fix: use BiRefNet-Lite as default model and fix JSON parsing

- Switch default from birefnet-general (973MB, 4min) to
  birefnet-general-lite (faster, still SOTA quality)
- Fix Python script stdout pollution — progress messages now go to
  stderr so the JSON result parser doesn't break
- Pre-download birefnet-general-lite in Docker build

apps/web/src/components/tools/remove-bg-settings.tsx

@@ -34,7 +34,7 @@ export function RemoveBgSettings() {
   const { processFiles, processing, error, downloadUrl, originalSize, processedSize } =
     useToolProcessor("remove-background");
 
-  const [model, setModel] = useState<BgModel>("birefnet-general");
+  const [model, setModel] = useState<BgModel>("birefnet-general-lite");
   const [bgColor, setBgColor] = useState("");
   const [elapsed, setElapsed] = useState(0);
   const timerRef = useRef<ReturnType<typeof setInterval> | null>(null);

docker/Dockerfile

@@ -71,8 +71,8 @@ RUN /opt/venv/bin/pip install --no-cache-dir --upgrade pip && \
 # This makes the Docker image fully self-contained — works offline
 RUN /opt/venv/bin/python3 -c "\
 from rembg import new_session; \
-print('Downloading BiRefNet model (SOTA)...'); \
-new_session('birefnet-general'); \
+print('Downloading BiRefNet-Lite model (SOTA, fast)...'); \
+new_session('birefnet-general-lite'); \
 print('Downloading u2net model (fallback)...'); \
 new_session('u2net'); \
 print('Background removal models ready') \

packages/ai/python/remove_bg.py

@@ -8,25 +8,26 @@ def main():
     output_path = sys.argv[2]
     settings = json.loads(sys.argv[3]) if len(sys.argv) > 3 else {}
 
-    model = settings.get("model", "birefnet-general")
+    model = settings.get("model", "birefnet-general-lite")
     bg_color = settings.get("backgroundColor", "")
 
     try:
         from rembg import remove, new_session
-        from PIL import Image
         import io
 
-        print(json.dumps({"progress": "loading_model"}), flush=True)
+        # Progress messages go to stderr (stdout reserved for JSON result)
+        sys.stderr.write(f"Loading model: {model}\n")
+        sys.stderr.flush()
 
-        # Create a session with the selected model
         session = new_session(model)
 
+        sys.stderr.write("Processing image...\n")
+        sys.stderr.flush()
+
         with open(input_path, "rb") as f:
             input_data = f.read()
 
-        print(json.dumps({"progress": "processing"}), flush=True)
-
-        # Try with alpha matting first for better edges
+        # Try with alpha matting for better edges, fall back without
         try:
             output_data = remove(
                 input_data,
@@ -40,6 +41,8 @@ def main():
 
         # If a background color is specified, composite onto it
         if bg_color and bg_color.startswith("#"):
+            from PIL import Image
+
             img = Image.open(io.BytesIO(output_data)).convert("RGBA")
             hex_color = bg_color.lstrip("#")
             r = int(hex_color[0:2], 16)