Research map / Video / Video generation
TM2D: Bimodality Driven 3D Dance Generation via Music-Text Integration
Kehong Gong, Dongze Lian, Heng Chang, Chuan Guo, Zihang Jiang, Xinxin Zuo, Michael Bi Mi, Xinchao Wang
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 HumanML3D text-motion data to mixed VQ-VAE training (ablation) | TM2D · Wild Audio | FID_k: 43.85 -> 27.65 | Improved |
| add HumanML3D text-motion data to mixed VQ-VAE training (ablation) | TM2D · AIST++ | FID_k: 23.94 -> 19.01 | Improved |
| add HumanML3D text-motion data to mixed training (ablation) | TM2D · AIST++ | FID_g: 9.53 -> 20.09 | Hurt |
| add mixed-data training (ablation) | TM2D · Wild Audio | PFF: 0.78 -> 0.12 | Improved |
| sample tokens from top-k instead of argmax (ablation) | TM2D · Wild Audio | PFF: 3.76 -> 0.61 (top-k 1 -> 10, dance-only, wild) | Improved |
| replace slerp transition with TM2D late fusion for text insertion | music-text dance generation · AIST++ | MPD (ft=30): 0.135 -> 0.102 | Improved |
| joint music2dance and text2motion training in shared decoder | TM2D text2motion · HumanML3D | FID: TM2D (t2m) 1.021 vs TM2T 1.501; R precision 0.319 vs 0.424 | No clear effect |
| apply late fusion ratio 0.8 with music features at inference (ablation) | TM2D text2motion · HumanML3D | R Precision: 0.319 -> 0.300 | Hurt |
| use VQ-VAE shared codebook for both datasets | TM2D · AIST++ and HumanML3D tokens | shared codebook vectors: 846 of 1024 vectors shared (98.9% AIST++, 92.8% HumanML3D) | Improved |
| lower VQ-VAE downsample rate (ablation) | motion VQ-VAE · AIST++ + HumanML3D (30% data) | token usage and val loss | Improved |
Closest papers by meaning
- OpenDance: Multimodal Controllable 3D Dance Generation with Large-scale Internet Data (CVPR 2026 · similarity 0.92)
- MEGADance: Mixture-of-Experts Architecture for Genre-Aware 3D Dance Generation (NeurIPS 2025 · similarity 0.91)
- DanceAnyWay: Synthesizing Beat-Guided 3D Dances with Randomized Temporal Contrastive Learning (AAAI 2024 · similarity 0.89)
- Bidirectional Autoregessive Diffusion Model for Dance Generation (CVPR 2024 · similarity 0.89)
- DanceCamera3D: 3D Camera Movement Synthesis with Music and Dance (CVPR 2024 · similarity 0.88)
- MoMu-Diffusion: On Learning Long-Term Motion-Music Synchronization and Correspondence (NeurIPS 2024 · similarity 0.87)
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.
