Update app.py

app.py

@@ -241,20 +241,23 @@ def outpaint(image, width, height, overlap_percentage, num_inference_steps, resi
     if not can_expand(background.width, background.height, width, height, alignment):
         alignment = "Middle"
 
-    # Extract original alpha from the input image
+    # Original alpha is 0 where transparent (or 255 where opaque)
     original_alpha = background.split()[3] if background.mode == "RGBA" else Image.new("L", background.size, 255)
-    combined_mask = ImageChops.logical_or(original_alpha, mask)  # 0 where to paint
-    # Combine original and generated mask
-    # combined_mask = ImageChops.logical_or(original_alpha_mask, mask)  # 0 where to fill
-    cnet_image = background.copy()
+    
+    # Generated mask is 0 where you want to paint (from image["layers"][0])
+    # If "layers"[0] is not alpha, convert it to L-mode
+    mask = mask.convert("L") if mask.mode != "L" else mask
+    
+    # Combine the two masks: 0 where either the original is transparent OR the generated mask is active
+    combined_mask = ImageChops.logical_or(original_alpha, mask)
+    # After generating the combined_mask
     if cnet_image.mode == "RGBA":
-        # Extract the alpha channel and apply the combined mask
-        alpha = combined_mask.convert("L")
-        cnet_image = cnet_image.convert("RGBA")
-        cnet_image.putalpha(alpha)  # Replace the alpha channel
+        # Create a new alpha channel based on the mask
+        new_alpha = combined_mask.convert("L")
+        cnet_image = Image.merge("RGBA", cnet_image.split()[:3] + (new_alpha,))
     else:
-        # If no alpha, just use the background
-        cnet_image = background
+        # If no alpha, you can't use the mask — just use the background as is
+        cnet_image = background.copy()
     # Use the combined_mask in the pipeline
     #cnet_image = background.copy()
     # cnet_image.paste(0, (0, 0), combined_mask)  # Overlay black on combined_mask area
@@ -276,12 +279,11 @@ def outpaint(image, width, height, overlap_percentage, num_inference_steps, resi
         num_inference_steps=num_inference_steps
     ):
         yield cnet_image, image
-    # Invert the combined_mask and paste the generated image back
-    filled_mask = combined_mask.point(lambda p: 255 - p)
-    # image = image.convert("RGBA")
-    # cnet_image.paste(image, (0, 0), filled_mask)
     image = image.convert("RGBA")
-    cnet_image.paste(image, (0, 0), combined_mask)
+    # If the pipeline uses alpha as the mask, invert it to get 255 where to paint
+    filled_mask = combined_mask.point(lambda p: 255 - p)
+    # Paste the result using the mask to preserve transparency
+    cnet_image.paste(image, (0, 0), filled_mask)  # Only modify areas where filled_mask is 255 (i.e., original transparent + generated mask area)
     yield background, cnet_image
     
 @spaces.GPU(duration=7)