Professional motion editing stops guessing from one instruction: an agent finds a demo, self-checks it, and Stoppie no longer becomes Wheelie

Professional motion video editing can now be improved without retraining the editing model, by supplying externally constructed multimodal guidance. Previously an editing model received only a short instruction such as a motion name, with no statement of technique or temporal structure, so a Stoppie request produced a Wheelie and reversed the required support relationship. On the 314 professional motion editing tasks of VPMEBench, the frozen Seedance 2.0 rises from 55.1 to 65.6 overall and the frozen MiniMax-H3 from 51.2 to 64.5, on a 100-point scale.

The agent finds the demonstration, clips it, checks it, and writes one grounded prompt plus one reference clip. Standard task formulations assume a suitable reference is already available and leave guidance acquisition and verification outside the task, while relevance from search alone does not guarantee reliable guidance. The agent makes multi-turn calls to search, inspect, and clip, then self-checks the clip for motion correctness, visibility, and completeness, re-clipping or reselecting when it fails; training uses MotionCoach-SFT-10k and MotionCoach-RL-2k, and evaluation uses VPMEBench, 314 tasks over 25 leaf categories in dance, sports, and martial arts.

The report covers the professional motion editing setting: a source video and a short edit instruction go in, an edited video comes out, and the editing model stays frozen. It is written for researchers and engineers working on video editing and multimodal agents, and the evaluated setting is VPMEBench, 314 tasks across dance, sports, and martial arts. The next step stated in the source is applying the guidance construction to professional motion terms outside the training data and benchmark, where the agent autonomously gathers domain knowledge to guide the editor.

The source reports that complex terrain can yield an incomplete G-turn with an insufficient turning radius, and that subtle local articulations such as the hip rotation required by the U-shaped Hip Swing remain difficult to perform accurately. Agreement between automatic evaluation and human judgments is measured by quadratic-weighted kappa, 0.846 overall and above 0.73 on every dimension, with 95% confidence intervals reported.