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

Jing Xu; Cunjian Chen; Qiuhong Ke

cs.CVOct 8, 2026

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

Authors: Jing Xu, Cunjian Chen, Qiuhong Ke

Organizations: Department of Data Science and Artificial Intelligence, Monash University, Australia

Abstract

Group dance generation aims to synthesize coordinated multi-dancer choreography from music, with broad applications in animation and interactive content creation. This task requires modeling dense inter-person dependencies to ensure spatial coordination, while naturally preserving individual dancer identities. Existing approaches model all dancers jointly with end-to-end transformers, which tie the architecture to a fixed group size and entangle per-dancer identities across frames. We propose ChainDance, a scalable framework that reformulates group dance generation as a Chain-of-Dancers: a sequential decomposition over per-dancer conditional distributions, allowing a single model to scale across variable group sizes without retraining and naturally preserving per-dancer identity. Built on a frozen single-dancer diffusion backbone, ChainDance introduces two lightweight modules: a Role-Aware Text Encoder (RATE) for per-dancer semantic conditioning, and a Group-Aware Motion Encoder (GAME) that aggregates previously generated dancers via a distance-weighted graph convolutional network, and incorporates a training-free noise optimization procedure at inference time to enforce global spatial coherence. Experiments on AIOZ-GDance demonstrate that ChainDance achieves state-of-the-art motion quality and group coordination while structurally preserving per-dancer identity, with $3$ - $4\times$ fewer parameters and requiring $3$ - $6\times$ less training time compared to prior approaches.

Figures & tables

Figure 1 : Key challenges in group dance generation. (a) Group-size rigidity: adapting to a new group size $N$ requires retraining from scratch, but self-attention over the concatenated sequence scales as $O((NL)^{2})$ , making per-size retraining prohibitively expensive. As a result, existing models are locked to a fixed group size. (b) Identity entanglement: joint representations obscure per-dancer identity, causing frequent swaps across frames. We visualize three consecutive frames from CoDancers [ 38 ] : between Frame 1 and Frame 2, the pink and blue dancers abruptly swap positions; in Frame 3, the blue dancer suddenly relocates to the leftmost position. When rendered as video, these inconsistencies appear as visible flickering and identity-switching artifacts.

Figure 2 : Overall framework of ChainDance. The model builds on a frozen single dance diffusion backbone and introduces lightweight multimodal encoders for text, music, and other dancers. Through sequential chain generation and inference-time global optimization, ChainDance produces coherent and coordinated group choreographies.

Figure 3 : Prompt generation and role-specific conditioning. The system converts structured multimodal captions into individualized text prompts for each dancer. Based on a global choreography description, dancer pairs are randomly sampled and assigned interaction keywords (e.g., approach , keep distance , face to , lead follow , swap , or passby ), providing diverse yet coherent textual conditions for ChainDance.

Figure 4 : Chain-of-Dancers generation design. At inference, ChainDance generates dancers sequentially and refines global consistency via diffusion-based optimization. Each dancer is conditioned on the previously generated one, and this process repeats until the whole group is complete. After that, optimize the shared initial noise via gradient descent until the global penalty $L$ converges below the target threshold. All pairs share the weights of the text and motion encoders.

Method	1 Dancer			2 Dancers			3 Dancers			4 Dancers
Method	FID $\downarrow$	Div $\uparrow$	PFC $\downarrow$	GMR $\downarrow$	GMC $\uparrow$	TIF $\downarrow$	GMR $\downarrow$	GMC $\uparrow$	TIF $\downarrow$	GMR $\downarrow$	GMC $\uparrow$	TIF $\downarrow$
GCD [ 16 ]	39.24	9.64	2.53	34.39	80.32	0.17	30.22	80.22	0.19	36.28	81.82	0.13
CoDancers [ 38 ]	23.98	9.48	3.53	24.55	72.52	0.08	26.34	74.22	0.08	26.44	75.34	0.10
TCDiff [ 3 ]	37.31	14.01	0.51	15.77	81.92	0.12	15.36	82.77	0.11	13.44	81.70	0.15
ST - GDance [ 35 ]	28.87	12.82	0.97	19.42	80.52	0.12	14.76	81.24	0.11	14.02	81.42	0.10
ChainDance*	23.12	13.85	0.50	14.10	81.52	0.11	12.44	82.32	0.11	13.28	82.13	0.12

Table 1 : Single and Group dance performance on AIOZ-GDance. Best results are bold ; second-best are underlined . * denotes models without the global noise optimization penalty.

Method	FLOPs (G) $\downarrow$	Params (M) $\downarrow$	Train/epoch (min:sec) $\downarrow$	Inf Time (min:sec) $\downarrow$
GCD [ 16 ]	27.69	62.16	1:04	0:05
Codancers [ 38 ]	58.95	59.32	0:42	0:02
TCDiff [ 3 ]	22.04	62.48	1:29	0:08
ST - GDance [ 35 ]	6.75	50.21	0:37	0:03
ChainDance*	16.92	15.27	0:14	0:06
ChainDance	16.92	15.27	0:14	0:15

Table 2 : Efficiency comparison (3 dancers group). Best results are bold ; second-best are underlined . * denotes models without the global noise optimization penalty.

Figure 5 : Qualitative results of generated 3-dancers group sequence compared with baselines.

Method	GMR $\downarrow$	GMC $\uparrow$	TIF $\downarrow$
w/o GAME	17.46	79.14	0.19
w/o RATE	14.32	81.98	0.11
w/o Global Penalty	12.44	82.32	0.11
ChainDance	11.63	83.74	0.10

Table 3 : Ablation study for group dance generation (3 dancers). Results are obtained on the Lodge backbone pretrained on FineDance. Bold indicates the best result.

Method	w/o Global Optimization						with Global Optimization
	Single-dance			Group-dance			Single-dance			Group-dance
	FID $\downarrow$	Div $\uparrow$	PFC $\downarrow$	GMR $\downarrow$	GMC $\uparrow$	TIF $\downarrow$	FID $\downarrow$	Div $\uparrow$	PFC $\downarrow$	GMR $\downarrow$	GMC $\uparrow$	TIF $\downarrow$
EDGE (AIST++)	24.74	12.98	1.04	15.23	80.21	0.13	24.74	12.98	1.04	12.32	82.49	0.11
EDGE (FineDance)	26.12	13.34	0.62	14.93	80.73	0.12	26.12	13.34	0.62	12.25	82.65	0.10
Lodge (AIST++)	27.31	14.41	1.01	14.63	80.99	0.12	27.31	14.41	1.01	11.69	83.30	0.11
Lodge (FineDance)	23.12	13.85	0.50	12.44	82.32	0.11	23.12	13.85	0.50	11.63	83.74	0.10

Table 4 : Comparison across backbones and pretraining datasets. Best results are bold , and second-best are underlined .

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Relation (REL)	Decision rule in window $\mathcal{W}$	Artifacts (for prompt)
approach	Monotonic decrease in distance: $\Delta d/W<-0.1$ and $d_{ij}(w_{0}+W-1)\in[d_{\text{near}},\,d_{\text{mid}}]$ .	WHEN, target distance $d^{\star}\!\approx d_{\text{near}}$
keep_distance	Low variance around mid distance: $\mathrm{Var}_{w\in\mathcal{W}}(d_{ij}(w))<0.05$ and $\mathrm{mean}(d_{ij})\in[d_{\text{mid}}-0.2,\,d_{\text{mid}}+0.2]$ .	WHEN, target distance $d^{\star}\!\approx d_{\text{mid}}$
face_to	Mutual facing with proximity: $\cos\!\big(\mathbf{f}_{i},\mathbf{p}_{j}-\mathbf{p}_{i}\big)>0.7$ and $\cos\!\big(\mathbf{f}_{j},\mathbf{p}_{i}-\mathbf{p}_{j}\big)>0.7$ and $d_{ij}<d_{\text{far}}$ .	WHEN
lead_follow	Cross-correlation of motion energies peaks at a positive lag: $\arg\max_{\Delta w\in[2,8]}\mathrm{xcorr}(E_{i},E_{j})=\Delta w>0$ . (Here $i$ leads, $j$ follows.)	WHEN, lag $\Delta w$ ; leader/follower IDs
swap	Start/end positions swap across the window: $\\|\mathbf{p}_{i}(w_{0})-\mathbf{p}_{j}(w_{0}+W-1)\\|_{2}<\epsilon$ and $\\|\mathbf{p}_{j}(w_{0})-\mathbf{p}_{i}(w_{0}+W-1)\\|_{2}<\epsilon$ , while trajectories do not collide.	WHEN
passby	Trajectories intersect near mid/near distance with opposite normal velocities: $\exists\,w\in\mathcal{W}$ s.t. $d_{ij}(w)\in[d_{\text{near}},\,d_{\text{mid}}]$ , $\langle\dot{\mathbf{p}}_{i}^{\perp},\dot{\mathbf{p}}_{j}^{\perp}\rangle<0$ , and distance increases rapidly after intersection.	WHEN

Appendix

Table 5: Window-based relation detection rules for prompt mining. WHEN records the time span(s) in

\mathcal{W}

where the rule is satisfied; artifacts (e.g.,

d^{\star}

, contacted parts, lag

\Delta t

) are injected into prompt templates.

Method	1 dancer			2 dancers			3 dancers			4 dancers			5 dancers
Method	FID $\downarrow$	Div $\uparrow$	PFC $\downarrow$	GMR $\downarrow$	GMC $\uparrow$	TIF $\downarrow$	GMR $\downarrow$	GMC $\uparrow$	TIF $\downarrow$	GMR $\downarrow$	GMC $\uparrow$	TIF $\downarrow$	GMR $\downarrow$	GMC $\uparrow$	TIF $\downarrow$
GCD [ 16 ]	39.24	9.64	2.53	34.39	80.32	0.17	30.22	80.22	0.19	36.28	81.82	0.13	38.43	81.44	0.17
CoDancers [ 38 ]	23.98	9.48	3.53	24.55	72.52	0.08	26.34	74.22	0.08	26.44	75.34	0.10	27.27	74.34	0.11
TCDiff [ 3 ]	37.31	14.01	0.51	15.77	81.92	0.12	15.36	82.77	0.11	13.44	81.70	0.15	14.62	81.40	0.11
ST - GDance [ 35 ]	28.87	12.82	0.97	19.42	80.52	0.12	14.76	81.24	0.11	14.02	81.42	0.10	23.22	80.76	0.12
ChainDance*	23.12	13.85	0.50	14.10	81.52	0.11	12.01	82.68	0.11	13.28	82.13	0.12	15.34	80.93	0.13

Appendix

Table 6 : Unified comparison across 1–5 dancers on the AIOZ - GDance dataset. For all baselines, separate models are trained and evaluated for each group size. In contrast, ChainDance uses a single unified model for all group sizes (1–5 dancers) . Best results are bold ; second-best are underlined . Results with * denote models without the global noise optimization penalty.

Method	FID $\downarrow$	Div $\rightarrow$	PFC $\downarrow$	GMR $\downarrow$	GMC $\uparrow$	TIF $\downarrow$
GT	–	15.67	0.36	–	–	–
Duolando [ 29 ]	12.42	14.35	16.22	14.52	81.32	0.15
GCD [ 16 ]	9.73	14.62	5.12	14.31	81.21	0.13
ChainDance (w/o RATE)	2.97	15.52	4.41	12.21	82.42	0.12

Appendix

Table 7 : Two-dancer generation results on the DD100 dataset. Best results are bold . “

\rightarrow

” indicates that closer to ground-truth (GT) is better.

Method	2			3			4
Method	GMR $\downarrow$	GMC $\uparrow$	TIF $\downarrow$	GMR $\downarrow$	GMC $\uparrow$	TIF $\downarrow$	GMR $\downarrow$	GMC $\uparrow$	TIF $\downarrow$
PINO [ 24 ]	14.10	80.83	0.12	12.69	82.92	0.11	13.83	80.93	0.14
ChainDance	13.67	81.28	0.12	11.63	83.74	0.10	12.59	82.16	0.13

Appendix

Table 8 : Comparison with PINO across different group sizes. Best results are in bold .

Figure 6 : User study results across four perceptual dimensions. ChainDance is consistently preferred over baselines on Realism, Music Correlation, Formation Aesthetics, and Dancer Harmony. The largest improvement is observed on Dancer Harmony (0.80 vs. 0.50 of the best baseline), which jointly captures inter-dancer coordination and per-dancer identity stability—both structurally improved by our sequential decomposition.

Method	Group Dance
Method	GMR $\downarrow$	GMC $\uparrow$	TIF $\downarrow$
ChainDance	11.63	83.74	0.10
Ours-Lodge(AIOZ-GDance)	12.08	82.32	0.11
Ours-Lodge (FineDance&AIST++)	11.97	82.41	0.12
ChainDance (w/o Motion Caption)	11.89	82.52	0.11
ChainDance (w/o bidirectional pairing)	12.93	79.42	0.13

Appendix

Table 9 : Comparison of 3 dancers performance under different ablation settings on the AIOZ - GDance dataset. Best results are bold .

$N$	GMR $\downarrow$		GMC $\uparrow$		TIF $\downarrow$
2	13.67	14.52	81.28	80.92	0.12	0.13
3	11.63	12.44	83.74	82.12	0.12	0.13
4	12.59	14.30	82.16	81.47	0.13	0.13
5	14.73	15.52	81.56	80.64	0.12	0.13

Appendix

Table 10 : Comparison of results with / without the global penalty across group sizes. Format: with penalty / without penalty .

Optimization Steps	Group Dance		Efficiency
Optimization Steps	GMR $\downarrow$	GMC $\uparrow$	Inf Time (min:sec) $\downarrow$
0 $\times$	12.01	82.68	0:06
0.1 $\times$	11.83	82.04	0:08
0.25 $\times$	11.78	81.72	0:10
0.5 $\times$	11.52	82.35	0:11
0.75 $\times$	11.72	82.59	0:13

Appendix

Table 11 : In-depth analysis of the trade-off between inference time and dance quality on the AIOZ-GDance dataset (3 dancers). Best results are bold ; second-best are underlined .

Number of Dancers	Group Dance			Efficiency
Number of Dancers	GMR $\downarrow$	GMC $\uparrow$	TIF $\downarrow$	Inf Time (min:sec) $\downarrow$
2	13.67	81.28	0.12	0:07
3	11.63	83.74	0.10	0:15
4	12.59	82.16	0.13	0:19
5	14.73	81.56	0.12	0:22
6	14.85	80.87	0.13	0:29

Appendix

Table 12 : In-depth analysis of the trade-off between inference time and dance quality on the AIOZ-GDance dataset for more dancers. Best results are bold ; second-best are underlined .

	Matched	Mismatched	Gap $\uparrow$
GT (BAS)	0.2463	–	–
Ours (BAS)	0.2247	0.0624	0.1623
GT (R-precision)	0.5106	–	–
Ours (R-precision)	0.4152	0.1093	0.3059

Appendix

Table 13 : Quantitative results for music and text alignment. Matched uses correctly aligned inputs; Mismatched uses randomly swapped inputs from other categories. A large gap indicates strong responsiveness to the conditioning signal.

Music genre	BAS		R-precision		Group Dance Quality
Music genre	GT	Ours	GT	Ours	GMR $\downarrow$	GMC $\uparrow$	TIF $\downarrow$
Disco	0.2871	0.2661	0.5234	0.4194	13.52	83.03	0.12
Electronic	0.2712	0.2562	0.5076	0.4212	13.21	82.83	0.13
Folk	0.2154	0.1961	0.5112	0.4287	13.84	81.97	0.13
Funk	0.2176	0.1923	0.5154	0.4431	12.87	82.42	0.12
Indian	0.2234	0.2001	0.5223	0.4076	13.48	83.13	0.12

Appendix

Table 14 : Quantitative results across different music genres. ChainDance maintains stable alignment scores (BAS, R-precision) and consistent group dance quality (GMR, GMC, TIF) across all 10 genres, demonstrating robust dual-condition control.

Order	GMR $\downarrow$	GMC $\uparrow$	TIF $\downarrow$
123	11.42	83.53	0.11
132	11.54	82.67	0.10
213	11.14	83.76	0.10
231	11.61	83.21	0.11
312	11.63	84.03	0.11
321	11.73	84.52	0.11

Appendix

Table 15 : Evaluation results for 3-dancer generation under all possible generation orders. The model is evaluated on 100 randomly sampled sequences. Standard deviations across orders are negligible compared to the gap between methods, indicating order-insensitivity.

Run	GMR $\downarrow$	GMC $\uparrow$	TIF $\downarrow$
1	14.87	81.21	0.13
2	14.63	81.63	0.13
3	14.32	81.83	0.13
4	14.21	82.02	0.12
Mean $\pm$ Std	$14.51\pm 0.31$	$81.67\pm 0.35$	$0.128\pm 0.005$

Appendix

Table 16 : Results for 5-dancer generation under 4 different random orders. ChainDance maintains stable performance across orders, confirming that order-insensitivity extends to larger group sizes.

Panel of Figure 7 — Figure 7 : Qualitative visualizations. (a) Ablation study showing the effect of each component. (b) Interaction control under different text prompts.

Figure 8 : ChainDance results on the same music under varying group sizes (from 6 to 10 dancers). The generated choreography remains coordinated and diverse as the group size scales up.

Figure 9 : Qualitative results of ChainDance under different group sizes.

Explore similar work

FlowerDance: MeanFlow for Efficient and Refined 3D Dance Generation
Nov 26, 2025Kaixing Yang, Xulong Tang, Ziqiao Peng +4MeanFlow Human Motion Generation
Renmin University of China · Malou Tech Inc · Wuhan University +1
Text Dictates, Music Decorates: Energy-based Attention for Editable Dance Motion Generation
Jun 22, 2026Seong Jong Yoo, Siyuan Peng, Felix Gu +2Diffusion Transformer Text-to-Motion Generation
University of Maryland, College Park, MD, 20742, USA · Perception and Robotics Group
Dance to Music Generation leveraging Pre-training with Unpaired data and Contrastive Alignment
Jul 12, 2026Ryota Kimura, Sangheon Park, Natalia Polouliakh +1Controllable Music Generation Cross-Modal Alignment
Sony Computer Science Laboratories · Keio University · Georgia Institute of Technology

Nov 26, 2025, cs.CVJ/K move · Enter open · S save

FlowerDance: MeanFlow for Efficient and Refined 3D Dance Generation

Kaixing Yang, Xulong Tang, Ziqiao Peng, Xiangyue Zhang+3

Renmin University of China · Malou Tech Inc · Wuhan University+1

MeanFlow Human Motion Generation

Open paper AskPDF arXiv

Music-to-dance generation aims to translate auditory signals into expressive human motion, with broad applications in virtual reality, choreography, and digital entertainment. Despite promising progress, the limited generation efficiency of existing methods leaves insufficient computational headroom for high-fidelity 3D rendering, thereby constraining the expressiveness of 3D characters during real-world applications. Thus, we propose FlowerDance, which not only generates refined motion with physical plausibility and artistic expressiveness, but also achieves significant generation efficiency on inference speed and memory utilization. Specifically, FlowerDance combines MeanFlow with Physical Consistency Constraints, which enables high-quality motion generation with only a few sampling steps. Moreover, FlowerDance leverages a simple but efficient model architecture with BiMamba-based backbone and Channel-Level Cross-Modal Fusion, which generates dance with efficient non-autoregressive manner. Meanwhile, FlowerDance supports motion editing, enabling users to interactively refine dance sequences. Extensive experiments on AIST++ and FineDance show that FlowerDance achieves state-of-the-art results in both motion quality and generation efficiency. Code will be released upon acceptance.

Abstract

Explore similar work

FlowerDance: MeanFlow for Efficient and Refined 3D Dance Generation

Text Dictates, Music Decorates: Energy-based Attention for Editable Dance Motion Generation

Dance to Music Generation leveraging Pre-training with Unpaired data and Contrastive Alignment