feat(api): allow painted-on mask via mask_is_painted flag (auto-diff + Otsu threshold)

api/main.py

@@ -10,6 +10,7 @@ from fastapi import FastAPI, UploadFile, File, HTTPException, Depends, Header, R
 from fastapi.responses import FileResponse, JSONResponse
 from pydantic import BaseModel
 from PIL import Image
+import cv2
 
 from src.core import process_inpaint
 
@@ -179,12 +180,29 @@ def inpaint_multipart(
     mask: UploadFile = File(...),
     request: Request = None,
     invert_mask: bool = True,
+    mask_is_painted: bool = False,  # if True, mask file is the painted-on image (e.g., black strokes on original)
     _: None = Depends(bearer_auth),
 ) -> Dict[str, str]:
     # Load in-memory
     img = Image.open(image.file).convert("RGBA")
-    m = Image.open(mask.file)
-    mask_rgba = _load_rgba_mask_from_image(m)
+    m = Image.open(mask.file).convert("RGBA")
+
+    if mask_is_painted:
+        # Derive mask by differencing painted image vs original
+        img_rgb = cv2.cvtColor(np.array(img), cv2.COLOR_RGBA2RGB)
+        m_rgb = cv2.cvtColor(np.array(m), cv2.COLOR_RGBA2RGB)
+        diff = cv2.absdiff(img_rgb, m_rgb)
+        gray = cv2.cvtColor(diff, cv2.COLOR_RGB2GRAY)
+        # Otsu threshold for robustness; fallback threshold if Otsu fails
+        try:
+            _, binmask = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
+        except Exception:
+            _, binmask = cv2.threshold(gray, 40, 255, cv2.THRESH_BINARY)
+        # Build RGBA mask where selected area has alpha=0
+        mask_rgba = np.zeros((binmask.shape[0], binmask.shape[1], 4), dtype=np.uint8)
+        mask_rgba[:, :, 3] = np.where(binmask > 0, 0, 255).astype(np.uint8)
+    else:
+        mask_rgba = _load_rgba_mask_from_image(m)
 
     result = process_inpaint(np.array(img), mask_rgba, invert_mask=invert_mask)
     result_name = f"output_{uuid.uuid4().hex}.png"