PROFESSIONAL MOTION EDITING

Constructing Multimodal Guidance
for Video Professional Motion Editing

Siyi Qian^1,2,*Hanqing Yang^2,*,†,‡Weihuang Lin²Zihao Yang^3,2Yichen Lu⁴Yuanhong Zheng¹Sheng Li^1,‡

¹Peking University²Alibaba Token Hub, Alibaba Group³University of Science and Technology of China⁴Zhejiang University

^*Equal contribution^†Corresponding author^‡Project lead

Professional motion editing

EDIT PROMPT

Motivation

Beyond general scenarios, we study video professional motion editing, where edited videos must show technically demanding and complex motions while preserving other source content. However, existing video editing models often generate an incorrect motion. For example, a Stoppie request instead produces a Wheelie, reversing the required support relationship. Correct motions require appropriate technique and temporal structure, including body coordination, support, and stage order or a key pose. A short instruction rarely specifies these details, while the editing models lack the knowledge coverage required for professional motions in vertical scenarios. To address this problem, we introduce MotionCoach, a trained autonomous multimodal agent that learns to perform multi-turn tool use and verify references to obtain reliable and sufficient guidance for professional motion editing.

Figure 1. A Stoppie prompt is incorrectly rendered as a Wheelie with direct editing, while MotionCoach provides grounded guidance for correct wheel support and motion.

Abstract

Video editing models achieve strong performance in modifying appearance, backgrounds, styles, and motions, empowered by large-scale training data. Beyond general scenarios, professional motions are significant in performance, sports, and martial arts. This paper focuses on professional motion editing, which involves fine-grained motion technique and temporal structure.

Given professional motion names but lacking detailed descriptions, existing models often fail to generate professional motions, such as incorrectly lifting the front wheel under a Stoppie instruction. To address this challenge, we introduce MotionCoach, a trained multimodal agent that uses multi-turn tools and performs self-verification, producing a grounded prompt and a reference clip as reliable and sufficient multimodal guidance to inject external professional motion knowledge into video editing models.

To supply training data, we build a data pipeline to construct trajectories. We cold-start the agent on MotionCoach-SFT-10k and apply Group Relative Policy Optimization (GRPO) on MotionCoach-RL-2k with rewards on the motion guidance prompt, the motion outcome video, and the format. For evaluation, we introduce VPMEBench, comprising 314 professional motion editing tasks with fine-grained references and annotations. Extensive experiments show that MotionCoach improves the overall score from 55.1 to 65.6 on Seedance 2.0 and from 51.2 to 64.5 on MiniMax-H3. Learning to construct multimodal guidance enables precise professional motion editing without retraining video editing models.

Method

Professional motion editing requires motion technique and temporal structure. To provide richer external motion knowledge, MotionCoach takes a source video and an edit prompt, calls the search, inspect, and clip tools, and verifies the references itself to produce a grounded prompt and a verified reference clip as multimodal guidance. A frozen video editing model then generates the edited video.

We cold-start the policy on MotionCoach-SFT-10k to establish multi-turn formatting and basic tool use, and agentic RL on MotionCoach-RL-2k then improves the policy through the quality of the resulting guidance and the frozen model’s editing performance.

MotionCoach agentic framework and reinforcement learning overview. — Top: the multi-turn agentic loop during rollout and inference uses search, inspect, and clip tools with self-verification to produce reliable and sufficient motion guidance. Bottom: Agentic RL updates the policy with motion guidance prompt, motion outcome video, and format rewards while keeping the video editing model frozen.

Benchmark

General video editing benchmarks do not provide comparable references or annotations for named professional atomic motions. To enable fine-grained evaluation of professional motion editing, we construct VPMEBench with references to motion technique and temporal structure. VPMEBench contains 314 professional motion editing tasks across 3 top-level, 10 second-level, and 25 leaf categories, each taking a source video and an edit prompt as input and providing fine-grained visual references and textual annotations of the target motion.

VPMEBench taxonomy, motion annotations, reference keyframes and evaluation overview. — 314 professional motion editing tasks under a hierarchical taxonomy over dance, sports, and martial arts. Each task provides visual references and textual annotations of the target motion for evaluation along motion professionalism (MP), physical plausibility (PP), and overall video quality (VQ).

Semantic keyframes

Results

Applying MotionCoach substantially improves the frozen MiniMax-H3 across all reported metrics. Furthermore, when transferred to Seedance 2.0, our method generalizes well across most dimensions, underscoring its plug-and-play applicability. Overall, these results show that the learned guidance-construction strategy generalizes across video editing models.

The untrained workflow yields inconsistent gains across Seedance 2.0 and MiniMax-H3, showing that tool access alone does not provide reliable improvements. Subsequently, SFT consistently improves performance across both base editors, and RL further refines the SFT policy. The overall gains therefore stem from learning to construct useful guidance rather than merely executing the workflow.

Main results

Selected main benchmark scores. MP: motion professionalism, PP: physical plausibility, VQ: overall video quality.
Video editor	MP ↑	PP ↑	VQ ↑	Overall ↑
Seedance 2.0	29.6	85.8	77.1	55.1
+ MotionCoach	53.0	87.1	72.2	65.6
MiniMax-H3	16.2	84.7	87.3	51.2
+ MotionCoach	33.1	92.2	98.2	64.5

Learning the workflow

Overall score

Training stage ablation
Agent training	Seedance 2.0	MiniMax-H3
Base	55.1	51.2
Base + Workflow	54.0	52.2
Base + SFT	62.3	62.8
Base + RL	65.6	64.5

More motion edits

Citation

@misc{qian2026motioncoachconstructingmultimodalguidance,
  title={MotionCoach: Constructing Multimodal Guidance for Video Professional Motion Editing},
  author={Siyi Qian and Hanqing Yang and Weihuang Lin and Zihao Yang and Yichen Lu and Yuanhong Zheng and Sheng Li},
  year={2026},
  eprint={2610.12963},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2610.12963},
}

MMotonCoachConstructing Multimodal Guidance for Video Professional Motion Editing