Research map / Image generation / Editing and style transfer
Dynamic Motion Blending for Versatile Motion Editing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MotionCutMix, an online body-part blending augmentation, and MotionReFit, an auto-regressive diffusion model with a body part coordinator classifier guidance, for text-guided motion editing; introduces the STANCE dataset.
Problem
Text-guided motion editing is limited by scarce (original, edited, instruction) training triplets, which restricts generalization across style edits, body-part edits, and temporal edits.
Approach
Synthesize training triplets online by spatially blending body-part motions from a large unannotated motion base (MotionCutMix), then train an auto-regressive conditional diffusion editor whose outputs are refined by a discriminator-based body part coordinator used as classifier guidance.
Key results
- On the body part replacement task, Ours full reaches FID 0.20 and E2S AvgR 4.65, versus TMED at FID 0.52 and E2S AvgR 6.47 (lower is better for AvgR in the paper's convention).
- On MotionFix, even without MotionCutMix, the auto-regressive model (Ours w/o MCM) reaches E2S R@1 83.47 versus TMED 71.77.
- Ablations show MotionCutMix improves both the proposed model and TMED; performance scales with MotionCutMix ratio and the method keeps strong performance with reduced annotated data.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add MotionCutMix training augmentation (ablation) | MotionReFit · STANCE body part replacement | FID: 0.23±.026 -> 0.20±.025 | Helped |
| add MotionCutMix training augmentation (ablation) | MotionReFit · STANCE style transfer | FID: 0.23±.017 -> 0.14±.015 | Helped |
| add MotionCutMix training augmentation (ablation) | TMED · STANCE style transfer | FID: 1.54±.093 -> 0.84±.060 | Helped |
| add MotionCutMix training augmentation (ablation) | TMED · STANCE body part replacement | FID: 0.52±.034 -> 0.54±.028 | Hurt |
| add body part coordinator (ablation) | MotionReFit · STANCE body part replacement | FID: 0.23±.016 -> 0.20±.025 | Helped |
| add body part coordinator (ablation) | MotionReFit · STANCE style transfer | FID: 0.16±.018 -> 0.14±.015 | Helped |
| increase the MotionCutMix ratio (ablation) | MotionReFit · STANCE body part replacement and style transfer | AvgR | Helped |
| reduce annotated training data while using MotionCutMix (ablation) | MotionReFit · STANCE body part replacement and style transfer | AvgR | Helped |
| use a 16-frame sliding window (ablation) | MotionReFit · STANCE body part replacement and style transfer | AvgR | Helped |
Surprising findings
- Adding MotionCutMix to TMED (TMED w/ MCM) improves its retrieval results, indicating the augmentation transfers to other architectures.
Limitations
- Struggles with long-term temporal dependencies.
- Lacks spatial awareness for position-dependent instructions.
Experimental setup
Datasets: HumanML3D MotionFix STANCE
Benchmarks: STANCE MotionFix
Metrics: FID Diversity Foot Score Edited-to-Source Retrieval R@1/R@2/R@3/AvgR Edited-to-Target Retrieval R@1/R@2/R@3/AvgR Multimodality
Baselines: MDM-BP TMED TMED w/ MCM
Builds on: DDPM classifier-free guidance CutMix TMED MDM
Closest papers by meaning
- SimMotionEdit: Text-Based Human Motion Editing with Motion Similarity Prediction (CVPR 2025 · similarity 0.92)
- MMM: Generative Masked Motion Model (CVPR 2024 · similarity 0.89)
- Enhanced Fine-Grained Motion Diffusion for Text-Driven Human Motion Synthesis (AAAI 2024 · similarity 0.89)
- MotionEditor: Editing Video Motion via Content-Aware Diffusion (CVPR 2024 · similarity 0.88)
- SALAD: Skeleton-aware Latent Diffusion for Text-driven Motion Generation and Editing (CVPR 2025 · similarity 0.88)
- MotionMix: Weakly-Supervised Diffusion for Controllable Motion Generation (AAAI 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.