Skip empty video kwargs for image-only inference

app.py

@@ -65,6 +65,17 @@ def strip_reasoning_output(text: str) -> str:
     return text
 
 
+def clean_video_kwargs(video_kwargs: dict) -> dict:
+    cleaned = {}
+    for key, value in video_kwargs.items():
+        if value is None:
+            continue
+        if isinstance(value, list) and len(value) == 0:
+            continue
+        cleaned[key] = value
+    return cleaned
+
+
 def ensure_model_loaded():
     global _MODEL, _PROCESSOR
     if _MODEL is not None and _PROCESSOR is not None:
@@ -155,9 +166,9 @@ def run_inference(
         "return_tensors": "pt",
         "padding": False,
     }
-    if video_inputs is not None:
+    if video_inputs:
         processor_kwargs["videos"] = video_inputs
-    processor_kwargs.update(video_kwargs)
+        processor_kwargs.update(clean_video_kwargs(video_kwargs))
 
     inputs = processor(**processor_kwargs)
     target_device = next(model.parameters()).device