Research map / Video / Video generation
Language-Guided Transformer Tokenizer for Human Motion Generation
Sheng Yan, yong wang, Xin Du, Junsong Yuan, Mengyuan Liu
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 language guidance to tokenizer and detokenizer (ablation) | LG-Tok · Motion-X generation | FID: 0.257 -> 0.088 | Improved |
| add language guidance to tokenizer and detokenizer (ablation) | LG-Tok · HumanML3D generation | FID: 0.062 -> 0.057 | Improved |
| add language guidance (ablation) | LG-Tok · Codebook perplexity Motion-X | perplexity: 146.5 -> 103.1 | Improved |
| inject language into both tokenizer and detokenizer rather than none (ablation) | LG-Tok-tiny · HumanML3D reconstruction | FID: 0.064 -> 0.049 | Improved |
| add skip connections (ablation) | LG-Tok-tiny · HumanML3D reconstruction | FID: 0.059 -> 0.049 | Improved |
| use RMSNorm instead of LayerNorm (ablation) | LG-Tok-tiny · HumanML3D reconstruction | FID: 0.050 -> 0.049 | No clear effect |
| use SwiGLU instead of GeLU (ablation) | LG-Tok-tiny · HumanML3D reconstruction | FID: 0.050 -> 0.049 | No clear effect |
| use RoPE base 100 instead of learnable position embedding (ablation) | LG-Tok-tiny · HumanML3D reconstruction | FID: 0.065 -> 0.049 | Improved |
| use LLaMA instead of CLIP text encoder (ablation) | LG-Tok-tiny · HumanML3D reconstruction | FID: 0.051 -> 0.049 | Improved |
| add language-drop scheme (p=0.1) (ablation) | LG-Tok · HumanML3D, Motion-X, KIT-ML reconstruction | FID | Improved |
| add language-free guidance decoding (ablation) | LG-Tok · Motion-X text-conditioned generation | FID: 0.139 -> 0.088 | Improved |
| add language-free guidance decoding to free-form generation (ablation) | LG-Tok · Motion-X free-form generation | FID: 25.093 -> 14.715 | Improved |
Closest papers by meaning
- MMM: Generative Masked Motion Model (CVPR 2024 · similarity 0.91)
- Text-to-Motion Generation With Discrete Representations and Large Language Models (TPAMI 2026 · similarity 0.90)
- Language Model Beats Diffusion - Tokenizer is key to visual generation (ICLR 2024 · similarity 0.89)
- MoMask: Generative Masked Modeling of 3D Human Motions (CVPR 2024 · similarity 0.88)
- Efficient Long Video Tokenization via Coordinate-based Patch Reconstruction (CVPR 2025 · similarity 0.88)
- SnapMoGen: Human Motion Generation from Expressive Texts (NeurIPS 2025 · similarity 0.88)
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.
