Pith. sign in

REVIEW 3 major objections 5 minor 42 references

ST-GDance: Long-Term and Collision-Free Group Choreography from Music

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read ST-GDance proposes decoupling spatial and temporal modeling in diffusion-based group choreography, cutting denoising complexity from quadratic to near-linear in sequence length and reporting the best long-term group metrics on AIOZ-GDance.

desk verdict Sensible decoupling of spatial and temporal modeling for group dance, but the central efficiency claim is contradicted by the paper's own complexity analysis. read the letter →

arxiv 2507.21518 v2 pith:OWUY3LUM submitted 2025-07-29 cs.AI

Jing Xu , Weiqiang Wang , Cunjian Chen , Jun Liu , Qiuhong Ke This is my paper · ORCID
classification cs.AI
keywords groupdancegenerationmusic-conditionedmotionsynthesisdiffusionmodelsgraphconvolutionalnetworksspatial-temporaldecouplinglong-sequencecollisionavoidanceefficientattention
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper sets out to show that long, collision-free group dance can be generated from music without the quadratic explosion in cost that plagues transformer-based diffusion models. Its proposal, ST-GDance, decouples the generation task into a spatial graph over dancers and per-dancer temporal attention, so that denoising complexity drops from O((NL)^2) to roughly O($LN^{2}$)+O(NL). On the AIOZ-GDance dataset the method reports the best group-level realism and correlation scores on long sequences while using a fraction of the FLOPs and parameters of the main diffusion baselines. If the results hold, dance animation from full-length music with many performers becomes feasible on limited hardware.

What carries the argument

The load-bearing object is the weighted distance graph G with adjacency A_ij = 1/(|p_i-p_j|+epsilon), injected through a spectral GCN with top-k edge pruning. It carries the spatial coordination and collision-avoidance claim. On the temporal side, differential attention and the local dependency transformer (a windowed linear attention) carry the long-range coherence claim, and the two are interleaved so total denoising complexity is approximately O($LN^{2}$)+O(NL) instead of O((NL)^2).

What would settle it

Generate a long duet where the two dancers are given trajectories that converge at high speed; if the output shows them intersecting or passing through each other more often than a baseline that explicitly predicts future positions, the instantaneous-distance graph is not sufficient for the collision-free claim.

Watch

Extended reading notes

Core claim

ST-GDance claims that the monolithic N-dancer, L-frame token sequence used by prior diffusion choreography models is both computationally wasteful and spatially blind. The paper replaces it with a distance-aware graph convolutional network that treats dancers as nodes with edge weights A_ij = 1/(|p_i-p_j|+epsilon) and prunes to top-k edges, plus a temporal transformer that applies differential attention and local dependency attention per dancer. This separates spatial coordination from temporal coherence and, the paper argues, is why it achieves the best or runner-up group metrics (GMR, GMC, TIF) on long-term AIOZ-GDance sequences while cutting inference time and model size relative to GCD, EDGE, and Lodge.

Load-bearing premise

The collision-free claim rests on the assumption that a fully connected graph with edge weights determined by instantaneous 2D distances between dancers can predict and prevent collisions during generation, even though it ignores relative velocity, 3D limb positions, and future motion.

Editorial extensions

If this is right

  • Full-length pieces (hundreds of frames) with several dancers can be generated in seconds rather than minutes on a single workstation, based on the reported 17.8s inference for 400-frame sequences.
  • The spatial-temporal decoupling makes the architecture modular: the graph module can be swapped or upgraded without retraining the temporal transformer.
  • Because FLOPs scale with LN^2 rather than N^2L^2, adding dancers is more affordable than extending sequence length, which fits production settings with fixed choreography lengths.
  • The reported TIF values (0.11 short, 0.13 long) indicate collisions are reduced but not eliminated, so 'collision-free' should be read as 'fewer collisions' in the paper's own numbers.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same decoupling could be applied to other multi-agent motion tasks where inter-agent spacing matters, such as crowd simulation or formation control, though the distance-graph assumption would need velocities to avoid fast-approaching pairs.
  • Because the edge weights depend only on instantaneous 2D positions, adding a velocity- or acceleration-dependent term to A_ij is a natural testable extension that should lower TIF further on fast choreography.
  • The claimed near-linear temporal complexity rests on W being small and fixed; at very long L with large windows, the O(LW) term can grow, so the headline complexity number is a regime statement, not a universal bound.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes ST-GDance, a diffusion-based framework for music-driven group dance generation. It decouples the generation task into spatial modeling via a distance-aware graph convolutional network and temporal modeling via a transformer decoder that mixes Local Dependency Transformer (LDT) and Differential Attention (DiffAttn) blocks. The authors claim that this decoupling reduces denoising complexity from O((NL)^2) to nearly O(LN^2)+O(NL), enabling efficient long-sequence generation. They evaluate on the AIOZ-GDance dataset for short-term (120 frames, 3 dancers) and long-term (400 frames, 3 dancers) settings, reporting favorable group-level metrics (GMR, GMC, TIF) and efficiency numbers against EDGE, GCD, CoDancers, TCDiff, and Lodge.

Significance. If the efficiency and quality claims hold, ST-GDance would be a practically useful contribution: it is a relatively simple spatial-temporal decoupling for group choreography that reports strong long-sequence results and lower parameter counts than several diffusion baselines. The empirical results on AIOZ-GDance, especially on long sequences, are valuable even if the asymptotic complexity claim is overstated. However, the central complexity assertion is currently not established, and the ablation table contains internally inconsistent efficiency numbers. These issues must be resolved before the main contribution can be accepted.

major comments (3)
  1. [§3.4, Eqs. (9)–(11)] The claimed total complexity O(LN^2)+O(NL) is not supported by the equations. Differential Attention explicitly forms two L×L softmax matrices per dancer (Eqs. 9–11), so its theoretical cost is O(NL^2). Section 3.6 concedes that 'the theoretical complexity remains quadratic O(NL^2),' but then asserts near-linear empirical scaling via 'shared projections and attention filtering' without a derivation or any profiling data. The reference to [36] and 'Table 2' does not establish near-linear scaling for this architecture. Since the third contribution bullet in §1 rests on this complexity reduction, Eq. (15) should be corrected to O(LN^2)+O(NL^2) unless the authors either replace DiffAttn with a genuinely linear-attention mechanism or provide measured scaling experiments over L. As written, the central efficiency claim is not internally consistent.
  2. [Table 2] The ablation efficiency numbers are internally inconsistent. The full ST-GDance model (all three blocks) reports 8.12 GFLOPs and 17.84 s inference time, while the variant without DMA (row 1) reports 26.87 GFLOPs and 28.41 s, and the variant without LDT (row 2) reports 13.73 GFLOPs and 24.31 s. Adding the SMB to the row-3 configuration (LDT+DMA) should not reduce FLOPs or inference time relative to that configuration, yet row 3 reports 13.73 GFLOPs / 15.62 s and the full model reports 8.12 GFLOPs / 17.84 s. The training-time pattern (full 1:25 vs. row 1 1:35 vs. row 2 1:48) is similarly hard to reconcile with the stated module additions. These numbers undermine the efficiency analysis and need to be rechecked and reported with a clear description of what is included in each configuration.
  3. [§3.3] The claim of 'collision-free' choreography is stronger than what the model actually enforces. The adjacency matrix in Eq. (6) depends only on instantaneous 2D Euclidean distances between dancers, with no use of relative velocities or 3D limb configurations. This static, per-frame graph cannot anticipate collisions that arise from motion between frames. The paper's own TIF metric in Table 1 is 0.11 (short-term) and 0.13 (long-term), which is low but not zero. The authors should either add a temporal collision-avoidance term, evaluate long-term TIF after tracking, or soften the 'collision-free' claim in the title and abstract to 'collision-reduced' or 'spatially constrained.'
minor comments (5)
  1. [§3.4] The LDT attention formula is not fully specified: the projection of Q, K, V to W×d and the denominator in Eq. (12) need a clearer description of normalization and dimension matching. As written, the expression is ambiguous about whether it is a softmax variant or a linear attention with a ReLU kernel.
  2. [§4.1] The metric PFC appears in both tables but is never defined in Section 4.1. Please add a definition or remove it from the tables.
  3. [§1] Minor typos: 'as such as' appears in §1, and 'Figuree' appears in §4.3. These should be corrected.
  4. [Figure 1] The caption states the complexity reduction is 'to O(LN^2)+O(NL^2)' but then says the overall complexity 'can be reduced to approximately O(LN^2)+O(NL).' This inconsistency mirrors the unresolved complexity issue in the main text and should be aligned with the corrected complexity statement.
  5. [Table 1] CoDancers has no efficiency entries (FLOPs, Params, Train Time, Inf Time) in either setting. Please state why these are unavailable, or remove the blank entries and note this in the text.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ST-GDance's design and evaluation are stated and fit-to-data in the standard sense, with no fitted inputs renamed as predictions and no load-bearing self-citation.

full rationale

The paper's derivation chain is self-contained as a system design report: the spatial GCN and temporal LDT/DiffAttn modules are defined by their equations (Eqs. 6-13), the loss function is a standard reconstruction-plus-auxiliary objective (Eq. 14), and the complexity claim (Eq. 15) is an accounting of those modules' costs. The reported GMR, GMC, TIF, and FID values come from held-out AIOZ-GDance splits; no test metric is used as a training signal or fitted constant, so there is no fitted-input-called-prediction pattern. The only internally questionable step is the claim that differential attention achieves 'near-linear complexity O(L) in practice [36]' despite the paper's own statement that 'the theoretical complexity remains quadratic O(NL^2)'. That is a potential correctness or evidence-quality issue about an imported complexity property, not a circular dependency on the paper's own outputs or definitions. There are no load-bearing self-citations and no uniqueness theorems imported from the authors' prior work. The collision-avoidance mechanism is an explicit inductive bias (inverse-distance graph weights), not a prediction derived from its own conclusion. Therefore the paper exhibits no significant circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on two hand-tuned spatial-temporal decompositions: a distance-based graph with unspecified epsilon and top-k, and a temporal transformer with unspecified window size and loss weights. No fundamentally new physical entity is introduced; the design is a combination of existing modules.

free parameters (4)
  • epsilon (epsilon) in adjacency = not specified
    Small constant in Eq. 6 to avoid division by zero and keep edges non-zero; chosen by hand, affects all edge weights and thus spatial interaction strength.
  • top-k edges = not specified
    Number of edges retained after pruning the fully connected graph; controls the cost of spatial GCN and quality of local relationships.
  • LDT window size W = not specified
    Local window length for Local Dependency Transformer; trades temporal context for linear complexity.
  • loss weights lambda_pos, lambda_vel, lambda_contact = not specified; set differently for short/long
    Balance reconstruction, position, velocity, and contact losses in Eq. 14; values deferred to missing supplementary material.
assumptions (5)
  • standard math The DDPM forward and reverse diffusion processes (Eqs. 1-2) correctly model dance motion noise.
    Adopted from Ho et al. [10] and EDGE [29] without modification.
  • domain assumption The AIOZ-GDance dataset and the [18] split provide a valid benchmark for group dance generation.
    The paper relies on this dataset for all quantitative claims, but no analysis of dataset bias or protocol differences is given.
  • domain assumption Long-term sequences reconstructed by grouping clips with the same sequence ID are valid long sequences.
    Section 4.1; assumes clip boundaries align seamlessly enough to test long-term coherence.
  • ad hoc to paper Pairwise Euclidean distance in 2D between dancers is a sufficient spatial feature for collision avoidance.
    Introduced in Eq. 6 as the core spatial signal; not validated against 3D or velocity-based features.
  • standard math Differential attention and LDT behave as described in their source papers and combine without negative interaction.
    The method imports DIFF Transformer [36] and SANA [32] components; the paper does not re-derive them.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ST-GDance: Long-Term and Collision-Free Group Choreography from Music." pith.science (2026). https://pith.science/paper/OWUY3LUM

@misc{pith2026250721518,
  author       = {Pith},
  title        = {Pith review of: ST-GDance: Long-Term and Collision-Free Group Choreography from Music},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OWUY3LUM}},
  note         = {Machine review of arXiv:2507.21518}
}
read the original abstract

Group dance generation from music has broad applications in film, gaming, and animation production. However, it requires synchronizing multiple dancers while maintaining spatial coordination. As the number of dancers and sequence length increase, this task faces higher computational complexity and a greater risk of motion collisions. Existing methods often struggle to model dense spatial-temporal interactions, leading to scalability issues and multi-dancer collisions. To address these challenges, we propose ST-GDance, a novel framework that decouples spatial and temporal dependencies to optimize long-term and collision-free group choreography. We employ lightweight graph convolutions for distance-aware spatial modeling and accelerated sparse attention for efficient temporal modeling. This design significantly reduces computational costs while ensuring smooth and collision-free interactions. Experiments on the AIOZ-GDance dataset demonstrate that ST-GDance outperforms state-of-the-art baselines, particularly in generating long and coherent group dance sequences. Project page: https://yilliajing.github.io/ST-GDance-Website/.

Figures

Figures reproduced from arXiv: 2507.21518 by the authors.

Figure 1
Figure 1. Illustration of the overall framework for the proposed ST-GDance. We decouple [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The proposed framework for spatial-temporal group dance generation. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Qualitative comparison of generated dance sequences across different models. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 36 canonical work pages

  1. [36]

    Differential transformer

    Tianzhu Ye, Li Dong, Yuqing Xia, Yutao Sun, Yi Zhu, Gao Huang, and Furu Wei. Differential transformer. In The Thirteenth International Conference on Learning Representations (ICLR) ,

    work page

  2. [1]

    A spatio-temporal transformer for 3d human motion prediction

    Emre Aksan, Manuel Kaufmann, Peng Cao, and Otmar Hilliges. A spatio-temporal transformer for 3d human motion prediction. In 2021 International Conference on 3D Vision (3DV), pages 565–574. IEEE, 2021

    work page 2021

  3. [2]

    Groovenet: Real-time music- driven dance movement generation using artificial neural networks

    Omid Alemi, Jules Françoise, and Philippe Pasquier. Groovenet: Real-time music- driven dance movement generation using artificial neural networks. networks, 8(17): 26, 2017

    work page 2017

  4. [3]

    A stochastic conditioning scheme for diverse human motion predic- tion

    Sadegh Aliakbarian, Fatemeh Sadat Saleh, Mathieu Salzmann, Lars Petersson, and Stephen Gould. A stochastic conditioning scheme for diverse human motion predic- tion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5223–5232, 2020

    work page 2020

  5. [4]

    Interactive motion generation from examples.ACM Transactions on Graphics (TOG), 21(3):483–490, 2002

    Okan Arikan and David A Forsyth. Interactive motion generation from examples.ACM Transactions on Graphics (TOG), 21(3):483–490, 2002

    work page 2002

  6. [5]

    Beyond atten- tion: breaking the limits of transformer context length with recurrent memory

    Aydar Bulatov, Yuri Kuratov, Yermek Kapushev, and Mikhail Burtsev. Beyond atten- tion: breaking the limits of transformer context length with recurrent memory. In Pro- ceedings of the AAAI Conference on Artificial Intelligence , volume 38, pages 17700– 17708, 2024. XU, W ANG, CHEN, LIU, KE: ST-GDANCE 11

    work page 2024

  7. [6]

    Harmonious group choreography with trajectory-controllable diffusion

    Yuqin Dai, Wanlu Zhu, Ronghui Li, Zeping Ren, Xiangzheng Zhou, Xiu Li, Jun Li, and Jian Yang. Harmonious group choreography with trajectory-controllable diffusion. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025

    work page 2025

  8. [7]

    Learning to dance: A graph convolu- tional adversarial network to generate realistic dance motions from audio

    Joao P Ferreira, Thiago M Coutinho, Thiago L Gomes, José F Neto, Rafael Azevedo, Renato Martins, and Erickson R Nascimento. Learning to dance: A graph convolu- tional adversarial network to generate realistic dance motions from audio. Computers & Graphics, 94:11–21, 2021

    work page 2021

Show all 42 references
  1. [8]

    Audio generation with multiple conditional diffusion model

    Zhifang Guo, Jianguo Mao, Rui Tao, Long Yan, Kazushige Ouchi, Hong Liu, and Xi- angdong Wang. Audio generation with multiple conditional diffusion model. In Pro- ceedings of the AAAI Conference on Artificial Intelligence , volume 38, pages 18153– 18161, 2024

    2024

  2. [9]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017

    2017

  3. [10]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020

    2020

  4. [11]

    Avatarclip: Zero-shot text-driven generation and animation of 3d avatars

    Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. ACM Transactions on Graphics (TOG), 2022

    2022

  5. [12]

    An extensive review of computational dance automation techniques and applications

    Manish Joshi and Sangeeta Chakrabarty. An extensive review of computational dance automation techniques and applications. Proceedings of the Royal Society A , 477 (2251):20210071, 2021

    2021

  6. [13]

    Rhythmic-motion synthesis based on motion-beat analysis

    Tae-hoon Kim, Sang Il Park, and Sung Yong Shin. Rhythmic-motion synthesis based on motion-beat analysis. ACM Transactions on Graphics (TOG), 22(3):392–401, 2003

    2003

  7. [14]

    Semi-supervised classification with graph convolu- tional networks

    Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolu- tional networks. arXiv preprint arXiv:1609.02907, 2016

    2016 arXiv

  8. [15]

    Motion graphs

    Lucas Kovar, Michael Gleicher, and Frédéric Pighin. Motion graphs. In Proceedings of SIGGRAPH, 2002

    2002

  9. [16]

    Controllable group choreography using contrastive diffusion

    Nhat Le, Tuong Do, Khoa Do, Hien Nguyen, Erman Tjiputra, Quang D Tran, and Anh Nguyen. Controllable group choreography using contrastive diffusion. ACM Transac- tions on Graphics (TOG), 42(6):1–14, 2023

    2023

  10. [17]

    Tran, and Anh Nguyen

    Nhat Le, Thang Pham, Tuong Do, Erman Tjiputra, Quang D. Tran, and Anh Nguyen. Music-driven group choreography. 2023

    2023

  11. [18]

    Music-driven group choreography

    Nhat Le, Thang Pham, Tuong Do, Erman Tjiputra, Quang D Tran, and Anh Nguyen. Music-driven group choreography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8673–8682, 2023

    2023

  12. [19]

    Dancing to music

    Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming- Hsuan Yang, and Jan Kautz. Dancing to music. In Advances in Neural Information Processing Systems (NeurIPS), volume 32, 2019. 12 XU, W ANG, CHEN, LIU, KE: ST-GDANCE

    2019

  13. [20]

    When creators meet the metaverse: A survey on computational arts

    Lik-Hang Lee, Zijun Lin, Rui Hu, Zhengya Gong, Abhishek Kumar, Tangyao Li, Sijia Li, and Pan Hui. When creators meet the metaverse: A survey on computational arts. arXiv preprint arXiv:2111.13486, 2021

    2021 arXiv

  14. [21]

    Finedance: A fine-grained choreography dataset for 3d full body dance generation

    Ronghui Li, Junfan Zhao, Yachao Zhang, Mingyang Su, Zeping Ren, Han Zhang, Yan- song Tang, and Xiu Li. Finedance: A fine-grained choreography dataset for 3d full body dance generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10234–10243, 2023

    2023

  15. [22]

    Lodge: A coarse to fine diffusion network for long dance gen- eration guided by the characteristic dance primitives

    Ronghui Li, Yuxiang Zhang, Yachao Zhang, Hongwen Zhang, Jie Guo, Yan Zhang, Yebin Liu, and Xiu Li. Lodge: A coarse to fine diffusion network for long dance gen- eration guided by the characteristic dance primitives. InIEEE/CVF Conf. on Computer Vision and Pattern Recognition (...

    2024

  16. [23]

    Ai choreographer: Music conditioned 3d dance generation with aist++

    Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13401–13412, 2021

    2021

  17. [24]

    Film: Visual reasoning with a general conditioning layer

    Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018

    2018

  18. [25]

    Diffdance: Cascaded human motion diffusion model for dance generation

    Qiaosong Qi, Le Zhuo, Aixi Zhang, Yue Liao, Fei Fang, Si Liu, and Shuicheng Yan. Diffdance: Cascaded human motion diffusion model for dance generation. In Pro- ceedings of the 31st ACM International Conference on Multimedia, pages 1374–1382, 2023

    2023

  19. [26]

    Construction and optimal search of interpolated motion graphs

    Alla Safonova and Jessica K Hodgins. Construction and optimal search of interpolated motion graphs. In Proceedings of ACM SIGGRAPH, pages 106–es, 2007

    2007

  20. [27]

    Photorealistic text-to-image diffusion models with deep language under- standing

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Den- ton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Sali- mans, et al. Photorealistic text-to-image diffusion models with deep language under- standing. Advances in neural infor...

    2022

  21. [28]

    Deepdance: music-to-dance motion choreography with adversarial learning

    Guofei Sun, Yongkang Wong, Zhiyong Cheng, Mohan S Kankanhalli, Weidong Geng, and Xiangdong Li. Deepdance: music-to-dance motion choreography with adversarial learning. IEEE Transactions on Multimedia, 23:497–509, 2020

    2020

  22. [29]

    Edge: Editable dance generation from music

    Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 448–458, 2023

    2023

  23. [30]

    Aist dance video database: Multi-genre, multi-dancer, and multi-camera database for dance information processing

    Shuhei Tsuchida, Satoru Fukayama, Masahiro Hamasaki, and Masataka Goto. Aist dance video database: Multi-genre, multi-dancer, and multi-camera database for dance information processing. In ISMIR, volume 1, page 6, 2019

    2019

  24. [31]

    Transflower: probabilistic autoregressive dance generation with multimodal attention

    Guillermo Valle-Pérez, Gustav Eje Henter, Jonas Beskow, Andre Holzapfel, Pierre- Yves Oudeyer, and Simon Alexanderson. Transflower: probabilistic autoregressive dance generation with multimodal attention. ACM Transactions on Graphics (TOG), 40(6):1–14, 2021. XU, W ANG, CHEN, L...

    2021

  25. [32]

    SANA: Efficient high- resolution image synthesis with linear diffusion transformer

    Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, and Song Han. SANA: Efficient high- resolution image synthesis with linear diffusion transformer. In Proceedings of the In- ternational Conference on Learning Rep...

    2025 arXiv

  26. [33]

    Weakly-supervised deep recurrent neural networks for basic dance step generation

    Nelson Yalta, Shinji Watanabe, Kazuhiro Nakadai, and Tetsuya Ogata. Weakly-supervised deep recurrent neural networks for basic dance step generation. In2019 International Joint Conference on Neural Networks (IJCNN), pages 1–8. IEEE, 2019

    2019

  27. [34]

    Codancers: Music-driven coherent group dance generation with choreographic unit

    Kaixing Yang, Xulong Tang, Ran Diao, Hongyan Liu, Jun He, and Zhaoxin Fan. Codancers: Music-driven coherent group dance generation with choreographic unit. In Proceedings of the 2024 International Conference on Multimedia Retrieval, pages 675–683, 2024

    2024

  28. [35]

    Dance with you: The diversity controllable dancer generation via diffusion models

    Siyue Yao, Mingjie Sun, Bingliang Li, Fengyu Yang, Junle Wang, and Ruimao Zhang. Dance with you: The diversity controllable dancer generation via diffusion models. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8504–8514, 2023

    2023

  29. [37]

    Graph-based normalizing flow for human motion generation and reconstruction

    Wenjie Yin, Hang Yin, Danica Kragic, and Mårten Björkman. Graph-based normalizing flow for human motion generation and reconstruction. In 2021 30th IEEE international conference on robot & human interactive communication (RO-MAN), pages 641–648. IEEE, 2021

    2021

  30. [38]

    Dance style transfer with cross-modal transformer

    Wenjie Yin, Hang Yin, Kim Baraka, Danica Kragic, and Mårten Björkman. Dance style transfer with cross-modal transformer. In Proceedings of the IEEE/CVF Winter Conference on Applica- tions of Computer Vision, pages 5058–5067, 2023

    2023

  31. [39]

    Motion anything: Any to motion generation

    Zeyu Zhang, Yiran Wang, Wei Mao, Danning Li, Rui Zhao, Biao Wu, Zirui Song, Bohan Zhuang, Ian Reid, and Richard Hartley. Motion anything: Any to motion generation. arXiv preprint arXiv:2503.06955, 2025

    2025 arXiv

  32. [40]

    Informer: Beyond efficient transformer for long sequence time-series forecasting

    Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. Informer: Beyond efficient transformer for long sequence time-series forecasting. In Proceedings of the AAAI conference on artificial intelligence , volume 35, pages 11106–11115, 2021

    2021

  33. [41]

    Human motion generation: A survey

    Wentao Zhu, Xiaoxuan Ma, Dongwoo Ro, Hai Ci, Jinlu Zhang, Jiaxin Shi, Feng Gao, Qi Tian, and Yizhou Wang. Human motion generation: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(4):2430–2449, 2023

    2023

  34. [2025]

    https://openreview.net/forum?id=OvoCm1gGhN

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.