CRITICAL FIX: Remove redundant checkpoint loading that was corrupting model weights. build_sam3_image_model() already loads them correctly

app.py

@@ -86,33 +86,9 @@ processor = None
 if model_installed:
     print(f"Loading SAM 3 onto {device}...")
     model = build_sam3_image_model()
-    # ── SYSTEM RAM OOM PROTECTION ────────────────────────────────────
-    # HF Spaces strictly enforce a 16GB System RAM limit. Loading 6.9GB models 
-    # out of nested dictionary checkpoints will spike RAM past 20GB during copies!
-    # Aggressively force Python to garbage collect unpinned tensors instantly.
-    import gc
-    
-    state_dict = torch.load(ckpt_path, map_location="cpu", weights_only=True)
-    state_dict_inner = state_dict.pop("model") if "model" in state_dict else state_dict
-    
-    image_state_dict = {}
-    for k, v in state_dict_inner.items():
-        if k.startswith("detector."):
-            image_state_dict[k.replace("detector.", "", 1)] = v
-        else:
-            image_state_dict[k] = v
-            
-    # Purge the original 3.45GB raw checkpoint instantly
-    del state_dict
-    del state_dict_inner
-    gc.collect()
-    
-    model.load_state_dict(image_state_dict, strict=False)
-    
-    # Purge the 3.45GB filtered checkpoint array instantly
-    del image_state_dict
-    gc.collect()
-    
+    # NOTE: build_sam3_image_model() already loads checkpoint weights internally
+    # via _load_checkpoint(). DO NOT load the checkpoint again manually—it corrupts
+    # the weights with incorrect key mappings and causes zero masks!
     model.to(device)
     model.to(torch.float32)  # Lossless 32-bit execution prevents BFloat16 exponent overflows (NaNs)