Conditioning arms for audio-visual object editing

Every arm except the bare noun is an oracle. The anchor frame is the ground-truth target's first frame; the captions were written by Qwen3-VL-8B looking at that same target clip. They bound how much conditioning could help — they are not a deployable setup and not comparable to any baseline.

All arms, n=30

gap = L1(generation → ground-truth target) − L1(generation → the clip the model was given). Negative = the edit moved toward the target. Identical per-clip seeds across every arm, so these are paired.

Paired sign tests

Two-sided exact sign test over the 30 paired per-clip differences. Bold p = significant at 0.05.

Examples

Per direction: input · anchor frame (outlined, a still — what --prepend-ref prepends) · the four arms · ground truth. The caption fed to the caption arms is printed above. Audio muted by default.