Research map / Video / Video generation
BAMM: Bidirectional Autoregressive Motion Model
Ekkasit Pinyoanuntapong, Muhammad Usama Saleem, Pu Wang, Minwoo Lee, Srijan Das, Chen Chen
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 |
|---|---|---|---|
| use bidirectional autoregressive refinement (second iteration) after unidirectional decoding (ablation) | BAMM · HumanML3D | FID: 0.064 -> 0.055 | Improved |
| add a third decoding iteration (ablation) | BAMM · HumanML3D | FID: 0.055 -> 0.055 | No clear effect |
| mask every other token instead of low-confidence tokens in refinement (ablation) | BAMM · HumanML3D | FID: 0.065 -> 0.055 | Improved |
| use suffix masking instead of every-other-token masking (ablation) | BAMM · HumanML3D | R-Precision Top-1: 0.525 -> 0.519 | Hurt |
| raise first-iteration CFG scale from 2 to 4 (ablation) | BAMM · HumanML3D | FID: 0.105 -> 0.055 | Improved |
| raise first-iteration CFG scale from 4 to 5 (ablation) | BAMM · HumanML3D | mixed: FID 0.055 -> 0.052; Top-1 0.525 -> 0.522 | Mixed |
| use predicted instead of ground-truth length for MoMask | MoMask · HumanML3D | FID: 0.045 -> 0.090 | Hurt |
| use predicted instead of ground-truth length for MMM | MMM · HumanML3D | R-Precision Top-1: 0.515 -> 0.504 | Hurt |
| sample length multinomially instead of top-1 from a separate length estimator | MoMask · HumanML3D | FID: 0.090 -> 0.120 | Hurt |
| sample length multinomially instead of top-1 | MMM · HumanML3D | FID: 0.080 -> 0.099 | Hurt |
| use BAMM bidirectional autoregressive model instead of MoMask for editing | temporal editing · HumanML3D prefix/suffix | R-Precision Top-1: prefix 0.536 vs 0.532; suffix 0.532 vs 0.527 | No clear effect |
| use BAMM instead of MDM for temporal editing | temporal editing · HumanML3D inpainting | FID: 2.362 -> 0.056 | Improved |
Later papers that compared with this method
6 later accepted papers report BAMM as a baseline in their results tables; 5 of them report a better number on most of the results they share with it.
- ParTY: Part-Guidance for Expressive Text-to-Motion Synthesis (CVPR 2026 · better on 5 of 5 shared results)
- K-Mask: Kinematic-Aware Masked Modeling for Controllable Text-to-Motion Synthesis (ECCV 2026 · better on 4 of 4 shared results)
- Think Before You Move: Latent Motion Reasoning for Text-to-Motion Generation (TPAMI 2026 · better on 1 of 1 shared result)
- HERO: Human Reaction Generation from Videos (ICCV 2025 · better on 2 of 3 shared results)
- Dual Reciprocal Learning of Language-based Human Motion Understanding and Generation (ICCV 2025 · better on 0 of 1 shared result)
- DisCoRD: Discrete Tokens to Continuous Motion via Rectified Flow Decoding (ICCV 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.
Closest papers by meaning
- MMM: Generative Masked Motion Model (CVPR 2024 · similarity 0.93)
- AMD: Autoregressive Motion Diffusion (AAAI 2024 · similarity 0.91)
- Priority-Centric Human Motion Generation in Discrete Latent Space (ICCV 2023 · similarity 0.91)
- Rethinking Diffusion for Text-Driven Human Motion Generation: Redundant Representations, Evaluation, and Masked Autoregression (CVPR 2025 · similarity 0.91)
- VersatileMotion: A Unified Framework for Motion Synthesis and Comprehension (ECCV 2026 · similarity 0.91)
- SnapMoGen: Human Motion Generation from Expressive Texts (NeurIPS 2025 · similarity 0.90)
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.
