Paper deep dive
ST-GDance++: A Scalable Spatial-Temporal Diffusion for Long-Duration Group Choreography
Jing Xu, Weiqiang Wang, Cunjian Chen, Jun Liu, Qiuhong Ke
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 96%
Last extracted: 3/26/2026, 1:29:15 AM
Summary
ST-GDance++ is a scalable, streaming spatial-temporal diffusion framework designed for long-duration group dance generation. It addresses the quadratic computational complexity of existing transformer-based models by decoupling spatial and temporal dependencies. The framework utilizes lightweight distance-aware graph convolutions for spatial modeling and a combination of a segment-based triangular noise schedule (TNS) and an efficient alignment attention mask (AAM) for temporal modeling, enabling stable, high-fidelity, and efficient streaming generation.
Entities (5)
Relation Signals (3)
ST-GDance++ â evaluatedon â AIOZ-GDance
confidence 100% · Experiments on the AIOZ-GDance dataset show that ST-GDance++ achieves competitive generation quality
ST-GDance++ â uses â Segment-Based Triangular Noise Schedule
confidence 95% · we introduce a segment-based triangular noise schedule (TNS) that enables progressive denoising
ST-GDance++ â uses â Alignment Attention Mask
confidence 95% · we design an alignment attention mask (AAM) inspired by sparse dependency modeling
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract: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.
Tags
Links
- Source: https://arxiv.org/abs/2603.22316v1
- Canonical: https://arxiv.org/abs/2603.22316v1
Trouble viewing inline? Open PDF directly â
Full Text
71,913 characters extracted from source content.
Expand or collapse full text
ST-GDance++: A Scalable SpatialâTemporal Diffusion for Long-Duration Group Choreography Jing Xu 1 , Weiqiang Wang 1* , Cunjian Chen 1* , Jun Liu 2 , Qiuhong Ke 1 1 Faculty of Information Technology, Monash University, Melbourne, VIC, 3800, Australia . 2 School of Computing and Communications, Lancaster University, Lancaster, LA1 4YW, UK . *Corresponding author(s). E-mail(s): weiqiang.wang@monash.edu; cunjian.chen@monash.edu; Contributing authors: jing.xu1@monash.edu; j.liu81@lancaster.ac.uk; qiuhong.ke@monash.edu; Abstract 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 anima- tion. 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 align- ing music conditions with motion sequences grows quadratically, leading to reduced efficiency and increased risk of motion collisions. Effectively modeling dense spatialâtemporal interactions is there- fore 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 computa- tional 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. Project page. Keywords: Music-driven Dance Generation, Group Choreography, SpatialâTemporal Modeling, Diffusion Model 1 Introduction Dance is a highly expressive art form that uses body movements to convey emotion. With the rise of generative AI, music-to-dance generation has found increasing applications in film, cultural pro- ductions, gaming, and beyond (Lee et al. 2021; Valle-P Ìerez et al. 2021; Yao et al. 2023; Liang et al. 2024), assisting artists in crafting immersive expe- riences for both dancers and audiences. To engage 1 arXiv:2603.22316v1 [cs.LG] 20 Mar 2026 the audience, dance must align with the rhythm and emotional tone of the music. Compared to single dance, group dance offers richer visual impact through spatial coordination, synchronized actions, and dynamic formation changes, making choreography significantly more complex. While existing works (Le et al. 2023; Dai et al. 2025; Yang et al. 2024) and datasets (Tsuchida et al. 2019; Le et al. 2023; Zhang et al. 2025) have advanced the field, two major challenges remain for group dance generation. Generation of long sequences. The goal of music-to-dance generation is to synthesize coher- ent motion sequences aligned with an entire musical composition. Short sequences often fail to capture structural progression and emotional dynamics, such as the buildup before a beat drop. To improve scalability, several works adopt hierar- chical or staged strategies that first generate high- level motion structures and then refine them into full sequences (Li et al. 2024b). Lodge (Li et al. 2024b) proposes a two-stage framework for solo dance synthesis by predicting key poses and inter- polating intermediate frames, while Lodge++ (Li et al. 2024a) extends this idea through a coarse- to-fine choreography design, generating global dance primitives and refining them via diffusion. Although such staged designs ease long-horizon modeling, they rely on stage-wise decomposi- tion, which may accumulate errors and introduce inconsistencies between global structure and local motion. Moreover, these methods focus on single- dancer choreography and do not explicitly address multi-dancer spatial coordination. For group dance generation, existing meth- ods often produce partially overlapping motion segments and apply smoothing during stitch- ing to maintain continuity (Tseng et al. 2023). Diffusion-based models such as GCD (Le et al. 2023) and TCDiff (Dai et al. 2025) improve segment consistency during sampling. However, stochastic diffusion sampling may assign incon- sistent spatial identities across segments, leading to abrupt positional shifts when merged (Dai et al. 2026). An alternative is to generate the entire sequence in a single pass. Yet long-range sequence modeling is inherently challenging due to increasing temporal dependencies and computa- tional cost (Zhou et al. 2021; Bulatov et al. 2024). In multi-dancer choreography, complexity grows with both sequence length and dancer count. Most end-to-end approaches rely on diffusion transform- ers, resulting in quadratic complexity (N 2 Ă L 2 ). Efficient long-duration group dance generation therefore remains an open challenge. Multi-dancer interaction modeling. As previously mentioned, the quality of group dance performances depends on precise inter-dancer coordination. Most existing methods (Le et al. 2023; Tseng et al. 2023; Yang et al. 2024) represent group sequences by concatenating body-motion features and 3D trajectories of all dancers into a unified vector, making it difficult to explicitly model spatial correlations across performers. GCD (Le et al. 2023) employs cross-attention to capture inter-dancer relationships, but this mechanism models only coarse spatialâtemporal interactions and may lead to overly uni- form motion patterns. TCDiff (Dai et al. 2025) addresses dancer ambiguity through trajectory- controllable diffusion, yet trajectory generation still lacks explicit spatial correlation modeling, which may result in motion overlap or col- lisions. Its additional dance-beat navigator also increases computational cost. TCDiff++ (Dai et al. 2026) extends this line of work with an end- to-end trajectory-controllable diffusion framework that improves stability for long-duration chore- ography. By incorporating positioning embed- dings and swap-mode encoding, it enhances global temporal coherence. However, it still relies on transformer-based diffusion modules for spa- tialâtemporal modeling, leading to substantial computational overhead and increased inference latency as sequence length grows. Explicit user- level control over multi-dancer spatial coordina- tion also remains limited. Our previous conference version published in BMVC 2025, ST-GDance (Xu et al. 2025), addresses this issue through graph-based spa- tial modeling to capture inter-dancer relation- ships. By combining graph neural networks with a lightweight temporal attention module, it reduces computational complexity from O((N Ă L) 2 ) to nearly O(NL) while producing coherent short sequences. However, although global coordination is pre- served, fine-grained motion detailsâespecially subtle footwork and rapid transitionsâare less accurate. This limitation may stem from the use of linear global attention, which can smooth local 2 PriorView OurView N Dancers L Frames LĂN 2 NĂL 2 Attentionbetween temporaltokens Spatialrelationships betweendancers N 2 ĂL 2 Attentionover onelongsequence L NĂL Frames Frames Decouple Dancers SegmentsS s 2s Fig. 1 Comparison between vanilla Group Dance Transformer (top) and ours. We decouple the long-duration group dance sequenceNĂL into spatial and temporal parts to reduce the computational complexity toO(LN 2 ) +O(NL 2 ). The overall complexity can be reduced into nearO(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). motion dynamics and remove useful details, result- ing in occasional unnatural or jittery motions in longer sequences. This issue negatively affects generation quality. Additionally, ST-GDance may struggle in interactive applications due to its bidi- rectional attention dependencies, which introduce first-token latency under a non-streaming genera- tion paradigm. In summary, to improve interactiv- ity in real-world applications, ST-GDance needs to be adapted into a streaming framework with enhanced motion fidelity. To address these limitations, we propose ST-GDance++, a scalable and streaming spatialâtemporal diffusion framework for long- duration group dance generation. First, inspired by motion generation work (Yin et al. 2025), we introduce a segment-based triangular noise sched- ule (TNS) that enables progressive denoising over long sequences, improving motion continuity and robustness during streaming generation. Second, to balance efficiency and motion fidelity, we design an alignment attention mask (AAM) inspired by sparse dependency modeling (Cai et al. 2025) to focus on locally relevant conditioning signals. An efficient state space model (SSM) is further employed to aggregate global temporal dependen- cies. Together with TNS, this design stabilizes long-term generation while reducing uncertainty accumulation. Third, instead of dense attention over N Ă L tokens, which incurs quadratic com- plexity and ignores explicit spatial coordination, we retain the graph-based spatial modeling block (SMB) from ST-GDance. It encode distance-aware inter-dancer relationships, enabling accurate spa- tial interaction with low computational overhead. Overall, ST-GDance++ generates long, synchro- nized group dance sequences in a streaming man- ner while preserving precise inter-dancer coordi- nation and temporal coherence. Our contributions can be summarized as fol- lows: âą Segment-Based Triangular Noise Sched- ule (TNS). We reformulate the noise sched- ule by introducing segment-wise correlated noise during training. This strategy significantly enhances model robustness and enables seam- less streaming inference, thereby improving the overall practical utility and user experience in real-world applications. 3 âą Efficient Temporally Aligned Attention Mask (AAM). We introduce a masked cross- attention mechanism that restricts each frame to locally relevant conditioning signals, improv- ing motionâcondition alignment while reduc- ing redundant dependencies. This design lowers attention complexity from quadratic to linear time without sacrificing bidirectional informa- tion flow. âą Scalable SpatialâTemporal Framework for Group Dance Modeling. We intro- duce a novel dance decoder that innovatively decouples spatial coordination from temporal dynamics (Figure 1). This decomposition trans- forms quadratic token interactions into an addi- tive spatialâtemporal complexity, enabling both computationally efficient and globally coherent long-duration motion synthesis. âą State-of-the-ArtGroupChoreogra- phy Quality. Extensive experiments on AIOZ-GDance dataset demonstrate that ST- GDance++ scales to longer sequences while achieving superior motion quality and sig- nificantly improved computational efficiency compared with existing group choreography methods. 2 Related Work 2.1 Music-driven Single-Dancer Generation Music-to-dance generation aims to synthesize motion sequences that are both natural and tem- porally coherent with the input music (Joshi and Chakrabarty 2021; Sui et al. 2026). Early studies mainly relied on rule-based or motion- graph techniques (Kovar et al. 2002; Kim et al. 2003; Safonova and Hodgins 2007), while recent work has increasingly adopted deep learning approaches, especially generative models (Zhu et al. 2023; Yin et al. 2023; Sun et al. 2020; Fer- reira et al. 2021). Most existing methods take music features together with previously generated motion as inputs. Although this design enables sequential prediction, it often leads to error accu- mulation over time and limits motion diversity. In addition, many approaches struggle to capture long-range temporal dependencies, which reduces the coherence and novelty of generated sequences. With the emergence of large-scale dance datasets (Lee et al. 2019; Li et al. 2021, 2023), generative modeling has become the dominant paradigm. Auto-regressive methods (Alemi et al. 2017; Yalta et al. 2019) generate motion step by step, but are prone to drift over long sequences. To improve diversity and realism, multimodal gen- erative frameworks based on GANs (Lee et al. 2019), VAEs (Hong et al. 2022; Siyao et al. 2022, 2023), and diffusion models (Tseng et al. 2023) have been proposed. Despite these advances, two major challenges remain: (1) most models are trained on short motion clips, making it diffi- cult to generalize to long-duration sequences; and (2) methods designed for single dancers often fail to handle group scenarios, where both tempo- ral synchronization and spatial coordination are required. 2.2 Music-driven Multi-Dancer Generation Compared with single-dancer motion synthesis, music-driven multi-dancer generation remains rel- atively underexplored. Only a limited number of works (Yao et al. 2023; Le et al. 2024, 2023; Yang et al. 2024; Dai et al. 2025) consider scenarios involving more than two dancers, where both tem- poral synchronization and spatial coordination are essential. Early approaches such as GDanceR (Le et al. 2023) and GCD (Le et al. 2023) extend single- dancer models to group settings, but they do not explicitly address motion representation imbal- ance across dancers, often resulting in ambigu- ous formations. CoDancers (Yang et al. 2024) decomposes group choreography into individ- ual motion streams and progressively introduces dancers, reducing computational cost but poten- tially overlooking global formation consistency. TCDiff (Dai et al. 2025) proposes a trajectory- guided framework that first predicts dancer coor- dinates and then generates motions conditioned on these trajectories, improving spatial consis- tency and reducing collisions. However, its two- stage pipeline separates trajectory prediction from motion synthesis, which may introduce inconsis- tencies in longer sequences. TCDiff++ (Dai et al. 2026) reformulates this design into an end-to- end trajectory-controllable diffusion framework, enhancing stability and temporal coherence for 4 long-duration choreography. Nevertheless, it still relies on transformer-based diffusion modules with bidirectional attention, leading to quadratic com- plexity with respect to sequence length and dancer count. Our previous work, ST-GDance (Xu et al. 2025), adopts a spatialâtemporal graph formu- lation to model inter-dancer interactions and motion dynamics jointly. While it improves coor- dination compared with earlier methods, its scalability for long-duration generation remains limited. In this work, we extend it to ST- GDance++, a scalable spatialâtemporal diffusion framework that explicitly models spatial interac- tions while enabling efficient long-sequence gener- ation, achieving improved coordination and scala- bility in multi-dancer scenarios. 2.3 Streaming Generation with Diffusion Model Streaming generation produces sequences progres- sively, allowing the model to adjust outputs when new control signals arrive instead of synthesiz- ing the entire sequence at once (Zhang et al. 2025; Kodaira et al. 2025; Jiang et al. 2025). This enables faster feedback and more practical interactive creation. Recent studies incorporate streaming into dif- fusion models via diffusion forcing, assigning dif- ferent noise levels to frames or tokens for flexible denoising (Benoit 2025). Subsequent work miti- gates trainâtest mismatch through explicit roll- out strategies such as self forcing and rolling forcing (Huang et al. 2025; Liu et al. 2025). Other methods enforce causal or monotonic timestep constraints to preserve temporal consis- tency (Chen et al. 2025; Sun et al. 2025). However, most approaches are designed for video generation with spatially large 2D inputs. Motion synthesis, by contrast, is temporally one-dimensional and often driven by rapidly changing controls, making direct adaptation suboptimal. Streaming diffusion has recently been extended to motion generation. FloodDiffusion (Cai et al. 2025) proposes a tailored diffusion-forcing strat- egy for text-driven streaming motion synthesis. Yet existing methods primarily address single- agent settings. Group dance generation introduces additional challenges, including spatial coordina- tion and synchronized transitions among multi- ple dancers. Our work adapts streaming diffu- sion to multi-dancer choreography, enabling long- duration generation while maintaining coherent spatial structures. 3 Preliminary Knowledge 3.1 Problem Definition Given an input music sequence M = m l L l=1 , the objective of group dance generation is to synthesize a corresponding multi-dancer motion sequence X =x l L l=1 , where l denotes the frame index and L is the sequence length. Each frame x l contains the poses of all dancers, defined as x l = x l,n N n=1 , where x l,n represents the pose of the n-th dancer at frame l, and N is the total number of dancers. Accordingly, the motion sequence of the n-th dancer can be written as X n = x l,n L l=1 . Each dancer pose is represented by a 151-dimensional vector, including 24-joint SMPL (Loper et al. 2023) with 6D rotations (Zhou et al. 2019), binary contact indicators, and 3D root position (McFee et al. 2015). Therefore, the group motion sequence can be represented as a tensor X â R LĂNĂ151 . Similarly, the music sequence is encoded into frame-level feature vectors using the audio encoder, resulting in a music feature tensor M â R LĂd m where d m denotes the dimensional- ity of the music feature from Librosa (McFee et al. 2015). Under the diffusion formulation, we denote x t as the noisy motion at diffusion timestep t, and the model is trained to progressively denoise x t toward the clean motion sequence conditioned on the music input M. 3.2 Diffusion for Group Dance Generation Diffusion models have demonstrated strong capa- bilities in high-dimensional generative tasks such as images (Saharia et al. 2022), audio (Guo et al. 2024), and videos (Qi et al. 2023). Fol- lowing the denoising diffusion probabilistic model (DDPM) (Ho et al. 2020) and its application to dance generations (Tseng et al. 2023; Le et al. 2023; Dai et al. 2025, 2026), we adopt a con- ditional diffusion framework for group motion synthesis. 5 The diffusion process constructs a Markov chain that gradually perturbs the clean motion sequence x 0 into a Gaussian noise distribution through T steps. The forward process is defined as: q(x t |x tâ1 ) =N (x t ; p 1â ÎČ t x tâ1 ,ÎČ t I),(1) where x t denotes the noisy sample at timestep t, and ÎČ t controls the variance of the injected noise. During generation, the reverse process aims to recover the clean motion from noise. The reverse transition is modeled as: p(x tâ1 |x t ) =N (x tâ1 ; 1 â 1â ÎČ t (x t â ÎČ t ÎŒ t ),Ï 2 t I), (2) where ÎŒ t is the model-predicted mean and Ï 2 t denotes the variance of the reverse process. To generate group dance sequences condi- tioned on music and spatial context, we extend the diffusion model into a conditional formulation. Given the music conditionM and the spatial con- ditionS, the denoising network D learns to reverse the diffusion trajectory: x tâ1 = D(x t ,t,M,S),(3) where the network predicts the clean motion or its equivalent parameterization at each timestep. In this way, the model progressively transforms random noise into a coherent multi-dancer motion sequence aligned with the input music. 3.3 State Space Model (SSM) State Space Models (SSMs) have recently shown strong capability in modeling long-range sequen- tial data due to their linear-time complexity and effective global context aggregation (Gu et al. 2022; Weng et al. 2024). Compared with conven- tional attention-based architectures, SSMs pro- vide a more scalable solution for long-duration motion generation due to its superior feature selection ability. Given an input sequence x(t) â R, an SSM transforms it into an output y(t) â R through a latent state h(t) â R N governed by a linear dynamical system: h âČ (t) = Ah(t) + Bx(t), y(t) = Ch(t),(4) where A â R NĂN controls the state transition, and B â R NĂ1 and C â R 1ĂN denote the input and output projection matrices, respectively. To enable efficient computation, we adopt the Mamba-style discretization with zero-order hold (ZOH) (Gu and Dao 2024). The continuous system is converted into a discrete form: Ì A = exp(âA), Ì B = (âA) â1 (exp(âA)â I)âB, (5) where â is an input-dependent step size that determines the temporal resolution. Smaller â focuses on short-term variations, while larger â captures longer-range dependencies. The dis- cretized system can be expressed as a convolution with a structured kernel: Ì K = ( Ì C Ì B, Ì C Ì A Ì B,..., Ì C Ì A Lâ1 Ì B), y = xâ Ì K, (6) where L denotes the sequence length. This formu- lation enables the model to efficiently aggregate information across the entire temporal horizon. In our framework, the SSM backbone serves as the temporal modeling module, capturing long- range motion dynamics while maintaining linear complexity with respect to sequence length O(L). This property is particularly beneficial for long- duration group dance generation. 4 Methodology In this section, we introduce ST-GDance++, a spatialâtemporal diffusion framework for scal- able and high-precision generation of long- duration group dance sequences, as illustrated in Fig. 2. Unlike conventional approaches that treat dancerâframe tokens as a single sequence, we explicitly decouple the dancer and tem- poral dimensions, modeling spatial interactions and temporal dynamics with dedicated mod- ules. The spatial modeling block (SMB) employs a distance-aware graph convolutional net- work to capture inter-dancer relationships through distance-aware graph interactions. The temporal branch focuses on motion evolution over time. To support stable long-duration generation, we fur- ther introduce an Alignment Attention Mask (AAM) and a Segment-based Triangular 6 NoisedGroup DanceSequence Linear Group-fusion (b)Spatial Modeling c a b w ab w bc w ac GCN íș Weighted DistanceGraph Linear Motion Feature Top-kEdges (íŽ,í) toofar toonear Spatial Modeling Temporal Transformer Dance Decoder Linear KĂ (c)Temporal Transformer ConcatSquashLinear Linear Foot correction WĂ â í„" ! GeneratedMulti-DancerSequence FiLM Feedforward FiLM Linear SumAndGate Rearrange FiLM Self-Attention (d)Cross- Attention Linear Linear Conv1D SSM ... MLP Jukebox Encoder Audio Condition Diffusion Timestep í 0 1 2 1 2 1 x y Linear 012 0 1 2 012 0 1 2 SwapType (d)AlignmentAttention Mask(AAM) ... ... z í„ " ! í„ " # í„ " $%# ... (a)GenerationArchitecture x y x y z TemporalSegments â Fig. 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 spatialâtemporal 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. Noise Schedule (TNS), which improve tempo- ral consistency and robustness during diffusion. 4.1 Group Fusion Projection To reduce dancer ambiguity and encourage group- level interaction, we use a lightweight group fusion module following previous work (Dai et al. 2025). Instead of processing each dancer independently, the features of all dancers are first concatenated along the channel dimension to form a shared rep- resentation because the data in high-dimensional feature spaces can be more easily differentiated. Specifically, the input group feature tensor X â R LĂNĂd , representing N dancers with sequence length L and feature dimension d, is first reshaped into X âČ â R LĂ(NĂd) . This fused feature is then projected with a multi-layer perceptron (MLP) to produce higher-dimensional latent space fea- ture for all dancers. Finally, the features are reshaped back into dancer-specific representations for subsequent spatialâtemporal modeling. 4.2 Spatial-temporal Group Dance Decoder As shown in Fig 1, we decouple multi-dancer gen- eration task into spatial and temporal dynamics and use different blocks in designed dance decoder to model separately in Fig. 2. 4.2.1 Spatial Distance-aware Graph Convolutional Network In group dance generation, each frame contains two essential factors: the interactions among dancers and the global formation of the group. Many existing approaches primarily focus on tem- poral motion synthesis, while spatial interactions between dancers are only implicitly modeled or even neglected. This often leads to ambiguous spa- tial relationships or motion collisions in crowded scenes. To explicitly capture intra-group spatial dependencies, we represent the dancers at each frame as nodes in a fully connected weighted graph as shown in Fig. 2 (b). The edge weights 7 are determined by pairwise Euclidean distances, allowing the model to encode distance-aware inter- actions. Let N denote the number of dancers. The adjacency matrix Aâ R NĂN is defined as: A ij = 1 â„p i â p j â„ + Δ ,(7) where p i and p j are the 2D coordinates of dancers i and j, and Δ is a small constant that prevents numerical instability. This formulation assigns larger weights to closer dancers, encour- aging stronger interaction modeling in spatially relevant regions. Given the latent motion feature sequence X group â R LĂNĂd after group fusion projection, where L is the sequence length and d is the feature dimension, spatial features are propagated using a graph convolutional layer: H (l+1) = ReLU Ì AH (l) W (l) ,(8) where Ì A = D â1/2 AD â1/2 is the normalized adjacency matrix and W (l) is a learnable transfor- mation. To control computational cost, we apply a top- k edge selection strategy, keeping only the most significant connections for each node. This reduces the per-layer complexity to: O GCN = O(LNk)â O(LN ),(9) where k is a fixed proportion of retained edges. In addition, extremely close connections are masked to prevent overly strong correlations that may lead to spatial collisions. By introducing this lightweight distance-aware GCN, spatial interactions are modeled explic- itly without relying on dense cross-attention over all dancerâframe tokens. This design reduces computational redundancy, improves scalability, and enables more stable and collision-free group motion generation. 4.2.2 Temporal Transformer The temporal module models long-range motion dynamics along the time dimension, as illustrated in Fig. 2(c). Self-attention. To capture long-range tempo- ral dependencies, we adopt Differential Attention from the DIFF Transformer (Ye et al. 2025) as the temporal self-attention mechanism. Unlike stan- dard multi-head attention, this design computes attention as the difference between two parallel attention branches, which enhances discriminative information filtering. The module is illustrated in the temporal decoder (Fig. 2(c)). Given a single dancer feature sequence X â R LĂd from spatially processed group feature X GCN â R LĂNĂd , we obtain two sets of queryâkey pairs by splitting the channel dimen- sion of linear projections: [Q 1 ;Q 2 ] = XW Q , [K 1 ;K 2 ] = XW K ,V = XW V , (10) where XW Q ,XW K âR LĂ2d ,and Q 1 ,Q 2 ,K 1 ,K 2 â R LĂd are obtained by evenly dividing the last dimension, while V â R LĂ2d . The two attention maps are computed as A 1 = softmax Q 1 K †1 â d ,A 2 = softmax Q 2 K †2 â d , (11) and the differential attention output is defined as DiffAttn(X) = (A 1 â λA 2 )V,(12) where λ is a learnable suppression coefficient that balances the two branches. Because the projections for the two branches share the same linear transformations, the pro- jection cost remains O(Ld 2 ) and does not domi- nate computation. Forming the attention matrices Q 1 K †1 and Q 2 K †2 requires O(L 2 d) operations in theory. However, the difference matrix A 1 âλA 2 is empirically sparseâmost entries are close to zero and are pruned during implementationâso each query retains at most s significant responses with sâȘ L. As a result, the effective number of queryâkey interactions reduces from L 2 to Ls, leading to a per-layer complexity of O(Lsd), which further simplifies to O(Ld) when s is treated as a constant. Therefore, the temporal modeling scales linearly with respect to the sequence length. Cross-attention with alignment mask. To incorporate music conditions, we apply cross- attention between motion tokens and music fea- tures, as shown in Fig. 2(d). A key challenge in long-duration and streaming generation is that dense cross-attention allows each motion 8 frame to attend to the entire music sequence, which may introduce temporal inconsistency and cause abrupt motion changes when the sequence is processed segment by segment. To address this issue, we introduce an Alignment Atten- tion Mask (AAM) that restricts each frame to attend only to a temporally aligned local music window, enforcing consistent musicâmotion corre- spondence across segments. Let M â R LĂd denote the music features. The cross-attention is formulated as: CA(X,M ) = Softmax QK †m + A â d V m ,(13) where Q = XW m Q â R LĂd , K m = MW m K â R LĂd , V m = MW m V â R LĂd , and A â R LĂL is the Alignment Attention Mask (AAM). This design provides two advantages. First, it enforces temporally consistent musicâmotion alignment, reducing abrupt transitions when generating long sequences. Second, it significantly reduces compu- tational cost. Assuming each frame attends only to a local window of size w, the complexity becomes: O cross = O(Lwd),(14) which scales linearly with the sequence length when w is fixed and w âȘ L. SSM-enhanced temporal modeling. To efficiently aggregate global context without re- introducing quadratic overhead, the features are subsequently processed by an SSM block (intro- duced in Sec. 3.3) acting as a lightweight temporal backbone: X SSM = SSM(X),(15) with complexity linear in the sequence length: O SSM = O(Ld).(16) Swap mode and rearrangement. Following TCDiff++(Dai et al. 2026), we explicitly model dancer permutation within each temporal segment via a swap-mode condition S, and use a canon- ical rearrangement to stabilize dancer identities across samples. Given the clean group motionX = x l L l=1 , we first rearrange dancers in the initial frame by sorting them according to their hori- zontal positions (x-axis), so that dancer indices are consistent across the dataset (i.e., dancers are ordered from left to right in the first frame). The diffusion process is then applied to the rearranged motion to obtain the noisy input x t . To describe whether dancers exchange their relative spatial orders within a segment, we com- pute the swap type by tracking the ordering changes between the start and end frames of each segment. Concretely, we assign dancer IDs accord- ing to their initial spatial ranks (left-to-right and top-to-bottom, as illustrated in Fig. 2(c)). At the end of the segment, we re-rank dancers along the x- and y-axes and record the result- ing index orders, which are concatenated into an index sequence. This index sequence captures the permutation pattern within the segment and is further embedded into a high-dimensional vector S â R d , which is fused with other conditions (e.g., music features and diffusion timestep) to guide denoising. Overallcomplexity.Combiningself- attention, cross-attention, and the SSM block, the overall temporal complexity is: O temporal = O(Lsd) + O(Lwd) + O(Ld)â O(Ld), (17) which scales linearly with the sequence length and enables efficient long-duration group dance generation. 4.3 Segment-based Triangular Noise Schedule (TNS) To bridge the discrepancy between training and inference phases, we reformulate the diffusion process using a segment-based triangular noise schedule inspired by (Cai et al. 2025), as shown in Fig. 3. Unlike vanilla Diffusion Forcing that assigns stochastic noise levels to individual frames, our approach ensures temporal coherence by pro- cessing the sequence in structured segments. We divide the length L group dance sequence x into s segments, where each segment x s consists of a fixed number of frames. During training, for a sampled global timestep t â [1,T ], the noised segment Ì x s is formulated as: Ì x s = â Ì Î± t x s + â 1â Ì Î± t Δ, ΔâŒN (0, I)(18) Crucially, all frames within the segment x s share a consistent noise level â Ì Î± t . As illustrated in the triangular schedule, the model is trained to 9 Noise level Denoise timestep Training GroupDanceDiffusion (Vanilla) Dancers í„" ! ... Frames â Segments â í„" " í„" # í„" $ í„" ! ... í„" " í„" # í„" $ í„" ! ... í„" " í„" # í„" % í„" ! ... í„" " í„" # í„" % â â â â â â â â â GroupDanceDiffusion (Ours) Fig. 3 Noise schedule comparison. Diffusion Forcing typi- cally samples training schedules with stochastic active win- dows, 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 con- sistent noise level to all frames within a single segment, our approach ensures temporal coherence and enhances the stability of long-duration generation. denoise the current active window while condi- tioned on the historical context x <s . This design offers two key advantages for long-term stabil- ity: (1) By applying a uniform noise level within each segment x s , the model maintains tem- poral invariance and avoids learning spurious frame-to-frame noise transitions, which effectively enhances local motion smoothness; (2) The tri- angular progression provides error robustness by explicitly simulating the auto-regressive infer- ence process, thereby training the model to predict coherent subsequent segments from varying his- torical noise levels and significantly mitigating the accumulation of drifting errors during extended generation. 4.4 Training Losses Our ST-GDance++ is trained in an end-to-end manner under a conditional diffusion framework. Following previous works (Tseng et al. 2023; Dai et al. 2026), the objective combines the standard diffusion reconstruction loss with several motion- level constraints to ensure temporal coherence, physical plausibility, and stable group interac- tions. Diffusion Reconstruction loss. We adopt the standard diffusion loss to encourage the denoised motion to match the ground-truth sequence: L simple = E x,t â„x 0 â D(x t ,t,M,S)â„ 2 2 ,(19) where x 0 denotes the clean motion, x t is the noised motion at timestep t, M is the music condition, and D(·) is the denoising network. Joint Velocity loss. To enforce temporal smoothness between consecutive frames, we intro- duce a velocity loss: L vel = 1 Lâ 1 Lâ1 X l=1 (x l+1 0 â x l 0 )â ( Ë x l+1 0 â Ë x l 0 ) 2 2 . (20) Forward Kinematics loss. To preserve the spatial consistency of the skeletal structure, we apply a forward kinematics constraint: L FK = 1 L L X l=1 FK(x l 0 )â FK( Ë x l 0 ) 2 2 ,(21) where FK(·) denotes the forward kinematics func- tion. Foot Contact loss. To improve physical real- ism and reduce foot sliding, we employ a foot contact constraint: L contact = 1 Lâ 1 Lâ1 X l=1 FK( Ë x l+1 0 )â FK( Ë x l 0 ) · Ë d l 2 2 , (22) where Ë d l indicates the predicted contact state. Distance consistency loss. To preserve sta- ble spatial relationships among dancers, Dai et al. (2026) introduce a distance-consistency loss that penalizes deviations in pairwise distances between predicted and ground-truth motions: L D = 1 N â 1 L X l=1 N 2 · p (l),i â p (l),j â Ë p (l),i â Ë p (l),j 2 2 . (23) where p (l),i and Ë p (l),i denote the ground-truth and predicted positions of dancer i at frame l, respectively. This loss encourages the model to maintain consistent inter-dancer distances, reducing colli- sions and preserving group formations. 10 Overall objective. The final training loss is a weighted combination of the above terms: Loss = λ simple L simple + λ FK L FK + λ vel L vel + λ contact L contact + λ D L D . (24) where λ simple , λ FK , λ vel , λ contact , and λ D are balancing coefficients. 5 Experiments 5.1 Experimental Settings Implementation Details. All experiments are implemented in PyTorch and conducted on NVIDIA RTX 4090 and V100 GPUs. For fair com- parison, all efficiency evaluations are performed on RTX 4090. Following previous work (Dai et al. 2026), we set the loss weights as λ simple = 0.636, λ vel = 2.964, λ FK = 0.646, λ D = 100, and λ contact = 10.942. The batch sizes are 64, 32, 24, and 8 per GPU for 2, 3, 4, and 5 dancers, respec- tively. We optimize the model using the Adam optimizer with a learning rate of 5Ă 10 â5 . Dataset.Weconductexperimentson the open-source group dance dataset AIOZ- GDance (Le et al. 2023), which provides 6.7 hours of group dance videos with synchronized music and 3D motion data, covering 4,000 dancers, 7 dance styles, and 16 music genres. All videos are decoded at 30 FPS. Following the protocol of Li et al. (Le et al. 2023), we split the dataset into training, validation, and test sets with a ratio of 8:1:1. Baselines.WeevaluateST-GDance++ against the following five state-of-the-art group and single dance generation method: âą EDGE (Tseng et al. 2023) is a representative and publicly available single-dancer generation model. For fair comparison, we re-train it on the AIOZ-GDance dataset to evaluate its general- ization to group dance settings. âą GCD (Le et al. 2023) incorporates contrastive learning into a diffusion framework to model dancerâgroup associations and enables control- lable trade-offs between motion consistency and diversity. âą CoDancers (Yang et al. 2024) is a retrieval- based group dance generation framework that models choreographic units and explicitly decomposes spatialâtemporal coherence into individual movements and group interactions. âą TCDiff++ (Dai et al. 2026) adopts a diffusion- based framework with enhanced temporal coherence modeling to generate stable long- duration dance sequences. âą ST-GDance (Xu et al. 2025) is our previous work, which adopts a spatialâtemporal decou- pling strategy with graph-based interaction modeling for efficient and collision-aware group choreography generation. Metrics. For evaluation, we compute both single- and group-level metrics to comprehen- sively assess the performance of our method: For group evaluation (Le et al. 2023), (1) Group Motion Realism (GMR) quantifies the realism of group formations by computing the Frechet Inception Distance (FID) between generated and real motion features. (2) Group Motion Correla- tion (GMC) quantifies group coherence by com- puting the cross-correlation between the motion sequences of generated dancers. (3) Trajectory Intersection Frequency (TIF) quantifies collision frequency by counting the number of trajec- tory intersections among dancers. For single- dancer generation, (4) Frechet Inception Distance (FID) (Li et al. 2021; Huang et al. 2020) mea- sures the distribution distance between generated and real motion features. (5) Generation Diver- sity (Div) (Li et al. 2021; Huang et al. 2020) quantifies motion diversity based on kinematic features. In addition, to evaluate computational efficiency, we report both device-dependent and device-independent metrics, including floating- point operations (FLOPs), parameters count, training time per epoch, and inference time. 5.2 Quantitative Comparison Evaluating short-duration scenarios Table 1 presents the quantitative comparison under the short-duration setting with 120 frames. On the group dance task, our method achieves the best performance across two core metrics, yielding 10.85 GMR and 82.83 GMC. Compared with existing baselines, this demonstrates more sta- ble motion reconstruction and better inter-dancer coordination. In particular, relative to TCDiff++, we further reduce GMR from 11.85 to 10.85 and improve GMC from 82.04 to 82.83, while main- taining the same low TIF, indicating fewer abrupt 11 Table 1 Performance and efficiency comparison on the short-duration group dance setting. â indicates lower is better, â indicates higher is better. Best and runner-upscores are highlighted. Method Group DanceSingle DanceEfficiency GMRâ GMCâ TIFâ FIDâDivâ PFCâ FLOPs (G) Params (M) Train (min/epoch) Inf (s) EDGE70.2662.240.4233.2410.113.0712.5052.130:248.6 GCD30.2280.220.1939.249.642.5327.6962.161:044.8 CoDancers26.3474.22 0.1023.98 9.483.2653.9556.830:564.5 TCDiff++11.8582.040.10 22.42 17.431.3232.5670.421:329.3 ST-GDance14.7681.240.1128.8712.83 0.97 6.75 50.210:373.3 ST-GDance++ 10.85 82.830.1125.34 18.761.2411.3251.640:424.2 swaps and collisions during generation. Consis- tent improvements are also observed in single- dancer evaluation. Our model achieves the highest motion diversity, outperforming all baselines by a clear margin, while preserving competitive real- ism with an FID of 25.34 and a low PFC of 1.24. These results suggest that the generated sequences exhibit both richer dynamics and stable temporal consistency. Overall, our method per- forms favorably on most group-level coordination metrics and remains competitive on single-dancer quality measures. Evaluating long-duration scenarios. To assess robustness under long-horizon generation, we extend the sequence length to 720 frames and report quantitative results in Table 2. As the generation horizon increases, all baselines exhibit noticeable performance degradation, man- ifested by reduced group coherence, accumulated temporal drift, and increased collision frequency. These issues become particularly severe when spatial coordination or temporal consistency is not explicitly modeled. Single-dancer diffusion methods such as EDGE and GCD, which lack explicit spatial interaction modeling, tend to pro- duce inconsistent dancer layouts over time, lead- ing to frequent position swaps and degraded group formation realism. Although EDGE main- tains relatively stable individual motion qual- ity, the absence of global structural constraints results in more collisions and poorer group- level metrics. CoDancers focuses primarily on individual-level synthesis without sufficiently cap- turing inter-dancer dependencies, which weakens global coordination. ST-GDance improves tempo- ral smoothness through spatialâtemporal factor- ization, yet its autoregressive decoding accumu- lates errors over long sequences, causing gradual drift and reduced stability. TCDiff++ achieves stronger coherence by jointly modeling positions and motions in an end-to-end manner; however, its dense temporal modeling introduces higher computational complexity and still struggles to fully eliminate long-horizon inconsistencies. In contrast, our ST-GDance++ consistently achieves the best or second-best performance across nearly all metrics in long-duration settings. By explic- itly decoupling spatial and temporal modeling and adopting segment-wise parallel diffusion, our approach preserves intra-group consistency while preventing long-term error accumulation. Con- sequently, it produces more coherent formations (lower GMR), stronger inter-dancer correlation (higher GMC), fewer collisions (lower TIF), and richer motion diversity (higher Div), demonstrat- ing superior robustness and scalability for long- duration group choreography generation. Efficiency Analysis. As shown in Tables 1 and 2, methods built on standard Transformer- style sequence modeling exhibit a clear scaling bottleneck when extending the generation hori- zon. This is most evident for TCDiff++ and GCD: their FLOPs increase sharply from the short setting to the long setting (e.g., TCDiff++ from 32.56 to 168.72 GFLOPs; GCD from 27.69 to 92.93 GFLOPs), and the training/inference latency grows accordingly (TCDiff++ reaches 5:43 per epoch and 124.4 s for inference in the long setting). Such growth is consistent with the quadratic attention cost with respect to sequence length, making long-duration generation substan- tially more expensive even when batch size and dancer count remain unchanged. CoDancers, in contrast, follows a VAE-style formulation and benefits from relatively faster inference in the short setting, but its overall cost still scales aggressively in the long setting (53.95 to 153.70 GFLOPs), and its efficiency advantage is less stable once the sequence becomes long. This suggests that while VAE-based designs can reduce 12 Table 2 Performance and efficiency comparison on the long-duration group dance setting. â indicates lower is better, â indicates higher is better. Best and runner-upscores are highlighted. Method Group DanceSingle DanceEfficiency GMRâ GMCâ TIFâ FIDâDivâ PFCâ FLOPs (G) Params (M) Train (min/epoch) Inf (s) EDGE67.2457.650.3835.407.973.8768.2155.272:56104.6 GCD40.6879.250.2851.257.243.5292.9364.424:1292.4 CoDancers35.2070.53 0.1443.625.484.23153.7059.325:2372.3 TCDiff++14.67 81.640.1522.4716.231.53168.7272.645:43124.4 ST-GDance34.2574.420.1531.4214.353.21 15.92 51.421:3130.5 ST-GDance++ 14.32 82.47 0.14 20.37 16.192.3229.3252.831:4532.5 Table 3 Quantitative comparison of different methods on group dance generation with varying numbers of dancers. MethodDancersGMRâGMCâTIFâ GCD 234.3980.320.17 330.2280.220.19 436.2881.820.13 538.4381.440.17 CoDancers 224.5572.52 0.08 326.3474.22 0.10 426.4475.34 0.10 527.2774.34 0.11 TCDiff++ 214.5281.460.08 311.8582.040.11 412.4282.920.11 512.9581.340.13 ST-GDance 219.4280.520.12 314.7681.240.11 414.0281.42 0.10 523.2280.760.12 ST-GDance++ 213.01 81.520.11 310.92 82.83 0.10 412.0682.460.12 512.75 81.860.12 â indicates lower is better, while â indicates higher is better. Best and runner-upscores are highlighted. per-step sampling overhead, they may still suf- fer from heavy temporal decoding and reduced scalability when the target horizon is extended. ST-GDance and ST-GDance++ are designed specifically to improve long-horizon efficiency within the diffusion framework. Comparing ST- GDance and ST-GDance++ in Tables 1 and 2 shows a clear trade-off. ST-GDance is highly efficient, requiring only 6.75 GFLOPs and 0:37 training time in the short setting, and remaining the most lightweight method in the long setting. However, its generation quality is slightly behind the best-performing diffusion models. But ST- GDance++ improves these metrics while keeping the model compact. It achieves the best group coordination scores in both short and long set- tings, and maintains strong diversity. At the same time, its computational cost remains moderate, with 29.32 GFLOPs and 1:45 training time in the long setting, far lower than TCDiff++. This indicates that ST-GDance++ achieves stronger generation quality while still maintaining high efficiency. Comparison of Multi-Dancer Genera- tion. To evaluate the robustness of our model in handling different crowd densities, we report the quantitative results for varying numbers of dancers (from 2 to 5) in Table 3. Our proposed ST-GDance++ consistently achieves the state- of-the-art performance across nearly all metrics and configurations. Notably, as the complexity of group dynamics increases with the number of dancers, baseline methods such as CoDancers exhibit a significant performance degradation; for instance, its GMR increases sharply from 24.55 to 27.27 when scaling from 2 to 5 dancers. In contrast, our model maintains a remarkably sta- ble and superior GMR (12.75 for N = 5), demonstrating its effective modeling of com- plex spatio-temporal dependencies among multi- ple agents. Furthermore, ST-GDance++ leads in Group Motion Consistency (GMC) across all set- tings (e.g., reaching 82.83 for N = 3), which underscores its capability to generate highly syn- chronized and harmonious group choreographies. While TCDiff++ and CoDancers show competi- tive results in Trajectory Intersection Frequency (TIF) in some cases, they often do so at the cost of motion fidelity or coordination. Our approach strikes a superior balance, achieving the best GMR and GMC while maintaining a comparable collision avoidance rate. 13 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. MethodGroup DanceSingle DanceEfficiency SMBAAMTNSGMRâGMCâTIFâFIDâDivâPFCâ FLOPs (G) Params (M) Train (min/ep) Inf (s) â12.5381.490.1124.5318.021.4411.3251.640:424.2 â14.3280.570.1225.5217.941.5214.6358.420:535.2 â15.4579.430.1325.4518.541.3830.4268.381:278.4 ST-GDance++(Full)10.85 82.83 0.1025.34 18.76 1.2411.3251.640:424.2 5.3 Ablation Study Impact of Triangular Noise Schedule (TNS) The Triangular Noise Schedule (TNS) primar- ily affects the quality of generated sequences. As shown in Table 4, when TNS is applied, both group dance and single dance generation per- formance improve. Specifically, the GMR score decrease from 12.53 to 10.85, which represents a 13.4% reduction. Additionally, diversity (Div) increases from 18.02 to 18.76, marking a 4.1% improvement. During training, we apply a vary- ing noise schedule for different segments of the sequence, which introduces more complexity and variability, helping the model learn to handle a wider range of noisy scenarios. In contrast, dur- ing testing, we use the traditional uniform noise addition method, which simplifies the noise sched- ule. This discrepancy makes the training process more challenging, but it also increases the modelâs robustness, as it is exposed to more diverse noise patterns. Since the introduction of TNS does not modify the network architecture, it does not introduce any additional computational overhead during either training or inference. Therefore, the efficiency-related metrics remain unchanged. The increased robustness during training trans- lates into more stable and higher-quality outputs during testing, without affecting the efficiency of inference. Impact of Spatial Modeling Block (SMB) The inclusion of the Spatial Modeling Block (SMB) significantly improves the perfor- mance of group dance generation. As shown in the table, when SMB is added, the GMR decreases from 15.45 to 10.85, which is a 29.8% reduc- tion, indicating that the group motion generation becomes more coherent. Similarly, the TIF score reduces from 0.13 to 0.10, showing a 23.1% reduc- tion in trajectory intersection frequency. These improvements highlight the ability of SMB to enhance spatial coordination among dancers in group settings. However, for single dance genera- tion, the impact of SMB is minimal. Thus, SMB primarily benefits group dance generation. Impact of Alignment Attention Mask (AAM) The use of the Alignment Attention Mask (AAM) improves both group and single dance generation quality, with a more noticeable effect in group scenarios. As shown in Table 4, configurations with AAM consistently achieve bet- ter motion consistency and alignment, leading to improved group coordination metrics. In the sin- gle dance setting, AAM also contributes to lower FID and more stable motion patterns. In terms of efficiency, AAM also helps to decrease model complexity by removing future attention calcula- tions. The training time per epoch decrease from 0:53 to 0:42, corresponding to a 20.8% decrease, and the number of parameters decrease from 58.42 M to 51.64 M. Overall, these results show that AAM improves efficiency while retaining motion alignment and generation quality. 5.4 Qualitative Visual Comparison Baseline comparison. We qualitatively com- pare ST-GDance++ with representative baselines, including GCD and the original ST-GDance. As shown in Fig. 4, each row presents the generated motion sequences of different methods under the same music condition. GCD tends to produce noticeable foot pen- etration artifacts, where the feet unrealistically intersect with the ground or exhibit abrupt posi- tional shifts. This issue becomes more apparent in sequences involving fast or complex steps, indicat- ing that the method struggles to maintain stable contact constraints and consistent spatial coordi- nation among dancers. The original ST-GDance alleviates some spatial coordination issues, but still suffers from foot sliding artifacts. In several 14 ST - Gdance ++ ( Ours ) GCD ST - GDance Fig. 4 Qualitative results compared with baselines. frames, the feet appear to drift along the ground instead of maintaining stable contact, resulting in visually unnatural motion. This phenomenon sug- gests that, although ST-GDance captures overall group structure, its temporal modeling is insuffi- cient for fine-grained motion accuracy, especially in long or complex sequences. In contrast, ST-GDance++ generates more stable and physically plausible motions. The feet maintain consistent ground contact, and the tran- sitions between steps are smoother and more natural. Moreover, inter-dancer spacing remains coherent throughout the sequence, with fewer collisions or abrupt adjustments. These visual improvements demonstrate the effectiveness of the proposed spatial modeling, alignment attention, and streaming-oriented temporal design in pro- ducing more realistic and coordinated group dance motions. Varying Group Sizes. We further evaluate the scalability of ST-GDance++ by testing the model under different group sizes, ranging from two to five dancers. As illustrated in Fig. 5, ST- GDance++ consistently produces stable and coor- dinated motions across all group configurations. When the number of dancers increases, the spatial layout becomes more crowded and the interac- tion complexity grows. Despite this, the generated motions remain well-organized, with consistent inter-dancer spacing and synchronized movement patterns. Even in the five-dancer setting, the model avoids collisions and maintains coherent formations throughout the sequence. This result demonstrates that ST-GDance++ scales effec- tively with the number of dancers. The explicit spatial modeling mechanism allows the model to capture inter-dancer relationships without rely- ing on dense attention over concatenated tokens, preventing the quadratic growth of computational complexity. As a result, the model preserves both motion quality and spatial consistency as the group size increases, highlighting its suitability for long-duration and large-group choreography generation. Visualization of the ablation experi- ment. We further present qualitative comparisons of different ablated variants of ST-GDance++ to analyze the effect of each component. As shown in Fig. 6, removing each module leads to distinct motion artifacts. Without the Triangular Noise Schedule (w/o TNS), the generated motions become overly simi- lar across dancers, with reduced diversity in poses and transitions. This phenomenon is consistent with the quantitative results in Table 4, where the diversity metric drops compared to the full model, indicating that TNS contributes to richer motion variations. When the Alignment Attention Mask is removed (w/o AAM), we observe abrupt swap- ping behaviors between dancers, as highlighted in the red dashed regions. The dancers occasionally exchange spatial positions in an unnatural man- ner, leading to inconsistent temporal alignment. 15 2 3 4 5 Fig. 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. w/o TNS w/o AAM w/o SMB ST - Gdance ++ Fig. 6 Qualitative comparison under different component ablations of ST-GDance++. This visual artifact corresponds to the degrada- tion in group motion metrics reported in Table 4, suggesting that AAM is essential for maintaining stable inter-dancer alignment over time. With- out the Spatial Modeling Block (w/o SMB), the generated sequences exhibit frequent collisions, particularly around the arms and upper body. As shown in the highlighted regions, dancersâ limbs intersect due to the lack of explicit spatial interac- tion modeling. This observation is consistent with the increased GMR and TIF values in Table 4, indicating poorer spatial coordination. In contrast, the full ST-GDance++ model pro- duces stable, well-coordinated motions with clear inter-dancer spacing, smooth transitions, and diverse poses. These qualitative results, together with the quantitative improvements in Table 4, demonstrate that each component plays a com- plementary role in improving motion diversity, alignment, and spatial consistency. 6 Limitation Despite the improvements achieved by ST- GDance++ in long-duration group dance gen- eration, several limitations remain. First, the current framework is designed for a single cross- modal setting, where dance motions are generated solely from music input. It does not yet sup- port more flexible or user-driven controls, such as textual descriptions, formation constraints, or motion-level editing signals. While focusing on this basic setting allows us to study the core spatialâtemporal generation problem, more prac- tical applications would require richer and more interactive control mechanisms. Second, although 16 the proposed spatial modeling improves overall coordination, the model still struggles with cer- tain complex interaction patterns, such as rapid role exchanges or formation swaps among dancers. These behaviors are inherently difficult to model, and are also underrepresented in existing datasets, which typically contain limited examples of such interactions. As a result, the modelâs performance in highly dynamic group choreography remains constrained. Future work will explore incorpo- rating richer control signals and improving the modeling of complex interaction patterns, as well as leveraging more diverse and better-annotated datasets for large-scale group choreography gener- ation. 7 Conclusion In this work, we presented ST-GDance++, a scalable spatialâtemporal diffusion framework for long-duration group dance generation. By decou- pling spatial interaction modeling from temporal motion dynamics, the proposed method avoids the quadratic complexity of dense attention over dancerâframe tokens. Specifically, we introduce a lightweight spatial modeling block to explicitly capture inter-dancer relationships, an alignment- aware attention mechanism for stable coordina- tion, and a streaming-oriented noise schedule to support long-sequence generation. Extensive experiments demonstrate that ST-GDance++ achieves improved motion quality, spatial consis- tency, and computational efficiency across both single and group dance scenarios. Qualitative and quantitative results show that the proposed com- ponents complement each other, leading to more stable, diverse, and well-coordinated motions, especially in long-duration and multi-dancer set- tings. Supplementary information. The supple- mentary material includes (i) the original con- ference version of this work (ST-GDance) for reference and (i) additional qualitative results in the form of generated multi-dancer dance videos. Declarations Conflict of interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Data availability The experiments in this study are conducted using the publicly available AIOZ-GDance dataset (Le et al. 2023), which can be accessed at https://huggingface.co/datasets/ aiozai/AIOZ-GDANCE. References Alemi, O., Fran ̧coise, J., Pasquier, P.: Groovenet: Real-time music-driven dance movement gener- ation using artificial neural networks. networks 8(17), 26 (2017) Benoit, H.: Diffusion Forcing. Accessed: 2025- 12-18 (2025). https://haroldbenoit.com/notes/ ml/diffusion-models/frameworks---theory/ diffusion-forcing Bulatov, A., Kuratov, Y., Kapushev, Y., Burtsev, M.: Beyond attention: breaking the limits of transformer context length with recurrent mem- ory. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, p. 17700â17708 (2024) Chen, G., Lin, D., Yang, J., Lin, C., Zhu, J., Fan, M., Zhang, H., Chen, S., Chen, Z., Ma, C., et al.: Skyreels-v2: Infinite-length film generative model. arXiv preprint arXiv:2504.13074 (2025) Cai, Y., Wu, Y., Li, K., Zhou, Y., Zheng, B., Liu, H.: Flooddiffusion: Tailored diffusion forcing for streaming motion generation. arXiv preprint arXiv:2512.03520 (2025) Dai, Y., Zhu, W., Li, R., Ren, Z., Zhou, X., Li, X., Li, J., Yang, J.: Harmonious group chore- ography with trajectory-controllable diffusion. In: Proceedings of the AAAI Conference on Artificial Intelligence (2025) Dai, Y., Zhu, W., Li, R., Li, X., Zhang, Z., Li, J., Yang, J.: Tcdiff++: An end-to-end trajectory- controllable diffusion model for harmonious music-driven group choreography. International Journal of Computer Vision 134(2), 61 (2026) Ferreira, J.P., Coutinho, T.M., Gomes, T.L., Neto, J.F., Azevedo, R., Martins, R., Nascimento, 17 E.R.: Learning to dance: A graph convolutional adversarial network to generate realistic dance motions from audio. Computers & Graphics 94, 11â21 (2021) Gu, A., Dao, T.: Mamba: Linear-time sequence modeling with selective state spaces. In: First Conference on Language Modeling (2024) Gu, A., Goel, K., R Ìe, C.: Efficiently modeling long sequences with structured state spaces. In: International Conference on Learning Represen- tations (ICLR) (2022) Guo, Z., Mao, J., Tao, R., Yan, L., Ouchi, K., Liu, H., Wang, X.: Audio generation with multiple conditional diffusion model. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, p. 18153â18161 (2024) Huang, R., Hu, H., Wu, W., Sawada, K., Zhang, M., Jiang, D.: Dance revolution: Long-term dance generation with music via curriculum learning. In: International Conference on Learn- ing Representations (2020) Ho, J., Jain, A., Abbeel, P.: Denoising diffu- sion probabilistic models. Advances in neural information processing systems 33, 6840â6851 (2020) Huang, X., Li, Z., He, G., Zhou, M., Shechtman, E.: Self forcing: Bridging the train-test gap in autoregressive video diffusion. arXiv preprint arXiv:2506.08009 (2025) Hong, F., Zhang, M., Pan, L., Cai, Z., Yang, L., Liu, Z.: Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. ACM Transactions on Graphics (TOG) (2022) Joshi, M., Chakrabarty, S.: An extensive review of computational dance automation techniques and applications. Proceedings of the Royal Soci- ety A 477(2251), 20210071 (2021) Jiang, B., Chen, X., Zeng, A., Sun, X., Yin, F., Zeng, X., Zhang, X., Yu, G., Chen, T.: Causal motion tokenizer for streaming motion generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 2024â2034 (2025) Kovar, L., Gleicher, M., Pighin, F.: Motion graphs. In: Proceedings of SIGGRAPH (2002) Kodaira, A., Hou, T., Hou, J., Georgopoulos, M., Juefei-Xu, F., Tomizuka, M., Zhao, Y.: Streamdit: Real-time streaming text-to-video generation. arXiv preprint arXiv:2507.03745 (2025) Kim, T.-h., Park, S.I., Shin, S.Y.: Rhythmic- motion synthesis based on motion-beat analysis. ACM Transactions on Graphics (TOG) 22(3), 392â401 (2003) Le, N., Do, K., Bui, X., Do, T., Tjiputra, E., Tran, Q.D., Nguyen, A.: Scalable group choreogra- phy via variational phase manifold learning. In: European Conference on Computer Vision, p. 293â311 (2024). Springer Le, N., Do, T., Do, K., Nguyen, H., Tjiputra, E., Tran, Q.D., Nguyen, A.: Controllable group choreography using contrastive diffusion. ACM Transactions on Graphics (TOG) 42(6), 1â14 (2023) Liu, K., Hu, W., Xu, J., Shan, Y., Lu, S.: Rolling forcing: Autoregressive long video diffusion in real time. arXiv preprint arXiv:2509.25161 (2025) Lee, L.-H., Lin, Z., Hu, R., Gong, Z., Kumar, A., Li, T., Li, S., Hui, P.: When creators meet the metaverse: A survey on computational arts. arXiv preprint arXiv:2111.13486 (2021) Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, p. 851â866 (2023) Le, N., Pham, T., Do, T., Tjiputra, E., Tran, Q.D., Nguyen, A.: Music-driven group choreography. In: IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR) (2023) Lee, H.-Y., Yang, X., Liu, M.-Y., Wang, T.-C., Lu, Y.-D., Yang, M.-H., Kautz, J.: Dancing to music. In: Advances in Neural Information Processing Systems (NeurIPS), vol. 32 (2019) 18 Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Ai choreographer: Music conditioned 3d dance gen- eration with aist++. In: Proceedings of the IEEE/CVF International Conference on Com- puter Vision, p. 13401â13412 (2021) Liang, H., Zhang, W., Li, W., Yu, J., Xu, L.: Intergen: Diffusion-based multi-human motion generation under complex interactions. Inter- national Journal of Computer Vision 132(9), 3463â3483 (2024) Li, R., Zhao, J., Zhang, Y., Su, M., Ren, Z., Zhang, H., Tang, Y., Li, X.: Finedance: A fine-grained choreography dataset for 3d full body dance generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 10234â10243 (2023) Li, R., Zhang, H., Zhang, Y., Zhang, Y., Zhang, Y., Guo, J., Zhang, Y., Li, X., Liu, Y.: Lodge++: High-quality and long dance gen- eration with vivid choreography patterns. In: IEEE/CVF Conf. on Computer Vision and Pat- tern Recognition (CVPR) (2024) Li, R., Zhang, Y., Zhang, Y., Zhang, H., Guo, J., Zhang, Y., Liu, Y., Li, X.: Lodge: A coarse to fine diffusion network for long dance generation guided by the characteristic dance primitives. In: IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR) (2024) McFee, B., Raffel, C., Liang, D., Ellis, D.P., McVicar, M., Battenberg, E., Nieto, O., et al.: librosa: Audio and music signal analysis in python. SciPy 2015(18-24), 7 (2015) Qi, Q., Zhuo, L., Zhang, A., Liao, Y., Fang, F., Liu, S., Yan, S.: Diffdance: Cascaded human motion diffusion model for dance generation. In: Proceedings of the 31st ACM Interna- tional Conference on Multimedia, p. 1374â 1382 (2023) Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffu- sion models with deep language understanding. Advances in neural information processing sys- tems 35, 36479â36494 (2022) Sui, K., Ghosh, A., Hwang, I., Zhou, B., Wang, J., Guo, C.: A survey on human interaction motion generation. International Journal of Computer Vision 134(3), 113 (2026) Safonova, A., Hodgins, J.K.: Construction and optimal search of interpolated motion graphs. In: Proceedings of ACM SIGGRAPH, p. 106 (2007) Sun, G., Wong, Y., Cheng, Z., Kankanhalli, M.S., Geng, W., Li, X.: Deepdance: music-to-dance motion choreography with adversarial learning. IEEE Transactions on Multimedia 23, 497â509 (2020) Sun, M., Wang, W., Li, G., Liu, J., Sun, J., Feng, W., Lao, S., Zhou, S., He, Q., Liu, J.: Ar- diffusion: Asynchronous video generation with auto-regressive diffusion. In: Proceedings of the Computer Vision and Pattern Recognition Con- ference, p. 7364â7373 (2025) Siyao, L., Yu, W., Gu, T., Lin, C., Wang, Q., Qian, C., Loy, C.C., Liu, Z.: Bailando: 3d dance gen- eration by actor-critic gpt with choreographic memory. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 11050â11059 (2022) Siyao, L., Yu, W., Gu, T., Lin, C., Wang, Q., Qian, C., Loy, C.C., Liu, Z.: Bailando++: 3d dance gpt with choreographic memory. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(12), 14192â14207 (2023) https: //doi.org/10.1109/TPAMI.2023.3319435 Tseng, J., Castellon, R., Liu, K.: Edge: Editable dance generation from music. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 448â458 (2023) Tsuchida, S., Fukayama, S., Hamasaki, M., Goto, M.: Aist dance video database: Multi-genre, multi-dancer, and multi-camera database for dance information processing. In: ISMIR, vol. 1, p. 6 (2019) Valle-P Ìerez, G., Henter, G.E., Beskow, J., Holzapfel, A., Oudeyer, P.-Y., Alexanderson, S.: Transflower: probabilistic autoregressive dance generation with multimodal attention. ACM 19 Transactions on Graphics (TOG) 40(6), 1â14 (2021) Weng, J., Yan, Z., Tai, Y., Qian, J., Yang, J., Li, J.: Mamballie: Implicit retinex-aware low light enhancement with global-then-local state space. Advances in neural information processing sys- tems 37, 27440â27462 (2024) Xu, J., Wang, W., Chen, C., Liu, J., Ke, Q.: St-gdance: Long-term and collision-free group choreography from music. In: Proceedings of the British Machine Vision Conference (BMVC) (2025) Ye, T., Dong, L., Xia, Y., Sun, Y., Zhu, Y., Huang, G., Wei, F.: Differential transformer. In: The Thirteenth International Conference on Learning Representations (ICLR) (2025). https: //openreview.net/forum?id=OvoCm1gGhN Yao, S., Sun, M., Li, B., Yang, F., Wang, J., Zhang, R.: Dance with you: The diversity con- trollable dancer generation via diffusion mod- els. In: Proceedings of the 31st ACM Inter- national Conference on Multimedia, p. 8504â 8514 (2023) Yang, K., Tang, X., Diao, R., Liu, H., He, J., Fan, Z.: Codancers: Music-driven coherent group dance generation with choreographic unit. In: Proceedings of the 2024 International Con- ference on Multimedia Retrieval, p. 675â683 (2024) Yalta, N., Watanabe, S., Nakadai, K., Ogata, T.: Weakly-supervised deep recurrent neural networks for basic dance step generation. In: 2019 International Joint Conference on Neural Networks (IJCNN), p. 1â8 (2019). IEEE Yin, W., Yin, H., Baraka, K., Kragic, D., Bj Ìorkman, M.: Dance style transfer with cross- modal transformer. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, p. 5058â5067 (2023) Yin, T., Zhang, Q., Zhang, R., Freeman, W.T., Durand, F., Shechtman, E., Huang, X.: From slow bidirectional to fast autoregressive video diffusion models. In: Proceedings of the Com- puter Vision and Pattern Recognition Confer- ence, p. 22963â22974 (2025) Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 5745â5753 (2019) Zhang, Y., Feng, Y., Cseke, A., Saini, N., Bajan- das, N., Heron, N., Black, M.J.: Primal: Phys- ically reactive and interactive motor model for avatar learning. In: Proceedings of the IEEE/CVF International Conference on Com- puter Vision, p. 12725â12736 (2025) Zhu, W., Ma, X., Ro, D., Ci, H., Zhang, J., Shi, J., Gao, F., Tian, Q., Wang, Y.: Human motion generation: A survey. IEEE Transac- tions on Pattern Analysis and Machine Intelli- gence 46(4), 2430â2449 (2023) Zhang, Z., Wang, Y., Mao, W., Li, D., Zhao, R., Wu, B., Song, Z., Zhuang, B., Reid, I., Hartley, R.: Motion anything: Any to motion generation. arXiv preprint arXiv:2503.06955 (2025) Zhou, H., Zhang, S., Peng, J., Zhang, S., Li, J., Xiong, H., Zhang, W.: Informer: Beyond effi- cient transformer for long sequence time-series forecasting. In: Proceedings of the AAAI Con- ference on Artificial Intelligence, vol. 35, p. 11106â11115 (2021) 20