Research map / Video / Video generation
Motion Mamba: Efficient and Long Sequence Motion Generation
Zeyu Zhang, Akide Liu, Ian Reid, RICHARD HARTLEY, Bohan Zhuang, Hao Tang
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 |
|---|---|---|---|
| Use hierarchical scan schedule {S_2N-1..S_1} instead of ascending scans {S_1..S_N} (ablation) | Motion Mamba HTM block · HumanML3D | FID: 1.278 -> 0.281 | Improved |
| Reverse the scan order to decrease with depth (ablation) | Motion Mamba HTM block · HumanML3D | FID: 1.278 -> 0.962 | Improved |
| Use block-level bidirectional scan in BSM instead of single scan (ablation) | Motion Mamba BSM block · HumanML3D | FID: 1.063 -> 0.281 | Improved |
| Use block-level instead of layer-level bidirectional scan (ablation) | Motion Mamba BSM block · HumanML3D | FID: 0.789 -> 0.281 | Improved |
| Set latent dimension to 2 instead of 1 (ablation) | Motion Mamba · HumanML3D | FID: 0.652 -> 0.281 | Improved |
| Increase latent dimension from 2 to 10 (ablation) | Motion Mamba · HumanML3D | FID: 0.281 -> 0.628 | Hurt |
| Increase depth from 11 to 27 or 37 layers (ablation) | Motion Mamba · HumanML3D | FID: 0.281 -> 0.975 (27 layers), 0.809 (37 layers) | Hurt |
| Reduce depth from 11 to 9 layers (ablation) | Motion Mamba · HumanML3D | FID: 0.281 -> 1.080 | Hurt |
| Replace transformer diffusion denoiser with Mamba-based Motion Mamba | MLD · HumanML3D | inference time per sequence: 0.217 s -> 0.058 s | Improved |
| Use Motion Mamba instead of MDM on diversity-style metrics | Motion Mamba · HumanML3D | MModality: 2.799 -> 2.294 | Hurt |
| Use Motion Mamba instead of MLD on long sequences | Motion Mamba · HumanML3D-LS | FID: 0.952 -> 0.668 | Improved |
Later papers that compared with this method
1 later accepted paper reports Motion Mamba as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- SFM: Taming State Space Models for Text-to-Motion via Spatial-Frequency Modeling (ECCV 2026 · better on 3 of 3 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- PoseMamba: Monocular 3D Human Pose Estimation with Bidirectional Global-Local Spatio-Temporal State Space Model (AAAI 2025 · similarity 0.87)
- Scaling Large Motion Models with Million-Level Human Motions (ICML 2025 · similarity 0.87)
- Vision Mamba Mender (NeurIPS 2024 · similarity 0.87)
- LC-Mamba: Local and Continuous Mamba with Shifted Windows for Frame Interpolation (CVPR 2025 · similarity 0.86)
- MambaOut: Do We Really Need Mamba for Vision? (CVPR 2025 · similarity 0.86)
- Achilles' Heel of Mamba: Essential difficulties of the Mamba architecture demonstrated by synthetic data (NeurIPS 2025 · similarity 0.86)
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.
