Research map / Video / Video generation
DanceAnyWay: Synthesizing Beat-Guided 3D Dances with Randomized Temporal Contrastive Learning
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 DanceAnyWay, a two-stage hierarchical beat-pose and repletion-pose generator for audio-driven 3D dance synthesis, using a randomized temporal contrastive loss to keep motion diverse over long sequences.
Problem
Generating long, beat-aligned 3D dance sequences from music that remain diverse and avoid motion freezing or mode collapse.
Approach
Separates dance synthesis into a coarse beat-pose transformer predictor and a fine repletion-pose GAN, and adds a randomized temporal contrastive loss on latent decoder features between a randomly chosen segment and its least-similar non-overlapping segment to enforce fine-grained diversity in continuous space without quantization.
Key results
- On AIST++, DanceAnyWay reports FID_k 17.98 and FID_g 9.42 versus EDGE's 20.55 and 9.49, and BAS 0.33 versus 0.27.
- The perceptual study (31 participants) gives DAW motion quality 3.58 vs EDGE 3.36 and FACT 2.86 on the 5-point Likert scale.
- Ablations show removing the RTC loss drops MD_k from 10.62 to 6.91 and removing BPS drops BAS from 0.33 to 0.22.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the BPS network (ablation) | DanceAnyWay · AIST++ | BAS: 0.22 -> 0.33 | Helped |
| add the RTC loss (ablation) | DanceAnyWay · AIST++ | MD_k: 6.91 -> 10.62 | Helped |
| randomize the RTC reference segment instead of fixing it (ablation) | DanceAnyWay · AIST++ | FID_k: 18.82 -> 17.98 | Helped |
| use DanceAnyWay instead of EDGE | audio-to-dance generation · AIST++ | FID_k: 20.55 -> 17.98 | Helped |
| use DanceAnyWay instead of EDGE | audio-to-dance generation · AIST++ | BAS: 0.27 -> 0.33 | Helped |
Surprising findings
- Removing the RTC loss sharply reduces motion diversity, and the ablation without BPS roughly halves the beat alignment score.
Limitations
- Evaluated only on AIST++ (ten dance genres) with motion metrics and a small perceptual study of 31 participants.
- Does not explicitly model dance styles or control; future work targets style understanding and controllability.
Experimental setup
Datasets: AIST++
Benchmarks: AIST++
Metrics: FID_k FID_g MD_k MD_g BAS PFC user Likert quality score user Likert synchronization score
Baselines: Dance Transformer DanceNet DanceRevolution FACT Bailando EDGE
Builds on: EDGE Bailando FACT
Closest papers by meaning
- OpenDance: Multimodal Controllable 3D Dance Generation with Large-scale Internet Data (CVPR 2026 · similarity 0.91)
- Walk Before You Dance: High-fidelity and Editable Dance Synthesis via Generative Masked Motion Prior (AAAI 2026 · similarity 0.91)
- MEGADance: Mixture-of-Experts Architecture for Genre-Aware 3D Dance Generation (NeurIPS 2025 · similarity 0.90)
- DanceCamera3D: 3D Camera Movement Synthesis with Music and Dance (CVPR 2024 · similarity 0.90)
- Bidirectional Autoregessive Diffusion Model for Dance Generation (CVPR 2024 · similarity 0.89)
- Breaking the Group Size Barrier: Parameter-Efficient Group Dance Generation with Chain-of-Dancers (NeurIPS 2026 · similarity 0.87)
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.