Research map / Video / Video generation

Breaking the Group Size Barrier: Parameter-Efficient Group Dance Generation with Chain-of-Dancers

NeurIPS 2026 · Video › Video generation · approach: Trajectory-guided motion control

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.

Reformulates group dance generation as a sequential Chain-of-Dancers over a frozen single-dancer diffusion backbone, adding lightweight role-aware text and group-aware graph motion encoders plus inference-time noise optimization.

Problem

Group dance generation from music is tied to fixed group sizes by joint transformer models, requiring retraining per size and causing per-dancer identity entanglement and swaps.

Approach

Decompose group choreography into a chain of pairwise conditional single-dancer generations using a shared frozen diffusion backbone with lightweight RATE (text) and GAME (distance-weighted GCN over previously generated dancers) encoders, plus shared-noise optimization with a global penalty at inference.

Key results

• On AIOZ-GDance with 3 dancers, ChainDance reaches GMR 11.63 and GMC 83.74, versus best baseline TCDiff GMR 15.36 and GMC 82.77.
• Single-dancer FID 23.12 and PFC 0.50, best among compared methods.
• Uses 15.27M parameters and 0:14 training time per epoch, reported as 3-4x fewer parameters and 3-6x less training time than baselines.

Surprising findings

• The variant without the global penalty still outperforms most baselines, suggesting the chain-of-dancers design drives most of the gains.
• Inference-time refinement raises inference time from 0:06 to 0:15 with only modest group-metric gains over the no-penalty variant.

Limitations

• Group-metric gains over baselines are modest and the 5-dancer GMR is slightly worse than TCDiff.
• Relies on rule-based weak-supervision interaction labels and ChatGPT-4o-generated instructions.
• Only one public large-scale group dance benchmark (AIOZ-GDance) is used for main evaluation, with limited 5-dancer test samples.

Experimental setup

Datasets: AIOZ-GDance AIST++ FineDance DD100

Benchmarks: AIOZ-GDance DD100

Metrics: FID Div PFC GMR GMC TIF FLOPs parameters training time per epoch inference time

Baselines: GCD CoDancers TCDiff ST-GDance PINO Duolando EDGE

Builds on: Lodge PINO DanceChat EDGE

Closest papers by meaning

• MEGADance: Mixture-of-Experts Architecture for Genre-Aware 3D Dance Generation (NeurIPS 2025 · similarity 0.87)
• DanceAnyWay: Synthesizing Beat-Guided 3D Dances with Randomized Temporal Contrastive Learning (AAAI 2024 · similarity 0.87)
• Harmonious Music-driven Group Choreography with Trajectory-Controllable Diffusion (AAAI 2025 · similarity 0.87)
• Duolando: Follower GPT with Off-Policy Reinforcement Learning for Dance Accompaniment (ICLR 2024 · similarity 0.87)
• Lodge: A Coarse to Fine Diffusion Network for Long Dance Generation Guided by the Characteristic Dance Primitives (CVPR 2024 · similarity 0.86)
• Bidirectional Autoregessive Diffusion Model for Dance Generation (CVPR 2024 · similarity 0.86)