Research map / Architectures and efficiency / RNNs and normalization
Light-T2M: A Lightweight and Fast Model for Text-to-motion Generation
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 Light-T2M, a lightweight diffusion-based text-to-motion model combining local 1D-conv modeling, Mamba with a pseudo-bidirectional scan, and an adaptive textual information injector, reaching 4.48M parameters and FID 0.040 on HumanML3D.
Problem
Text-to-motion generation models have large parameter counts and slow inference, raising deployment cost; existing methods underuse local motion information.
Approach
Combine lightweight 1D convolutional local modeling with Mamba-based global modeling on temporally downsampled segments, use a pseudo-bidirectional scan (concatenating the reversed sequence) to get bidirectional context without extra parameters, and inject text via a gated, segment-aware adaptive textual information injector.
Key results
- On HumanML3D, Light-T2M reaches FID 0.040 vs MoMask 0.045 with 4.48M vs 44.85M trainable parameters.
- On KIT-ML, FID 0.161 vs MoMask 0.204 (table) / 0.228 (abstract).
- Inference time 0.151s vs MoMask 0.180s on an RTX 3090Ti.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace MoMask with Light-T2M | Text-to-motion generation · HumanML3D test | FID: 0.045^{\pm.002} -> 0.040^{\pm{.002}} | Helped |
| Replace MoMask with Light-T2M | Text-to-motion generation · KIT-ML test | FID: 0.204^{\pm.011} -> 0.161^{\pm.009} | Helped |
| Replace MoMask with Light-T2M | Text-to-motion generation · HumanML3D test | AIT: 0.180s -> 0.151s | Helped |
| Replace Light-T2M with Transformer encoder blocks (ablation) | Light-T2M · HumanML3D test | FID: 0.040 -> 0.214 | Hurt |
| Use LTL blocks instead of LGL blocks (ablation) | Light-T2M · HumanML3D test | FID: 0.040 -> 0.160 | Hurt |
| Use original Mamba block design LML instead of LGL (ablation) | Light-T2M · HumanML3D test | FID: 0.040 -> 0.165 | Hurt |
| Use BiMamba blocks instead of LGL blocks (ablation) | Light-T2M · HumanML3D test | FID: 0.040 -> 0.114 | Hurt |
| Replace pseudo-bidirectional scan with single-directional scan (ablation) | Light-T2M · HumanML3D test | FID: 0.040 -> 0.058 | Hurt |
| Replace pseudo-bidirectional scan with bidirectional scan (ablation) | Light-T2M · HumanML3D test | FID: 0.040 -> 0.042 | Hurt |
| Add ATII (ablation) | Light-T2M · HumanML3D test | FID: 0.102 -> 0.040 | Helped |
| Add the ATII gating mechanism (ablation) | Light-T2M · HumanML3D test | FID: 0.074 -> 0.040 | Helped |
| Increase basic blocks from 2 to 4 (ablation) | Light-T2M · HumanML3D test | FID: 0.086 -> 0.040 | Helped |
Surprising findings
- A 2-block variant with 2.28M parameters still reaches FID 0.086, and the model's pure-transformer variant (TTT) reaches 0.175 FID.
- Predicting M0 rather than noise epsilon was found to yield better performance in T2M.
Limitations
- Falls short of MoMask on R-Precision and multimodal distance on HumanML3D
- Diversity (MultiModality) metric not included in main evaluation; Light-T2M has worse MM than several baselines
- Inference-time comparison only on HumanML3D and only against a subset of methods
Experimental setup
Datasets: HumanML3D KIT-ML
Benchmarks: HumanML3D KIT-ML
Metrics: FID R-Precision MM. Dist. MM average inference time #Params
Baselines: MoMask ReMoDiffuse T2M-GPT MotionDiffuse MDM MLD T2M TM2T Motion Mamba
Builds on: MoMask Mamba CLIP classifier-free guidance DDPM
Closest papers by meaning
- MMM: Generative Masked Motion Model (CVPR 2024 · similarity 0.87)
- Scaling Large Motion Models with Million-Level Human Motions (ICML 2025 · similarity 0.87)
- OMG: Towards Open-vocabulary Motion Generation via Mixture of Controllers (CVPR 2024 · similarity 0.86)
- Towards Detailed Text-to-Motion Synthesis via Basic-to-Advanced Hierarchical Diffusion Model (AAAI 2024 · similarity 0.86)
- SnapMoGen: Human Motion Generation from Expressive Texts (NeurIPS 2025 · similarity 0.86)
- Enhancing Motion in Text-to-Video Generation with Decomposed Encoding and Conditioning (NeurIPS 2024 · 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.