Research map / Video / Video generation
Generating Human Motion From Textual Descriptions With Discrete Representations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 discrete VQ-VAE representations with EMA and Code Reset, then autoregressively predict codes with GPT | T2M-GPT · HumanML3D test | FID: MotionDiffuse 0.630; T2M-GPT 0.116 | Helped |
| Corrupt 50% of training code indices (ablation) | T2M-GPT · HumanML3D test | R-Precision Top-1: 0.417 -> 0.491 | Helped |
| Corrupt 50% of training code indices (ablation) | T2M-GPT · HumanML3D test | FID: 0.140 -> 0.116 | Helped |
| Use Code Reset alongside EMA (ablation) | Motion VQ-VAE · HumanML3D test reconstruction | FID: 0.097 -> 0.070 | Helped |
| Use EMA alongside Code Reset (ablation) | Motion VQ-VAE · HumanML3D test reconstruction | FID: 0.102 -> 0.070 | Helped |
| Use EMA and Code Reset instead of neither (ablation) | T2M-GPT · HumanML3D test generation | FID: 42.797 -> 0.116 | Helped |
| Use L1Smooth reconstruction loss instead of L1 (ablation) | Motion VQ-VAE · HumanML3D test reconstruction | FID: 0.144 -> 0.070 | Helped |
| Use L1Smooth reconstruction loss instead of L2 (ablation) | Motion VQ-VAE · HumanML3D test reconstruction | FID: 0.292 -> 0.070 | Helped |
| Set the velocity-loss weight α to 0.5 (ablation) | Motion VQ-VAE with L1Smooth · HumanML3D test reconstruction | FID: 0.112 -> 0.070 | Helped |
| Use 512 codebook entries instead of 256 (ablation) | Motion VQ-VAE · HumanML3D test reconstruction | FID: 0.145 -> 0.070 | Helped |
| Use 512 codebook entries instead of 1024 (ablation) | Motion VQ-VAE · HumanML3D test reconstruction | FID: 0.090 -> 0.070 | Helped |
| Use dilation rates 9, 3, 1 instead of 1, 1, 1 (ablation) | Motion VQ-VAE · HumanML3D test reconstruction | FID: 0.145 -> 0.070 | Helped |
Closest papers by meaning
- Rethinking Diffusion for Text-Driven Human Motion Generation: Redundant Representations, Evaluation, and Masked Autoregression (CVPR 2025 · similarity 0.91)
- MMM: Generative Masked Motion Model (CVPR 2024 · similarity 0.90)
- MoMask: Generative Masked Modeling of 3D Human Motions (CVPR 2024 · similarity 0.89)
- Shape My Moves: Text-Driven Shape-Aware Synthesis of Human Motions (CVPR 2025 · similarity 0.89)
- Seamless Human Motion Composition with Blended Positional Encodings (CVPR 2024 · similarity 0.89)
- SnapMoGen: Human Motion Generation from Expressive Texts (NeurIPS 2025 · 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.