_process_saree_core_preserve_alpha UPDATED

app.py

@@ -71,7 +71,7 @@ def _process_saree_core_preserve_alpha(base_image: Image.Image, pattern_image: I
     model.eval()
 
     with torch.no_grad():
-        target_size = base_rgb.size[::-1]  # ✅ FIX: use base_rgb instead of img_rgb
+        target_size = base_rgb.size[::-1]
         depth_map = model(img_tensor.to(device), target_size=target_size)
         depth_map = depth_map.squeeze().cpu().numpy()
 
@@ -124,15 +124,55 @@ def _process_saree_core_preserve_alpha(base_image: Image.Image, pattern_image: I
     mask_blurred = cv2.GaussianBlur(mask_eroded, (15, 15), sigmaX=3, sigmaY=3)
     mask_final = mask_blurred.astype(np.float32) / 255.0
 
+    # ----------------------------------------
+    # 5.5 MAGIC-WAND STYLE CLEANUP for black background (connected to borders)
+    #     (Simulates Mac Preview's wand: threshold + keep only border-connected dark regions)
+    # ----------------------------------------
+    # Tune these two params if needed:
+    black_thresh = 30   # 0..255 – raise if some dark bg remains; lower if it eats shadows
+    feather_px  = 9     # odd kernel radius for blur/feathering
+
+    # Build a candidate black mask from the CURRENT composite (what will be visible)
+    preview_rgb_u8 = (pattern_folded * 255).astype(np.uint8)
+    gray = cv2.cvtColor(preview_rgb_u8, cv2.COLOR_RGB2GRAY)
+
+    # Binary mask of "near-black"
+    black_mask = cv2.threshold(gray, black_thresh, 255, cv2.THRESH_BINARY_INV)[1]
+
+    # Keep ONLY components touching image borders (likely the true background)
+    num_labels, labels = cv2.connectedComponents(black_mask)
+    h, w = gray.shape
+    border_labels = set(np.unique(np.concatenate([labels[0, :], labels[-1, :], labels[:, 0], labels[:, -1]])))
+    bg_touching = np.zeros_like(black_mask)
+    for lbl in border_labels:
+        if lbl == 0:
+            continue
+        bg_touching[labels == lbl] = 255
+
+    # Feather the background mask
+    if feather_px % 2 == 0:
+        feather_px += 1  # must be odd for GaussianBlur
+    bg_touching = cv2.GaussianBlur(bg_touching, (feather_px, feather_px), 0)
+
+    bg_touching_f = bg_touching.astype(np.float32) / 255.0  # 0..1
+
+    # Combine with the (refined) original alpha: remove black background regardless of original alpha
+    # old_alpha = mask_final  # (keeping for reference)
+    refined_alpha = np.clip(mask_final * (1.0 - bg_touching_f), 0.0, 1.0)
+
     # ----------------------------------------
     # 6. Apply mask to final pattern
     # ----------------------------------------
-    mask_stack = np.stack([mask_final] * 3, axis=-1)
+    mask_stack = np.stack([refined_alpha] * 3, axis=-1)
     pattern_final = pattern_folded * mask_stack
     pattern_rgb = (pattern_final * 255).astype(np.uint8)
-    alpha_channel = (mask_final * 255).astype(np.uint8)
-    pattern_rgba = np.dstack((pattern_rgb, alpha_channel))
 
+    # old line (kept, but not used now):
+    # alpha_channel = (mask_final * 255).astype(np.uint8)
+    # NEW alpha using magic-wand cleanup:
+    alpha_channel = (refined_alpha * 255).astype(np.uint8)
+
+    pattern_rgba = np.dstack((pattern_rgb, alpha_channel))
     return Image.fromarray(pattern_rgba, mode="RGBA")