Research map / Video / Video generation
Retrieving and Refining Winning Noise Tickets for Diffusion-Based Motion Generation
Sakuya Ota, Qing Yu, Kent Fujiwara, Satoshi Ikehata, Ikuro Sato
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 |
|---|---|---|---|
| retrieve semantically aligned noise from null-prompt dictionary (WIN-R) (ablation) | MDM-50step · HumanML3D | FID: 0.438 -> 0.250 | Improved |
| add frame-adjusted retrieval with perturbation (WIN-FAR) over plain retrieval (ablation) | WIN-R · HumanML3D MDM-50step | R-Top1: 0.505 -> 0.512 | Improved |
| add KL-regularized noise refinement after retrieval (WINRO) (ablation) | WIN-FAR · HumanML3D MDM-50step | FID: 0.248 -> 0.115 | Improved |
| add KL-regularized refinement (ablation) | WIN-FAR · HumanML3D MotionLCM | R-Top1: 0.572 -> 0.580 | Improved |
| add LoRA noise refiner instead of iterative optimization (ablation) | WINRO · HumanML3D MDM-50step | AITS (s): 81.408 -> 0.280 | Improved |
| add KL regularization term (lambda=250 vs 0) (ablation) | WINRO · HumanML3D MotionLCM | FID: 0.136 -> 0.066 | Improved |
| use cosine similarity instead of Euclidean distance (ablation) | WINRO · HumanML3D MotionLCM | FID: 0.095 -> 0.066 | Improved |
| increase dictionary size from 100 to 10k (ablation) | WIN-FAR · HumanML3D MotionLCM | FID: 0.094 -> 0.083 | Improved |
| increase dictionary size beyond 10k to 100k (ablation) | WIN-FAR · HumanML3D MotionLCM | FID: 0.083 -> 0.085 | No clear effect |
| vary LoRA rank from 16 to 128 (ablation) | WIN-LoRA · HumanML3D MotionLCM | R-Top1: 0.553-0.560 | No clear effect |
| apply WINRO to MTT single-text generation (ablation) | MDM-50step · MTT | R@1: 12.2 -> 20.7 | Improved |
| apply optimization (WIN-O) to STMC multi-track long generation (ablation) | STMC · MTT | R@1: 33.6 -> 62.1 | Improved |
Closest papers by meaning
- Towards Detailed Text-to-Motion Synthesis via Basic-to-Advanced Hierarchical Diffusion Model (AAAI 2024 · similarity 0.91)
- MotionMix: Weakly-Supervised Diffusion for Controllable Motion Generation (AAAI 2024 · similarity 0.90)
- AMD: Autoregressive Motion Diffusion (AAAI 2024 · similarity 0.89)
- Optimizing Diffusion Noise Can Serve As Universal Motion Priors (CVPR 2024 · similarity 0.89)
- MMM: Generative Masked Motion Model (CVPR 2024 · similarity 0.89)
- AMD: Anatomical Motion Diffusion with Interpretable Motion Decomposition and Fusion (AAAI 2024 · 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.
