ST-GDance++: A Scalable Spatial-Temporal Diffusion for Long-Duration Group Choreography
Authors: Jing Xu, Weiqiang Wang, Cunjian Chen, Jun Liu, Qiuhong Ke
Organizations: Faculty of Information Technology, Monash University, Melbourne, VIC, 3800, Australia . · School of Computing and Communications, Lancaster University, Lancaster, LA1 4YW, UK .
Group dance generation from music requires synchronizing multiple dancers while maintaining spatial coordination, making it highly relevant to applications such as film production, gaming, and animation. Recent group dance generation models have achieved promising generation quality, but they remain difficult to deploy in interactive scenarios due to bidirectional attention dependencies. As the number of dancers and the sequence length increase, the attention computation required for aligning music conditions with motion sequences grows quadratically, leading to reduced efficiency and increased risk of motion collisions. Effectively modeling dense spatial-temporal interactions is therefore essential, yet existing methods often struggle to capture such complexity, resulting in limited scalability and unstable multi-dancer coordination. To address these challenges, we propose ST-GDance++, a scalable framework that decouples spatial and temporal dependencies to enable efficient and collision-aware group choreography generation. For spatial modeling, we introduce lightweight distance-aware graph convolutions to capture inter-dancer relationships while reducing computational overhead. For temporal modeling, we design a diffusion noise scheduling strategy together with an efficient temporal-aligned attention mask, enabling stream-based generation for long motion sequences and improving scalability in long-duration scenarios. Experiments on the AIOZ-GDance dataset show that ST-GDance++ achieves competitive generation quality with significantly reduced latency compared to existing methods.
Figures & tables
Figure 1: Comparison between vanilla Group Dance Transformer (top) and ours. We decouple the long-duration group dance sequence N×L into spatial and temporal parts to reduce the computational complexity to O(LN2)+O(NL2) . The overall complexity can be reduced into near O(NL) (see Section 4.2.2 ). Also in our framework, the temporal dimension is divided into segments and the noise schedule in training will also change to capture local dependencies and support streaming generation (will be discussed in Section 4.3 ).
Figure 2: Overview of the proposed ST-GDance++ framework for efficient long-duration group dance generation. (a) Overall generation architecture, where noised group motion is processed by the spatial–temporal diffusion decoder to produce the final multi-dancer sequence. (b) Spatial distance-aware graph modeling, which constructs a weighted graph based on pairwise dancer distances and applies a lightweight GCN to capture inter-dancer interactions. (c) Temporal transformer module, responsible for modeling long-range motion dynamics under music and timestep conditions. (d) Alignment attention mask (AAM), which enforces temporally consistent interactions across segments to support stable streaming generation.
Figure 3: Noise schedule comparison. Diffusion Forcing typically samples training schedules with stochastic active windows, leading to a noticeable discrepancy between training and inference. To bridge this gap, we introduce a segment-based triangular schedule that exclusively denoises the current active window before advancing. By applying a consistent noise level to all frames within a single segment, our approach ensures temporal coherence and enhances the stability of long-duration generation.
Method
Group Dance
Single Dance
Efficiency
GMR ↓
GMC ↑
TIF ↓
FID ↓
Div ↑
PFC ↓
FLOPs (G)
Params (M)
Train (min/epoch)
Inf (s)
EDGE
70.26
62.24
0.42
33.24
10.11
3.07
12.50
52.13
0:24
8.6
GCD
30.22
80.22
0.19
39.24
9.64
2.53
27.69
62.16
1:04
4.8
CoDancers
26.34
74.22
0.10
23.98
9.48
3.26
53.95
56.83
0:56
4.5
TCDiff++
11.85
82.04
0.10
22.42
17.43
1.32
32.56
70.42
1:32
9.3
ST-GDance
14.76
81.24
0.11
28.87
12.83
0.97
6.75
50.21
0:37
3.3
Table 1: Performance and efficiency comparison on the short-duration group dance setting. ↓ indicates lower is better, ↑ indicates higher is better. Best and runner-up scores are highlighted.
Method
Group Dance
Single Dance
Efficiency
GMR ↓
GMC ↑
TIF ↓
FID ↓
Div ↑
PFC ↓
FLOPs (G)
Params (M)
Train (min/epoch)
Inf (s)
EDGE
67.24
57.65
0.38
35.40
7.97
3.87
68.21
55.27
2:56
104.6
GCD
40.68
79.25
0.28
51.25
7.24
3.52
92.93
64.42
4:12
92.4
CoDancers
35.20
70.53
0.14
43.62
5.48
4.23
153.70
59.32
5:23
72.3
TCDiff++
14.67
81.64
0.15
22.47
16.23
1.53
168.72
72.64
5:43
124.4
ST-GDance
34.25
74.42
0.15
31.42
14.35
3.21
15.92
51.42
1:31
30.5
Table 2: Performance and efficiency comparison on the long-duration group dance setting. ↓ indicates lower is better, ↑ indicates higher is better. Best and runner-up scores are highlighted.
Method
Dancers
GMR ↓
GMC ↑
TIF ↓
GCD
2
34.39
80.32
0.17
3
30.22
80.22
0.19
4
36.28
81.82
0.13
5
38.43
81.44
0.17
CoDancers
2
24.55
72.52
0.08
3
26.34
74.22
0.10
Table 3: Quantitative comparison of different methods on group dance generation with varying numbers of dancers. 0 0 footnotetext: ↓ indicates lower is better, while ↑ indicates higher is better. Best and runner-up scores are highlighted.
Method
Group Dance
Single Dance
Efficiency
SMB
AAM
TNS
GMR ↓
GMC ↑
TIF ↓
FID ↓
Div ↑
PFC ↓
FLOPs (G)
Params (M)
Train (min/ep)
Inf (s)
✓
✓
12.53
81.49
0.11
24.53
18.02
1.44
11.32
51.64
0:42
4.2
✓
✓
14.32
80.57
0.12
25.52
17.94
1.52
14.63
58.42
0:53
5.2
✓
✓
15.45
79.43
0.13
25.45
18.54
1.38
30.42
68.38
1:27
8.4
ST-GDance++(Full)
10.85
82.83
0.10
25.34
18.76
1.24
11.32
51.64
0:42
4.2
Table 4: Ablation study of ST-GDance++ on different components. SMB, AAM, and TNS denote the Spatial Modeling Block, Alignment Attention Mask, and Triangular Noise Schedule, respectively. Best results are highlighted in bold.
Figure 4: Qualitative results compared with baselines.
Figure 5: Qualitative results generated by ST-GDance++ across varying group sizes. We demonstrate the generated group choreographies with different numbers of dancers, ranging from 2 to 5.
Figure 6: Qualitative comparison under different component ablations of ST-GDance++.
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× fewer parameters and requiring 3-6× less training time compared to prior approaches.
Jing Xu, Cunjian Chen, Qiuhong Ke
Department of Data Science and Artificial Intelligence, Monash University, Australia
Dance is a form of human motion characterized by emotional expression and communication, playing a role in various fields such as music, virtual reality, and content creation. Existing methods for dance generation often fail to adequately capture the inherently sequential, rhythmical, and music-synchronized characteristics of dance. In this paper, we propose MambaDance, a new dance generation approach that leverages a Mamba-based diffusion model. Mamba, well-suited to handling long and autoregressive sequences, is integrated into our two-stage diffusion architecture, substituting off-the-shelf Transformer. Additionally, considering the critical role of musical beats in dance choreography, we propose a Gaussian-based beat representation to explicitly guide the decoding of dance sequences. Experiments on AIST++ and FineDance datasets for each sequence length show that our proposed method effectively generates plausible dance movements while reflecting essential characteristics, consistently from short to long dances, compared to the previous methods. Additional qualitative results and demo videos are available at \small{https://vision3d-lab.github.io/mambadance}.
Sangjune Park, Inhyeok Choi, Donghyeon Soon +2
Ulsan National Institute of Science and Technology, South Korea · Daegu Gyeongbuk Institute of Science and Technology, South Korea
Generating long-duration, high-definition, and rhythmically synchronized dance videos directly from music remains a significant challenge, primarily due to the temporal constraints of current diffusion models, which typically fail beyond 20 seconds. Existing approaches, whether they rely on intermediate 3D skeletons or on end-to-end video synthesis, suffer from temporal drift, identity inconsistency, and repetitive motion patterns when extended to longer horizons. To address these limitations, we propose a novel hierarchical framework for minute-scale coherent music-to-dance generation. Our method decouples the process into global keyframe planning and local temporal refinement, leveraging full-track musical context to ensure long-range coherence. Key innovations include dynamic frame rate adaptation via time-mapped RoPE embeddings for precise alignment, an optical-flow-based loss function to enhance motion continuity, and motion-speed control to preserve high-fidelity details during rapid movements. Extensive experiments demonstrate that our framework surpasses the conventional duration barrier, generating stable, 720p/30fps videos exceeding one minute with superior temporal stability. Furthermore, the model exhibits robust versatility across five distinct dance genres, conditioned on both audio and textual prompts, establishing a new state-of-the-art in coherent, long-form dance video synthesis.