
 #!/bin/bash
 # PixelGen Medical - Ablation Study Training Commands
-# All experiments: 100k steps, 8 GPUs
+# All experiments: 100k steps, 2x H800 80GB
 # Organized by dataset, then by experiment type
+# Visualization saved to training_vis/ every 5000 steps
 
 # ═══════════════════════════════════════════════════════════════
 # CVC-ClinicDB (550 train images, binary polyp masks)
 # ═══════════════════════════════════════════════════════════════
 
-# CVC - S/8 Spatial (39M params)
-torchrun --nproc_per_node=8 main.py fit --config configs_medical/PixelGen_Medical_CVC_S8_Spatial.yaml
+# CVC - S/8 Spatial (39M, bs=64/gpu)
+torchrun --nproc_per_node=2 main.py fit --config configs_medical/PixelGen_Medical_CVC_S8_Spatial.yaml
 
-# CVC - B/8 Spatial (131M params)
-torchrun --nproc_per_node=8 main.py fit --config configs_medical/PixelGen_Medical_CVC_B8_Spatial.yaml
+# CVC - B/8 Spatial (131M, bs=32/gpu)
+torchrun --nproc_per_node=2 main.py fit --config configs_medical/PixelGen_Medical_CVC_B8_Spatial.yaml
 
-# CVC - B/16 CrossAttn (159M params)
-torchrun --nproc_per_node=8 main.py fit --config configs_medical/PixelGen_Medical_CVC_B16_CrossAttn.yaml
+# CVC - B/16 CrossAttn (159M, bs=96/gpu)
+torchrun --nproc_per_node=2 main.py fit --config configs_medical/PixelGen_Medical_CVC_B16_CrossAttn.yaml
 
 # ═══════════════════════════════════════════════════════════════
 # Kvasir-SEG (900 train images, binary polyp masks)
 # ═══════════════════════════════════════════════════════════════
 
-# Kvasir - S/8 Spatial (39M params)
-torchrun --nproc_per_node=8 main.py fit --config configs_medical/PixelGen_Medical_Kvasir_S8_Spatial.yaml
+# Kvasir - S/8 Spatial (39M, bs=64/gpu)
+torchrun --nproc_per_node=2 main.py fit --config configs_medical/PixelGen_Medical_Kvasir_S8_Spatial.yaml
 
-# Kvasir - B/8 Spatial (131M params)
-torchrun --nproc_per_node=8 main.py fit --config configs_medical/PixelGen_Medical_Kvasir_B8_Spatial.yaml
+# Kvasir - B/8 Spatial (131M, bs=32/gpu)
+torchrun --nproc_per_node=2 main.py fit --config configs_medical/PixelGen_Medical_Kvasir_B8_Spatial.yaml
 
-# Kvasir - B/16 CrossAttn (159M params)
-torchrun --nproc_per_node=8 main.py fit --config configs_medical/PixelGen_Medical_Kvasir_B16_CrossAttn.yaml
+# Kvasir - B/16 CrossAttn (159M, bs=96/gpu)
+torchrun --nproc_per_node=2 main.py fit --config configs_medical/PixelGen_Medical_Kvasir_B16_CrossAttn.yaml
 
 # ═══════════════════════════════════════════════════════════════
 # REFUGE2 (720 train images, 3-class fundus masks)
 # ═══════════════════════════════════════════════════════════════
 
-# REFUGE2 - S/8 Spatial (39M params)
-torchrun --nproc_per_node=8 main.py fit --config configs_medical/PixelGen_Medical_REFUGE2_S8_Spatial.yaml
+# REFUGE2 - S/8 Spatial (39M, bs=64/gpu)
+torchrun --nproc_per_node=2 main.py fit --config configs_medical/PixelGen_Medical_REFUGE2_S8_Spatial.yaml
 
-# REFUGE2 - B/8 Spatial (131M params)
-torchrun --nproc_per_node=8 main.py fit --config configs_medical/PixelGen_Medical_REFUGE2_B8_Spatial.yaml
+# REFUGE2 - B/8 Spatial (131M, bs=32/gpu)
+torchrun --nproc_per_node=2 main.py fit --config configs_medical/PixelGen_Medical_REFUGE2_B8_Spatial.yaml
 
-# REFUGE2 - B/16 CrossAttn (159M params)
-torchrun --nproc_per_node=8 main.py fit --config configs_medical/PixelGen_Medical_REFUGE2_B16_CrossAttn.yaml
+# REFUGE2 - B/16 CrossAttn (159M, bs=96/gpu)
+torchrun --nproc_per_node=2 main.py fit --config configs_medical/PixelGen_Medical_REFUGE2_B16_CrossAttn.yaml
 
 # ═══════════════════════════════════════════════════════════════
 # OCTA500 (108k train images, 6-class layer masks)
 # ═══════════════════════════════════════════════════════════════
 
-# OCTA500 - B/16 CrossAttn (159M params)
-torchrun --nproc_per_node=8 main.py fit --config configs_medical/PixelGen_Medical_OCTA500_B16_CrossAttn.yaml
+# OCTA500 - B/16 CrossAttn (159M, bs=96/gpu)
+torchrun --nproc_per_node=2 main.py fit --config configs_medical/PixelGen_Medical_OCTA500_B16_CrossAttn.yaml
 
-# OCTA500 - L/16 Global (458M params)
-torchrun --nproc_per_node=8 main.py fit --config configs_medical/PixelGen_Medical_OCTA500_L16_Global.yaml
+# OCTA500 - L/16 Global (458M, bs=48/gpu)
+torchrun --nproc_per_node=2 main.py fit --config configs_medical/PixelGen_Medical_OCTA500_L16_Global.yaml
 
-# OCTA500 - XL/16 Global (676M params)
-torchrun --nproc_per_node=8 main.py fit --config configs_medical/PixelGen_Medical_OCTA500_XL16_Global.yaml
+# OCTA500 - XL/16 Global (676M, bs=32/gpu)
+torchrun --nproc_per_node=2 main.py fit --config configs_medical/PixelGen_Medical_OCTA500_XL16_Global.yaml
 
 # ═══════════════════════════════════════════════════════════════
 # Evaluation (FID / Precision / Recall)
@@ -62,21 +63,26 @@ torchrun --nproc_per_node=8 main.py fit --config configs_medical/PixelGen_Medica
 
 # CVC evaluations
 CUDA_VISIBLE_DEVICES=0 python scripts/evaluate_medical.py --dataset cvc --model S/8 --mask_mode spatial --ckpt medical_workdirs/exp_PixelGen_Medical_CVC_S8_Spatial/PATH_TO_CKPT --cfg
+
 CUDA_VISIBLE_DEVICES=0 python scripts/evaluate_medical.py --dataset cvc --model B/8 --mask_mode spatial --ckpt medical_workdirs/exp_PixelGen_Medical_CVC_B8_Spatial/PATH_TO_CKPT --cfg
+
 CUDA_VISIBLE_DEVICES=0 python scripts/evaluate_medical.py --dataset cvc --model B/16 --mask_mode cross_attention --ckpt medical_workdirs/exp_PixelGen_Medical_CVC_B16_CrossAttn/PATH_TO_CKPT --cfg
 
 # Kvasir evaluations
 CUDA_VISIBLE_DEVICES=0 python scripts/evaluate_medical.py --dataset kvasir --model S/8 --mask_mode spatial --ckpt medical_workdirs/exp_PixelGen_Medical_Kvasir_S8_Spatial/PATH_TO_CKPT --cfg
+
 CUDA_VISIBLE_DEVICES=0 python scripts/evaluate_medical.py --dataset kvasir --model B/8 --mask_mode spatial --ckpt medical_workdirs/exp_PixelGen_Medical_Kvasir_B8_Spatial/PATH_TO_CKPT --cfg
+
 CUDA_VISIBLE_DEVICES=0 python scripts/evaluate_medical.py --dataset kvasir --model B/16 --mask_mode cross_attention --ckpt medical_workdirs/exp_PixelGen_Medical_Kvasir_B16_CrossAttn/PATH_TO_CKPT --cfg
 
 # REFUGE2 evaluations
 CUDA_VISIBLE_DEVICES=0 python scripts/evaluate_medical.py --dataset refuge2 --model S/8 --mask_mode spatial --ckpt medical_workdirs/exp_PixelGen_Medical_REFUGE2_S8_Spatial/PATH_TO_CKPT --cfg
+
 CUDA_VISIBLE_DEVICES=0 python scripts/evaluate_medical.py --dataset refuge2 --model B/8 --mask_mode spatial --ckpt medical_workdirs/exp_PixelGen_Medical_REFUGE2_B8_Spatial/PATH_TO_CKPT --cfg
+
 CUDA_VISIBLE_DEVICES=0 python scripts/evaluate_medical.py --dataset refuge2 --model B/16 --mask_mode cross_attention --ckpt medical_workdirs/exp_PixelGen_Medical_REFUGE2_B16_CrossAttn/PATH_TO_CKPT --cfg
 
-# OCTA500 evaluations (evaluate_medical.py does not yet support OCTA500,
-# use Lightning predict mode or write custom eval script)
-# torchrun --nproc_per_node=8 main.py predict --config configs_medical/PixelGen_Medical_OCTA500_B16_CrossAttn.yaml --ckpt_path PATH_TO_CKPT
-# torchrun --nproc_per_node=8 main.py predict --config configs_medical/PixelGen_Medical_OCTA500_L16_Global.yaml --ckpt_path PATH_TO_CKPT
-# torchrun --nproc_per_node=8 main.py predict --config configs_medical/PixelGen_Medical_OCTA500_XL16_Global.yaml --ckpt_path PATH_TO_CKPT
+# OCTA500 evaluations (evaluate_medical.py does not yet support OCTA500)
+# torchrun --nproc_per_node=2 main.py predict --config configs_medical/PixelGen_Medical_OCTA500_B16_CrossAttn.yaml --ckpt_path PATH_TO_CKPT
+# torchrun --nproc_per_node=2 main.py predict --config configs_medical/PixelGen_Medical_OCTA500_L16_Global.yaml --ckpt_path PATH_TO_CKPT
+# torchrun --nproc_per_node=2 main.py predict --config configs_medical/PixelGen_Medical_OCTA500_XL16_Global.yaml --ckpt_path PATH_TO_CKPT

code/src/callbacks/medical_visualization.py

@@ -0,0 +1,190 @@
+# Medical Visualization Callback
+# Saves mask + generated image grids during training for visual quality monitoring
+
+import os
+import torch
+import numpy as np
+from PIL import Image, ImageDraw, ImageFont
+import lightning.pytorch as pl
+from lightning.pytorch import Callback
+from lightning_utilities.core.rank_zero import rank_zero_info
+
+
+class MedicalVisualizationCallback(Callback):
+    """
+    Periodically generates and saves visualization grids during training.
+
+    Each grid shows: [Mask | Generated (CFG) | Generated (No-CFG)]
+    for a fixed set of masks, allowing visual comparison across training.
+    """
+
+    def __init__(
+        self,
+        every_n_steps: int = 5000,
+        num_samples: int = 8,
+        num_sampling_steps: int = 50,
+        cfg_scale: float = 2.0,
+        save_dir: str = "training_vis",
+        t_eps: float = 0.05,
+    ):
+        super().__init__()
+        self.every_n_steps = every_n_steps
+        self.num_samples = num_samples
+        self.num_sampling_steps = num_sampling_steps
+        self.cfg_scale = cfg_scale
+        self.save_dir = save_dir
+        self.t_eps = t_eps
+        self._fixed_noise = None
+        self._fixed_masks = None
+
+    def _shift_respace_fn(self, t, shift=1.0):
+        return t / (t + (1 - t) * shift)
+
+    @torch.no_grad()
+    def _sample(self, model, noise, mask, cfg_scale=None):
+        """Euler sampling with optional CFG, directly passing mask to model."""
+        bs = noise.shape[0]
+        timesteps = torch.linspace(0.0, 1 - 1.0 / self.num_sampling_steps, self.num_sampling_steps)
+        timesteps = torch.cat([timesteps, torch.tensor([1.0])], dim=0)
+        timesteps = self._shift_respace_fn(timesteps, 1.0).to(noise.device)
+        y = torch.zeros(bs, dtype=torch.long, device=noise.device)
+        x = noise
+
+        for i in range(len(timesteps) - 1):
+            t_cur = timesteps[i]
+            t_next = timesteps[i + 1]
+            dt = t_next - t_cur
+            t_batch = t_cur.repeat(bs)
+
+            if cfg_scale is not None and cfg_scale > 1.0:
+                # CFG: concat unconditional + conditional
+                cfg_x = torch.cat([x, x], dim=0)
+                cfg_t = t_batch.repeat(2)
+                cfg_y = torch.cat([y, y], dim=0)
+                cfg_mask = torch.cat([torch.zeros_like(mask), mask], dim=0)
+                pred = model(cfg_x, cfg_t, cfg_y, mask=cfg_mask)
+                pred_v = (pred - cfg_x) / (1.0 - cfg_t.view(-1, 1, 1, 1)).clamp_min(self.t_eps)
+                v_uncond, v_cond = pred_v.chunk(2)
+                v = v_uncond + cfg_scale * (v_cond - v_uncond)
+            else:
+                # No CFG
+                pred = model(x, t_batch, y, mask=mask)
+                v = (pred - x) / (1.0 - t_batch.view(-1, 1, 1, 1)).clamp_min(self.t_eps)
+
+            x = x + v * dt
+
+        return x.clamp(-1, 1) * 0.5 + 0.5  # [-1,1] -> [0,1]
+
+    def _mask_to_rgb(self, mask_tensor):
+        """Convert single-channel mask [1, H, W] to RGB [H, W, 3] uint8."""
+        m = mask_tensor[0].cpu().numpy()
+        unique_vals = np.unique(m)
+
+        if len(unique_vals) <= 3:
+            # Multi-class: colorize (e.g., REFUGE2: 0=black, ~0.5=blue, ~1.0=red)
+            rgb = np.zeros((*m.shape, 3), dtype=np.uint8)
+            rgb[m < 0.1] = [0, 0, 0]           # background
+            rgb[(m >= 0.1) & (m < 0.7)] = [0, 120, 255]  # class 1 (blue)
+            rgb[m >= 0.7] = [255, 60, 60]       # class 2 (red)
+        else:
+            # Binary or continuous: grayscale
+            m_uint8 = (m * 255).astype(np.uint8)
+            rgb = np.stack([m_uint8] * 3, axis=-1)
+
+        return rgb
+
+    def _make_grid(self, masks, gen_cfg, gen_nocfg, step):
+        """Create visualization grid: Mask | CFG | No-CFG."""
+        n = masks.shape[0]
+        h, w = masks.shape[2], masks.shape[3]
+        pad = 4
+        col_labels = ["Mask", f"CFG={self.cfg_scale}", "No-CFG"]
+        n_cols = len(col_labels)
+        header_h = 30
+
+        canvas_w = n_cols * w + (n_cols + 1) * pad
+        canvas_h = n * h + (n + 1) * pad + header_h
+        canvas = np.ones((canvas_h, canvas_w, 3), dtype=np.uint8) * 40
+
+        # Header
+        try:
+            font = ImageFont.truetype("/usr/share/fonts/truetype/dejavu/DejaVuSans-Bold.ttf", 16)
+        except (IOError, OSError):
+            font = ImageFont.load_default()
+
+        img_pil = Image.fromarray(canvas)
+        draw = ImageDraw.Draw(img_pil)
+        for col_idx, label in enumerate(col_labels):
+            x_pos = pad + col_idx * (w + pad) + w // 2
+            draw.text((x_pos, 6), label, fill=(255, 255, 255), font=font, anchor="mt")
+        # Step info
+        draw.text((canvas_w - 10, 6), f"step={step}", fill=(180, 180, 180), font=font, anchor="rt")
+        canvas = np.array(img_pil)
+
+        for row in range(n):
+            y_pos = header_h + pad + row * (h + pad)
+
+            # Mask column
+            mask_rgb = self._mask_to_rgb(masks[row])
+            x_pos = pad
+            canvas[y_pos:y_pos + h, x_pos:x_pos + w] = mask_rgb
+
+            # CFG generated
+            img_cfg = (gen_cfg[row].permute(1, 2, 0).cpu().numpy() * 255).clip(0, 255).astype(np.uint8)
+            x_pos = pad + (w + pad)
+            canvas[y_pos:y_pos + h, x_pos:x_pos + w] = img_cfg
+
+            # No-CFG generated
+            img_nocfg = (gen_nocfg[row].permute(1, 2, 0).cpu().numpy() * 255).clip(0, 255).astype(np.uint8)
+            x_pos = pad + 2 * (w + pad)
+            canvas[y_pos:y_pos + h, x_pos:x_pos + w] = img_nocfg
+
+        return canvas
+
+    def on_train_batch_end(self, trainer, pl_module, outputs, batch, batch_idx):
+        step = trainer.global_step
+        if step == 0 or step % self.every_n_steps != 0:
+            return
+        if not trainer.is_global_zero:
+            return
+
+        # Initialize fixed noise and masks from first validation batch
+        if self._fixed_masks is None:
+            eval_dl = trainer.datamodule.val_dataloader()
+            for eval_batch in eval_dl:
+                xT, y, metadata = eval_batch
+                mask = metadata.get('mask', None) if isinstance(metadata, dict) else None
+                if mask is None:
+                    rank_zero_info("[MedicalVis] No mask in eval batch, skipping visualization")
+                    return
+                n = min(self.num_samples, mask.shape[0])
+                self._fixed_masks = mask[:n].to(pl_module.device)
+                self._fixed_noise = torch.randn(
+                    n, 3, pl_module.denoiser.input_size, pl_module.denoiser.input_size,
+                    device=pl_module.device, generator=torch.Generator(device=pl_module.device).manual_seed(42)
+                )
+                break
+
+        if self._fixed_masks is None:
+            return
+
+        # Generate samples using EMA model
+        model = pl_module.ema_denoiser
+        was_training = model.training
+        model.eval()
+
+        noise = self._fixed_noise
+        masks = self._fixed_masks
+
+        gen_cfg = self._sample(model, noise, masks, cfg_scale=self.cfg_scale)
+        gen_nocfg = self._sample(model, noise, masks, cfg_scale=None)
+
+        if was_training:
+            model.train()
+
+        # Save grid
+        save_path = os.path.join(trainer.default_root_dir, self.save_dir)
+        os.makedirs(save_path, exist_ok=True)
+        grid = self._make_grid(masks, gen_cfg, gen_nocfg, step)
+        Image.fromarray(grid).save(os.path.join(save_path, f"vis_step_{step:06d}.png"))
+        rank_zero_info(f"[MedicalVis] Saved visualization at step {step}")

code/src/diffusion/base/sampling.py

@@ -24,8 +24,8 @@ class BaseSampler(nn.Module):
     def _impl_sampling(self, net, noise, condition, uncondition):
         raise NotImplementedError
 
-    def forward(self, net, noise, condition, uncondition, return_x_trajs=False, return_v_trajs=False):
-        x_trajs, v_trajs = self._impl_sampling(net, noise, condition, uncondition)
+    def forward(self, net, noise, condition, uncondition, return_x_trajs=False, return_v_trajs=False, **kwargs):
+        x_trajs, v_trajs = self._impl_sampling(net, noise, condition, uncondition, **kwargs)
         if return_x_trajs and return_v_trajs:
             return x_trajs[-1], x_trajs, v_trajs
         elif return_x_trajs:

code/src/lightning_model.py

@@ -147,11 +147,22 @@ class LightningModel(pl.LightningModule):
         with torch.no_grad():
             condition, uncondition = self.conditioner(y, metadata)
 
+        # Extract mask for direct conditioning (spatial/cross_attention modes)
+        mask = None
+        if isinstance(metadata, dict):
+            mask = metadata.get('mask', None)
+        elif isinstance(metadata, (list, tuple)):
+            masks = [m.get('mask', None) for m in metadata if isinstance(m, dict)]
+            if len(masks) > 0 and masks[0] is not None:
+                mask = torch.stack(masks, dim=0)
+        if mask is not None:
+            mask = mask.to(xT.device)
+
         # sample images
         if self.eval_original_model:
-            samples = self.diffusion_sampler(self.denoiser, xT, condition, uncondition)
+            samples = self.diffusion_sampler(self.denoiser, xT, condition, uncondition, mask=mask)
         else:
-            samples = self.diffusion_sampler(self.ema_denoiser, xT, condition, uncondition)
+            samples = self.diffusion_sampler(self.ema_denoiser, xT, condition, uncondition, mask=mask)
 
         samples = self.vae.decode(samples)
         # fp32 -1,1 -> uint8 0,255