Research map / 3D vision / 3D generation
Motion-2-to-3: Leveraging 2D Motion Data for 3D Motion Generations
Ruoxi Guo, Huaijin Pi, Zehong Shen, Qing Shuai, Zechen Hu, Zhumei Wang, Yajiao Dong, Ruizhen Hu, Taku Komura, Sida Peng, Xiaowei Zhou
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 2D motion pretraining of the generator before multi-view fine-tuning (ablation) | Multi-view diffusion model · HumanML3D (model design ablation) | FID: 4.950 -> 0.321 | Improved |
| add consistency block (triangulate and reproject at each diffusion step) (ablation) | Multi-view diffusion model · HumanML3D (model design ablation) | FID: 0.321 -> 0.642 | Hurt |
| use 3 views instead of 4 (ablation) | Multi-view diffusion model · HumanML3D (model design ablation) | FID: 0.321 -> 0.654 | Hurt |
| use 5 views instead of 4 (ablation) | Multi-view diffusion model · HumanML3D (model design ablation) | FID: 0.321 -> 0.593 | Hurt |
| use multi-view attention generation instead of lifting generated 2D with MotionBERT (ablation) | 2D motion generator · HumanML3D (data utilization ablation) | FID: 23.292 -> 0.321 | Improved |
| use multi-view diffusion with consistency learned from 3D data instead of MAS (ablation) | 2D motion generator · HumanML3D (data utilization ablation) | FID: 11.893 -> 0.321 | Improved |
| use root-decoupled multi-view generation instead of 2D-condition multi-view model (ablation) | 2D motion generator · HumanML3D (data utilization ablation) | FID: 0.877 -> 0.321 | Improved |
| use 2D video motion then 3D fine-tuning instead of monocular 3D annotations (Motion-X++) (ablation) | text-to-motion model · HumanML3D (data utilization ablation) | FID: 0.893 -> 0.321 | Improved |
| use 2D motion pretraining and multi-view 3D fine-tuning versus 3D-only MDM | text-to-motion model · HumanML3D | FID: 0.544 -> 0.321 | Improved |
| use our method versus MLD | text-to-motion model · HumanML3D | R-Precision (top-3): 0.772 -> 0.697 | Hurt |
| use our method versus MDM | text-to-motion model · HumanML3D | Diversity: 9.559 -> 9.286 | Hurt |
| use our method versus MDM on novel text | text-to-motion model · Novel text prompts (user study) | Best Motion Rate: 24.08% -> 51.43% | Improved |
Closest papers by meaning
- Generating Diverse and Natural 3D Human Motions From Text (CVPR 2022 · similarity 0.92)
- Move-in-2D: 2D-Conditioned Human Motion Generation (CVPR 2025 · similarity 0.90)
- AMD: Autoregressive Motion Diffusion (AAAI 2024 · similarity 0.89)
- Scaling Large Motion Models with Million-Level Human Motions (ICML 2025 · similarity 0.89)
- SnapMoGen: Human Motion Generation from Expressive Texts (NeurIPS 2025 · similarity 0.89)
- VimoRAG: Video-based Retrieval-augmented 3D Motion Generation for Motion Language Models (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.
