Research map / Video / Video generation
K-Mask: Kinematic-Aware Masked Modeling for Controllable Text-to-Motion Synthesis
Faisal Ahmed, Chenqiu Zhao, Anup Basu
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 kinematic group masking (pkin 0.25) to random token masking (ablation) | K-Mask Base-T · HumanML3D | FID: 0.108 -> 0.041 | Improved |
| raise kinematic masking probability to 0.5 (ablation) | K-Mask Base-T · HumanML3D | FID: 0.041 -> 0.078 | Hurt |
| add LAKD loss to the group tokenizer (ablation) | KG-RVQ · HumanML3D short-run tokenizer | cross-group leakage: 2.484 -> 0.088 | Improved |
| add LAKD loss to the group tokenizer (ablation) | KG-RVQ · HumanML3D short-run tokenizer | MPJPE: 16.4 -> 17.2 | Hurt |
| use 6 anatomical groups instead of whole-body tokenization (ablation) | KG-RVQ · HumanML3D short-run tokenizer | MPJPE: 30.6 -> 17.2 | Improved |
| use 22 joint-level groups instead of 6 (ablation) | KG-RVQ · HumanML3D short-run tokenizer | cross-group leakage: 0.088 -> 1.489 | Hurt |
| proportional latent allocation instead of uniform split (ablation) | KG-RVQ · HumanML3D short-run tokenizer | MPJPE: 18.5 -> 17.2 | Improved |
| temporal-then-spatial factorized attention instead of spatial-then-temporal (ablation) | K-Mask transformers · HumanML3D | FID: 0.077 -> 0.041 | Improved |
| axis-factorized attention instead of full self-attention (ablation) | K-Mask transformers · HumanML3D | FID: 0.049 -> 0.041 | No clear effect |
| group-local token resampling for localized editing vs whole-body inpainting (MoMask) | K-Mask · HumanML3D 200-case editing | Spillover: .164 -> .108 | Improved |
| K-Mask vs MoGenTS on generation fidelity | K-Mask · HumanML3D test | FID: MoGenTS .028 vs K-Mask .041 (Table 2) while Table 1 text claims best FID | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- BAMM — BAMM: Bidirectional Autoregressive Motion Model (ECCV 2024) · better on 4 of 4 shared results
- ReMoDiffuse — ReMoDiffuse: Retrieval-Augmented Motion Diffusion Model (ICCV 2023) · better on 2 of 2 shared results
- ParCo — ParCo: Part-Coordinating Text-to-Motion Synthesis (ECCV 2024) · 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.
Closest papers by meaning
- MMM: Generative Masked Motion Model (CVPR 2024 · similarity 0.90)
- MotionDreamer: One-to-Many Motion Synthesis with Localized Generative Masked Transformer (ICLR 2025 · similarity 0.86)
- Enhanced Fine-Grained Motion Diffusion for Text-Driven Human Motion Synthesis (AAAI 2024 · similarity 0.86)
- MoMask: Generative Masked Modeling of 3D Human Motions (CVPR 2024 · similarity 0.86)
- InterMask: 3D Human Interaction Generation via Collaborative Masked Modeling (ICLR 2025 · similarity 0.86)
- Breaking the Group Size Barrier: Parameter-Efficient Group Dance Generation with Chain-of-Dancers (NeurIPS 2026 · similarity 0.85)
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.
