fix: pass file paths instead of PIL objects to ZeroGPU worker to avoid serialization TypeError

app.py

@@ -3,6 +3,7 @@ AD-Copilot Demo: Comparison-Aware Anomaly Detection with Vision-Language Model
 """
 
 import os
+import tempfile
 import traceback
 import spaces
 import gradio as gr
@@ -36,17 +37,18 @@ model = AutoModelForImageTextToText.from_pretrained(
 # ---------------------------------------------------------------------------
 @spaces.GPU(duration=120)
 def _predict_inner(
-    reference_image: Image.Image,
-    test_image: Image.Image,
+    ref_path: str,
+    test_path: str,
     prompt: str,
     max_new_tokens: int,
 ):
     with torch.inference_mode():
         max_new_tokens = int(max_new_tokens)
 
+        reference_image = Image.open(ref_path).convert("RGB")
+        test_image = Image.open(test_path).convert("RGB")
+
         # Resize long edge to 512 to save GPU memory / speed up
-        reference_image = reference_image.copy()
-        test_image = test_image.copy()
         reference_image.thumbnail((512, 512), Image.Resampling.LANCZOS)
         test_image.thumbnail((512, 512), Image.Resampling.LANCZOS)
 
@@ -97,7 +99,17 @@ def predict(
     if reference_image is None or test_image is None:
         return "Please upload both a reference (good) image and a test image."
     try:
-        return _predict_inner(reference_image, test_image, prompt, max_new_tokens)
+        # Save PIL images to temp files to avoid serialization issues with ZeroGPU
+        with tempfile.NamedTemporaryFile(suffix=".png", delete=False) as f:
+            reference_image.save(f, format="PNG")
+            ref_path = f.name
+        with tempfile.NamedTemporaryFile(suffix=".png", delete=False) as f:
+            test_image.save(f, format="PNG")
+            test_path = f.name
+        result = _predict_inner(ref_path, test_path, prompt, max_new_tokens)
+        os.unlink(ref_path)
+        os.unlink(test_path)
+        return result
     except Exception as e:
         tb = traceback.format_exc()
         print(tb, flush=True)