Addition conditioning — anchor frame vs captions

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

All six arms, n=50

video gap = L1(generation → ground-truth target) − L1(generation → the clip the model was given). Negative = the edit moved toward the target. Every arm used identical per-clip seeds (seed + crc32(video_id)), so these are paired comparisons.

What actually moved the needle

Two-sided exact sign test over the 50 paired per-clip differences.

Examples

Per clip: input (object absent) · the anchor frame itself, outlined · the three anchor-conditioned generations · ground truth. The captions fed to the model are printed above each row. Audio muted by default.