Stabilize auto mask generator (disable multi-scale crops)

app.py

@@ -39,7 +39,7 @@ print(f"Loading MedSAM model ({MODEL_TYPE})...")
 
 try:
     torch.load = patched_torch_load
-    sam = sam_model_registry[MODEL_TYPE](checkpoint=MODEL_CHECKPOINT)
+sam = sam_model_registry[MODEL_TYPE](checkpoint=MODEL_CHECKPOINT)
 finally:
     torch.load = original_torch_load
 
@@ -51,15 +51,19 @@ predictor = SamPredictor(sam)
 print("✓ SamPredictor initialized for interactive segmentation")
 
 # SamAutomaticMaskGenerator for automatic mask generation
-# Uses the same model but with automatic grid-based prompting
+# Uses the same model but with automatic grid-based prompting.
+# NOTE: We disable multi-scale cropping (crop_n_layers=0) because it can
+# sometimes produce empty crop_boxes on certain images, which leads to
+# IndexError inside torchvision.boxes.box_area. Using a single-scale grid
+# is more stable for our use case.
 mask_generator = SamAutomaticMaskGenerator(
     model=sam,
-    points_per_side=32,  # Grid density (32x32 = 1024 points)
-    pred_iou_thresh=0.88,  # IoU threshold for filtering
-    stability_score_thresh=0.95,  # Stability threshold
-    crop_n_layers=1,  # Number of crop layers for multi-scale
-    crop_n_points_downscale_factor=2,  # Downscale factor for crops
-    min_mask_region_area=100  # Minimum mask area in pixels
+    points_per_side=32,            # Grid density (32x32 = 1024 points)
+    pred_iou_thresh=0.88,          # IoU threshold for filtering
+    stability_score_thresh=0.95,   # Stability threshold
+    crop_n_layers=0,               # Disable multi-scale crops to avoid IndexError
+    crop_n_points_downscale_factor=2,
+    min_mask_region_area=100       # Minimum mask area in pixels
 )
 print("✓ SamAutomaticMaskGenerator initialized for automatic segmentation")
 print("✓ MedSAM model loaded successfully!")