Upload model.py

model.py

@@ -87,9 +87,21 @@ def _embed_image(image: Image.Image) -> torch.Tensor:
         dy = h // 8
         if dy > 0:
             image = image.crop((0, dy, w, h - dy))
-    inputs = _CLIP_PROCESSOR(images=[image.convert("RGB")], return_tensors="pt").to(_DEVICE)
+
+    # Newer HF processor outputs should be moved field-by-field for robustness.
+    inputs = _CLIP_PROCESSOR(images=[image.convert("RGB")], return_tensors="pt")
+    inputs = {k: v.to(_DEVICE) for k, v in inputs.items()}
+
     with torch.no_grad():
+        # Preferred API: projected CLIP image embeddings as a tensor.
         feats = _CLIP_MODEL.get_image_features(**inputs)
+
+        # Defensive fallback in case an HF-side change returns a structured output.
+        if hasattr(feats, "image_embeds"):
+            feats = feats.image_embeds
+        elif hasattr(feats, "pooler_output"):
+            feats = feats.pooler_output
+
         feats = feats / feats.norm(dim=-1, keepdim=True)
     return feats  # [1, d_in]