Research map / Video / Video generation
Harmonious Music-driven Group Choreography with Trajectory-Controllable Diffusion
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 TCDiff, a two-stage framework that first predicts non-overlapping dancer trajectories (DTN with distance-consistency loss) and then generates dance motion via a trajectory-conditioned diffusion model with a footwork adaptor and RFK loss, to reduce multi-dancer collisions and foot sliding.
Problem
Music-driven group dance generation suffers from multi-dancer collisions (dancer ambiguity from similar movements) and single-dancer foot sliding.
Approach
Decouple dancer coordinates from movements: first generate collision-free trajectories with an autoregressive navigator trained with a distance-consistency loss, then condition a diffusion model on these fixed trajectories, using a fusion projection to reduce dancer ambiguity, a footwork adaptor driven by trajectory displacement, and a relative forward-kinematic loss to reduce foot sliding.
Key results
- On AIOZ-GDance, TCDiff reaches GMR 13.86 versus 26.10 for the next best (CoDancers), and TIF 0.13 vs 0.10 for CoDancers (CoDancers is lower on TIF).
- Single-dancer PFC of 0.51 versus 2.53 to 3.26 for baselines, indicating reduced foot sliding.
- Div of 15.10 vs 10.87 for GCD among methods, with user study favoring TCDiff on realism, correlation, aesthetics and harmony.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TCDiff instead of CoDancers | Group choreography generation · AIOZ-GDance test | GMR: 26.10 -> 13.86 | Helped |
| Use TCDiff instead of GCD | Group choreography generation · AIOZ-GDance test | GMC: 80.97 -> 81.98 | Helped |
| Use TCDiff instead of CoDancers | Group choreography generation · AIOZ-GDance test | TIF: 0.10 -> 0.13 | Hurt |
| Use TCDiff instead of CoDancers | Group choreography generation · AIOZ-GDance test | FID: 23.98 -> 37.47 | Hurt |
| Use TCDiff instead of GCD | Group choreography generation · AIOZ-GDance test | Div: 10.87 -> 15.10 | Helped |
| Use TCDiff instead of EDGE | Group choreography generation · AIOZ-GDance test | MMC: 0.26 -> 0.25 | Hurt |
| Use TCDiff instead of CoDancers | Group choreography generation · AIOZ-GDance test | PFC: 3.26 -> 0.51 | Helped |
| Add Conditional Motion Denoising (ablation) | TCDiff · AIOZ-GDance test | GMR: 33.25 -> 13.86 | Helped |
| Add Conditional Motion Denoising (ablation) | TCDiff · AIOZ-GDance test | MMC: 0.23 -> 0.25 | Helped |
| Add Footwork Adaptor and Fusion Projection (ablation) | TCDiff · AIOZ-GDance test | GMR: 25.57 -> 13.86 | Helped |
| Add Footwork Adaptor (ablation) | TCDiff · AIOZ-GDance test | PFC: 3.25 -> 0.51 | Helped |
| Add Footwork Adaptor (ablation) | TCDiff · AIOZ-GDance test | GMC: 80.13 -> 81.98 | Helped |
Surprising findings
- CoDancers achieves lower TIF (0.10) than TCDiff (0.13), so TCDiff does not win on the collision metric in the main table.
Limitations
- FID is worse than some baselines (37.47 vs 23.98 for CoDancers), trading individual fidelity for group metrics.
- TIF is not best; CoDancers reports lower collision frequency.
Experimental setup
Datasets: AIOZ-GDance
Benchmarks: AIOZ-GDance
Metrics: GMR GMC TIF FID Div MMC PFC
Baselines: EDGE GDanceR GCD CoDancers
Builds on: EDGE denoising diffusion probabilistic models SMPL Librosa
Closest papers by meaning
- Breaking the Group Size Barrier: Parameter-Efficient Group Dance Generation with Chain-of-Dancers (NeurIPS 2026 · similarity 0.87)
- MEGADance: Mixture-of-Experts Architecture for Genre-Aware 3D Dance Generation (NeurIPS 2025 · similarity 0.86)
- Lodge: A Coarse to Fine Diffusion Network for Long Dance Generation Guided by the Characteristic Dance Primitives (CVPR 2024 · similarity 0.86)
- Duolando: Follower GPT with Off-Policy Reinforcement Learning for Dance Accompaniment (ICLR 2024 · similarity 0.85)
- Bidirectional Autoregessive Diffusion Model for Dance Generation (CVPR 2024 · similarity 0.85)
- DanceCamera3D: 3D Camera Movement Synthesis with Music and Dance (CVPR 2024 · 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.