Research map / Video / Video generation
SFM: Taming State Space Models for Text-to-Motion via Spatial-Frequency Modeling
Shang Gao, Haicheng Liao, Wenshuo Chen, Yumu Xie, Jiaxun Zhang, Bin Rao, Chengyue Wang, Yanchen Guan, Zhiyong Cui, Shiqi Ou, Yutao Yue, Zhenning Li
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 |
|---|---|---|---|
| add frequency state-space module (F-SSM) (ablation) | SFM · HumanML3D ablation | R-Precision Top-3: 0.733 -> 0.845 | Improved |
| add spatial state-space module (S-SSM) (ablation) | SFM · HumanML3D ablation | R-Precision Top-3: 0.779 -> 0.845 | Improved |
| add semantic attention after dual-axis fusion (ablation) | SFM · HumanML3D ablation | R-Precision Top-3: 0.758 -> 0.845 | Improved |
| add text-gated frequency modulation (ablation) | SFM · HumanML3D ablation | R-Precision Top-3: 0.804 -> 0.845 | Improved |
| use soft text-gated adaptation instead of hard DCT coefficient selection (k=64) (ablation) | SFM · HumanML3D ablation | R-Precision Top-3: 0.587 -> 0.845 | Improved |
| use DCT instead of DWT, FFT or STFT (ablation) | SFM · HumanML3D ablation | R-Precision Top-3: FFT 0.815, DWT 0.834 -> DCT 0.845 | Improved |
| replace vanilla Mamba (Motion Mamba) with SFM dual-axis design | SFM · HumanML3D | FID: 0.281 -> 0.066 | Improved |
| SFM vs Transformer MoMask on FID | SFM · HumanML3D | FID: 0.045 -> 0.066 | Hurt |
| SFM vs Transformer MDM on inference speed | SFM · HumanML3D inference (A100) | time per sentence: 15.33 s -> 0.19 s | Improved |
| SFM vs CNN StableMoFusion on inference speed | SFM · HumanML3D inference (A100) | time per sentence: 0.13 s -> 0.19 s | Hurt |
| SFM vs Motion Mamba on long sequences | SFM · HumanML3D-LS (>190 frames) | R-Precision Top-3: 0.713 -> 0.801 | Improved |
| SFM vs MoMask on long sequences FID | SFM · HumanML3D-LS (>190 frames) | FID: 0.125 -> 0.190 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Motion Mamba — Motion Mamba: Efficient and Long Sequence Motion Generation (ECCV 2024) · 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
- MoLingo: Motion-Language Alignment for Text-to-Human Motion Generation (CVPR 2026 · similarity 0.91)
- Retrieving and Refining Winning Noise Tickets for Diffusion-Based Motion Generation (ECCV 2026 · similarity 0.90)
- Priority-Centric Human Motion Generation in Discrete Latent Space (ICCV 2023 · similarity 0.90)
- MMM: Generative Masked Motion Model (CVPR 2024 · similarity 0.90)
- Towards Detailed Text-to-Motion Synthesis via Basic-to-Advanced Hierarchical Diffusion Model (AAAI 2024 · similarity 0.89)
- AMD: Autoregressive Motion Diffusion (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.
