Research map / Video / Video generation
Improving the Diffusability of Motion Tokenizer
Guanhe Huang, Songqiao Han, Tangzheng Lian, Oya Celiktutan
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Addresses the generation-reconstruction trade-off in latent diffusion for text-driven human motion by improving the diffusability (latent generability) of the motion VAE tokenizer.
Problem
Latent diffusion models with VAE motion tokenizers show a generation-reconstruction trade-off: increasing tokenizer capacity improves reconstruction but does not proportionally improve text-conditioned motion generation.
Approach
Hypothesizes the trade-off stems from the latent space structure being hard to diffuse, and proposes improving the tokenizer's latent diffusability rather than only its reconstruction capacity. Abstract truncated; specific method not stated.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DiMoT tokenizer (CSA+AMD) to MLD++ | MLD++ · InterHuman | FID: 5.046 -> 4.716 | Improved |
| add DiMoT tokenizer to MLD++ | MLD++ · InterX | FID: 0.446 -> 0.335 | Improved |
| add Critic Spectral Adaptation to MLD (ablation) | MLD · InterHuman | R-Top3: 0.636 -> 0.689 | Improved |
| add Critic Spectral Adaptation to MLD++ (ablation) | MLD++ · InterHuman | R-Top3: 0.688 -> 0.746 | Improved |
| add Alpha-Flow Motion Decoder to MLD++ (ablation) | MLD++ · InterHuman | FID: 5.046 -> 4.781 | Improved |
| add AMD to MLD (ablation) | MLD · InterHuman | FID: 5.477 -> 5.209 | Improved |
| combine CSA and AMD vs CSA alone on MLD++ (ablation) | MLD++ · InterHuman | R-Top3: 0.746 -> 0.739 | Hurt |
| add CSA to MLD++ tokenizer (ablation) | MLD++ · InterHuman reconstruction | rFID: 0.080 -> 0.086 | Hurt |
| add DiMoT to MLD++ | MLD++ · InterHuman reconstruction | MPJPE: 49.26 -> 41.36 | Improved |
| add DiMoT to MLD++ | MLD++ · InterHuman | foot sliding: 0.0186 -> 0.0168 | Improved |
| add DiMoT to MLD++ | MLD++ · HumanML3D | FID: 0.049 -> 0.038 | Improved |
| add CSA only to MLD++ (ablation) | MLD++ · HumanML3D | R-Top3: 0.544 -> 0.552 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- InterMask — InterMask: 3D Human Interaction Generation via Collaborative Masked Modeling (ICLR 2025) · better on 2 of 2 shared results
- TIMotion — TIMotion: Temporal and Interactive Framework for Efficient Human-Human Motion Generation (CVPR 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Surprising findings
- Simply increasing motion tokenizer capacity improves reconstruction but does not proportionally improve text-conditioned generation.
Closest papers by meaning
- Reconstruction-Anchored Diffusion Model for Text-to-Motion Generation (ECCV 2026 · similarity 0.91)
- Rethinking Diffusion for Text-Driven Human Motion Generation: Redundant Representations, Evaluation, and Masked Autoregression (CVPR 2025 · similarity 0.90)
- MMM: Generative Masked Motion Model (CVPR 2024 · similarity 0.90)
- Towards Detailed Text-to-Motion Synthesis via Basic-to-Advanced Hierarchical Diffusion Model (AAAI 2024 · similarity 0.89)
- MotionMix: Weakly-Supervised Diffusion for Controllable Motion Generation (AAAI 2024 · similarity 0.88)
- Enhanced Fine-Grained Motion Diffusion for Text-Driven Human Motion Synthesis (AAAI 2024 · 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.
