Grounded examples (video-frame positive refs), tighten GPU duration to measured 30s

app.py

@@ -145,7 +145,7 @@ def highlight_tags(caption: str) -> str:
 # ---------------------------------------------------------------------------
 # Inference
 # ---------------------------------------------------------------------------
-@spaces.GPU(duration=120)
+@spaces.GPU(duration=30)
 def generate_caption(
     video,
     ref_1,
@@ -291,12 +291,12 @@ with gr.Blocks(theme=gr.themes.Citrus(), css=CSS) as demo:
 
         gr.Examples(
             examples=[
-                # border collie running through a barley field; refs: a golden-retriever-like dog
-                # (loose visual match the model may or may not ground) + a husky (clear distractor).
-                ["examples/border_collie_running.mp4", "examples/dog.jpg", "examples/husky_dog.jpg"],
-                # man jumping rope on a city sidewalk; refs: white sneakers (worn in the video)
-                # + a night city skyline (distractor — the video is daytime, no skyline).
-                ["examples/man_jumping_rope.mp4", "examples/sneakers_white_studio.jpg", "examples/city_skyline_night.jpg"],
+                # Border collie running through a field. <Image_1> is a frame of the
+                # same dog (grounded); <Image_2> is a husky (distractor, omitted).
+                ["examples/border_collie_running.mp4", "examples/collie_frame.jpg", "examples/husky_dog.jpg"],
+                # Man jumping rope on a sidewalk. <Image_1> is a frame of the same
+                # man (grounded); <Image_2> is a night skyline (distractor, omitted).
+                ["examples/man_jumping_rope.mp4", "examples/jump_rope_frame.jpg", "examples/city_skyline_night.jpg"],
             ],
             # Examples fill video + <Image_1> + <Image_2>; ref_3..ref_6 stay at their defaults.
             inputs=[video_in, r1, r2],