Research map / Video / Video generation
Less is More: Improving Motion Diffusion Models with Sparse Keyframes
Jinseok Bae, Inwoo Hwang, Young-Yoon Lee, Ziyu Guo, Joseph Liu, Yizhak Ben-Shabat, Young Min Kim, Mubbasir Kapadia
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train on sparse keyframes (sMDM) instead of dense frames | MDM · HumanML3D | FID: 0.544 -> 0.130 | Improved |
| use reduction-selected keyframes instead of random keyframes (ablation) | sMDM · HumanML3D | FID: 0.249 -> 0.130 | Improved |
| add interpolation module with loss over all frames (ablation) | sMDM · HumanML3D | FID: 0.267 -> 0.130 | Improved |
| add Lipschitz MLP layers (ablation) | sMDM · HumanML3D | FID: 0.329 -> 0.130 | Improved |
| add Lipschitz MLP layers (ablation) | sMDM · HumanML3D | R-Precision Top1: 0.474 -> 0.494 | Improved |
| add dynamic mask update at inference (ablation) | sMDM (1000 steps) · HumanML3D | FID: 0.291 -> 0.246 | Improved |
| add dynamic mask update at inference (ablation) | sMDM (500 steps) · HumanML3D | FID: 0.322 -> 0.229 | Improved |
| add dynamic mask update at inference (ablation) | sMDM (10 steps) · HumanML3D | FID: 0.349 -> 0.385 | Hurt |
| add dynamic mask update at inference (ablation) | sMDM (50 steps) · HumanML3D | FID: 0.130 -> 0.134 | No clear effect |
| use sparse keyframe prior (sPriorMDM) instead of MDM prior | DoubleTake · HumanML3D long motion segments | FID: 0.971 -> 0.624 | Improved |
| use sparse keyframe prior (sPriorMDM) instead of MDM prior | DoubleTake · HumanML3D transitions | FID: 1.125 -> 1.705 | Hurt |
| train DiP with sparse keyframes at 60% reduction | DiP · HumanML3D | FID: 0.214 -> 0.175 | Improved |
Closest papers by meaning
- Generative Inbetweening: Adapting Image-to-Video Models for Keyframe Interpolation (ICLR 2025 · similarity 0.91)
- Enhanced Fine-Grained Motion Diffusion for Text-Driven Human Motion Synthesis (AAAI 2024 · similarity 0.91)
- Human Motion Diffusion as a Generative Prior (ICLR 2024 · similarity 0.89)
- EDEN: Enhanced Diffusion for High-quality Large-motion Video Frame Interpolation (CVPR 2025 · similarity 0.89)
- Single Motion Diffusion (ICLR 2024 · similarity 0.89)
- MotionMix: Weakly-Supervised Diffusion for Controllable Motion Generation (AAAI 2024 · similarity 0.89)
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.
