Videos
The full supplementary video and the baseline comparison reel.
Results on LTX-2
Hit Spark the motion! on any card to apply its edit.
Generalization to H3
Percentages are from a human study of 25 scenarios with 15 participants (375 judgements): on each card, how often ours was preferred, judged more natural, and judged the stronger edit, against the same for H3 alone. Ties make up the rest. These clips carry sound.
Audio, text, or both
The audio latent alone, the text residual alone, and the two together, beside H3 with no editing. The reel opens with the human study over all 12 scenarios (15 participants), then shows the 5 where the combination won clearly.
Baseline comparison
Twelve scenarios against seven baselines. Pick a scenario to compare. In a study of 21 participants, our result was ranked first in 56.3% of trials, against 21.4% for the next best method.
Motion transfer
The audio latent and text residual tuned on one clip, applied to another. Hit Transfer the motion! to carry them across.
Where each signal attends
Average cross-attention for the two tuned signals, over the first frame of the result. A2V is audio-to-video, T2V is text-to-video; switch between them on any card.
Ablations
Every variant of a scenario plays at once. Each reel replays its first scenario at half speed.
Rebuttal Results
Additional results referenced in our responses to the reviewers. All clips play muted; use the player controls to turn sound on. Each card uses the input clip’s soundtrack.
IVEBench: subject-motion edits
Edits from IVEBench’s subject-motion category, made with our method and with prompt-only LTX-2 Retake. The plot shows the official IVEBench metrics; each card shows the input and our result for the instruction above it.
Mean over items for our method, prompt-only Retake and the unedited input. The highlighted axes are the Compliance dimension, which measures whether the instructed motion appears.
Input scores 1.0 on semantic fidelity by construction, since it is compared with itself.
Best-of-8 vs. ours
Prompt-only LTX-2 Retake rendered with 8 seeds per scenario; the sample that the same critic scores highest is shown next to our result. The note under the best-of-8 column describes what that video shows.
Noise optimization vs. ours
Gradient-based optimization of the initial noise of the Retake segment, with the same critic and the audio and text conditioning frozen, shown for each learning rate tried. The default is the learning rate with the highest critic score in its own optimization (marked with a dot); ours is the result from the paper.
Motion transfer
An audio latent and text residual optimized on one clip, applied unchanged to a different video with no further tuning on it. Hit Transfer the motion! to switch from the new input to the transferred result.