Research map / Video / Video generation
Bidirectional Autoregessive Diffusion Model for Dance Generation
Read the paperCheck your idea against 108,967 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.
Proposes BADM, a bidirectional autoregressive diffusion model with slice-wise cross-attention, beat conditioning, and a 1D-conv local information decoder for music-to-dance generation on AIST++.
Problem
Music-to-dance generation with diffusion models often produces whole sequences unidirectionally, lacking local smoothness and inter-frame coherence and not aligning well with beats.
Approach
Split the noisy motion sequence into K slices denoised autoregressively with cross-attention to previously generated slices and future noise in both directions, add beat one-hot as an independent condition, and refine concatenated slices with a 1D convolutional local information decoder.
Key results
- On AIST++, BADM achieves PFC 1.424 vs EDGE 1.654 and Bailando 1.754, and Beat Align 0.2366 vs EDGE 0.2281.
- User study: BADM wins 64.9% against EDGE and 66.6% against ground truth on AIST++.
- FIDg of 20.24 for BADM is lower than EDGE's 24.71, per the paper's reported comparison.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add beat conditioning (ablation) | BADM · AIST++ | BA: 0.2235 -> 0.2366 | Helped |
| Set segmentation hyperparameter K to 3 (ablation) | BADM · AIST++ | Distk: 8.49 vs 8.29 | Hurt |
| Set segmentation hyperparameter K to 5 (ablation) | BADM · AIST++ | Distk: 8.38 vs 8.29 | Hurt |
| Set segmentation hyperparameter K to 10 (ablation) | BADM · AIST++ | PFC: 1.756 vs 1.424 | Hurt |
| Replace bidirectional encoder with unidirectional encoder (ablation) | BADM · AIST++ | PFC: 1.582 vs 1.424 | Hurt |
| Add local information decoder (ablation) | BADM · AIST++ | PFC: 1.542 -> 1.424 | Helped |
| Use BADM instead of EDGE | Music-to-dance generation · AIST++ | BA: 0.2281 -> 0.2366 | Helped |
| Use BADM instead of EDGE | Music-to-dance generation · AIST++ | PFC: 1.654 -> 1.424 | Helped |
| Use BADM instead of EDGE | Music-to-dance generation · AIST++ | Distk: 9.48 -> 8.29 | Hurt |
Surprising findings
- A global information encoder added to refine output made results less satisfying, constraining body movement and complex motions.
- The authors report FID is unreliable on AIST++ due to small test-set coverage, as noted by EDGE.
Limitations
- Fixed-step segmentation may split semantically meaningful movements.
- Evaluated mainly on AIST++; in-the-wild evaluation limited to user study with few songs.
- Some metrics (Distk, PFC) are mixed; BADM is not best on Distk versus FACT or ground truth.
Experimental setup
Datasets: AIST++
Benchmarks: AIST++
Metrics: Distk Distg Beat Align Score PFC FIDk FIDg user win rate
Baselines: FACT Bailando EDGE
Builds on: EDGE DDPM classifier-free guidance Jukebox features FiLM
Closest papers by meaning
- DanceAnyWay: Synthesizing Beat-Guided 3D Dances with Randomized Temporal Contrastive Learning (AAAI 2024 · similarity 0.89)
- Lodge: A Coarse to Fine Diffusion Network for Long Dance Generation Guided by the Characteristic Dance Primitives (CVPR 2024 · similarity 0.88)
- MEGADance: Mixture-of-Experts Architecture for Genre-Aware 3D Dance Generation (NeurIPS 2025 · similarity 0.88)
- MoMu-Diffusion: On Learning Long-Term Motion-Music Synchronization and Correspondence (NeurIPS 2024 · similarity 0.88)
- Enhancing Dance-to-Music Generation via Negative Conditioning Latent Diffusion Model (CVPR 2025 · similarity 0.88)
- MotionMix: Weakly-Supervised Diffusion for Controllable Motion Generation (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.