Pith. sign in

REVIEW 5 major objections 5 minor 4 cited by

FlowerDance: MeanFlow for Efficient and Refined 3D Dance Generation

T0 review · 5 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read FlowerDance claims that predicting interval-averaged flow velocities plus a physical consistency loss yields state-of-the-art 3D dance quality with as few as 5–20 sampling steps, running at 2008 FPS.

desk verdict A plausible few-step flow-matching system for music-to-dance with a real efficiency win, but the SOTA quality claim is undermined by sloppy baseline tables and missing error bars. read the letter →

arxiv 2511.21029 v3 pith:S4CH4BEM submitted 2025-11-26 cs.CV

Kaixing Yang , Xulong Tang , Ziqiao Peng , Xiangyue Zhang , Puwei Wang , Jun He , Hongyan Liu This is my paper · ORCID
classification cs.CV
keywords music-to-dancegenerationflowmatchingMeanfew-stepsamplingphysicalconsistencyconstraintBidirectionalMambanon-autoregressivemotionediting
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

The paper tries to establish that music-to-dance generation can reach state-of-the-art quality while running fast enough for real-time 3D rendering. Its route is to replace diffusion's long denoising chains with a flow-matching model that predicts the average velocity over each integration interval, so the training objective matches the Euler updates used at inference. To keep few-step outputs physically plausible, it adds a Physical Consistency Constraint that recovers the full motion from the predicted mean velocity and enforces reconstruction, velocity, and joint-position losses. The reported results on FineDance show the best kinetic FID (29.73), diversity, and beat alignment among compared methods at 2008 frames per second with 63M parameters, and the best kinetic FID on AIST++ (20.50). A sympathetic reader would take away that few-step flow matching plus a kinematic anchor can beat diffusion-based dance generators in both speed and quality.

What carries the argument

The interval-averaged velocity field u(z_t,r,t)—the mean of the instantaneous velocity over (r,t)—is the central object; the MeanFlow identity makes it trainable and lets one Euler update jump from t to r. The Physical Consistency Constraint is the stabilizer: it recovers the predicted motion at time zero and enforces reconstruction, velocity, and 3D joint-position losses, which the ablation shows is necessary to avoid NaN divergence. The architecture's carriers are the BiMamba backbone (forward and backward state-space scans fused with a multiplicative skip connection) and Channel-Level Cross-Modal Fusion (element-wise addition rather than cross-attention), giving O(n) temporal modeling and

What would settle it

One decisive check: run the released model on a fixed GPU against the strongest baselines reported in the paper with the same 1024-frame protocol and measure wall-clock time per sequence; if FlowerDance is not several times faster under identical hardware, the efficiency claim fails. For the quality claim, compute FID_k on the FineDance test set across 5 seeds with 95% confidence intervals; if the 29.73 value does not separate from the next best baseline, the SOTA claim is not established.

Watch

Extended reading notes

Core claim

The central claim is that a flow-matching generator can outperform diffusion-based dance models in both quality and efficiency if the network learns the interval-averaged velocity field and if training explicitly anchors intermediate states to the human motion manifold. The MeanFlow identity u(z_t,r,t)=v(z_t,t)-(t-r)d/dt u(z_t,r,t) lets a single Euler step z_r=z_t-(t-r)u(z_t,r,t) transport between any two times, so inference needs only 5–20 steps. In each training iteration the model also outputs the mean velocity to time zero, recovers the predicted motion, and applies reconstruction, forward-kinematic position, and velocity losses. The architecture uses a bidirectional Mamba backbone with

Load-bearing premise

The load-bearing premise is that the Physical Consistency Constraint, trained on random intervals, keeps 5–20-step Euler integration on the human-motion manifold at inference time; the paper asserts but does not directly measure this transfer at large step sizes, and without PCC training diverges to NaN.

Editorial extensions

If this is right

  • If the reported 2008 FPS holds under identical hardware, real-time, high-fidelity 3D dance generation becomes practical for interactive applications.
  • Few-step flow matching with a kinematic anchor could generalize to other motion-generation tasks (text-to-motion, gesture, character animation) where few-step diffusion outputs tend to jitter or drift.
  • Because training and inference share the same interval-averaged update, the model's quality degrades gracefully at 10 steps and stays usable at 5, so deployment can trade speed against quality.
  • Training-free editing through a time-decayed soft mask means users can interactively refine generated dances without retraining or fine-tuning.
  • The NaN failure without the physical consistency term indicates that flow-matching objectives alone under-constrain high-dimensional human-motion manifolds, making explicit kinematic anchors a design requirement for this domain.

Reading between the lines

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

  • If the speed comparison were rerun on identical hardware with matched sequence lengths and batching, the margin might shrink; the paper does not state the hardware or protocol behind the FPS numbers, so the headline efficiency claim should be verified independently.
  • The physical-consistency term is essentially a trajectory anchor; a natural extension is to feed the recovered motion back through the estimator for multiple iterations, effectively turning the few-step sampler into a predictor–corrector that could push toward one-step generation.
  • The FSR numbers are odd: the ground-truth FSR (0.216) is worse than several generated models' FSR, suggesting the foot-slide metric alone is a weak physical-plausibility signal; a better falsifier would compare joint accelerations or ground-contact forces.
  • Because the cross-modal fusion assumes frame-aligned music and dance, the method may not transfer to settings with loose temporal alignment; testing on misaligned audio–motion pairs would reveal how much the alignment assumption carries the beat-alignment advantage.
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

5 major / 5 minor

Summary. FlowerDance proposes a music-to-dance generation framework combining MeanFlow with a Physical Consistency Constraint (PCC) and a BiMamba backbone with channel-level cross-modal fusion. The paper claims state-of-the-art quality and efficiency on FineDance and AIST++, reporting e.g., FID_k=29.73 at 2008 FPS on FineDance (Table 1) and FID_k=20.50 on AIST++ (Table 2). The method also supports training-free motion editing via a time-decayed soft mask, and the text argues that few-step sampling (5–20 steps) with MeanFlow matches or exceeds 50-step diffusion baselines. The central claims are that the model achieves both high motion quality and high generation efficiency, enabling real-time 3D dance generation.

Significance. The paper is clearly written and the proposed architecture is simple and plausible. The candid reporting of NaN divergence in ablations (Table 5) is a strength, and the parameter-free fusion and O(n) BiMamba backbone are practical contributions. If the reported numbers are reproducible, the work would be a meaningful step toward real-time, high-quality dance generation with a small memory footprint. However, the quantitative evidence for the SOTA claim is currently fragile: apparent anomalies in the baseline tables (duplicate FID_k, implausible FSR), unexplained non-monotonicity in the sampling-step ablation, missing error bars, and an unspecified FPS measurement protocol all undermine the headline comparisons. The skeptics' concerns about metric reliability land and must be addressed.

major comments (5)
  1. [Table 1] The duplicate FID_k=50.00 for Lodge and MEGA, together with Lodge's FSR=0.028 (vs ground-truth 0.216), strongly suggests a transcription or evaluation-protocol error. Since Table 1 carries the FineDance SOTA claim, these entries must be corrected and verified against the original papers. All baseline numbers should be rechecked carefully.
  2. [Table 4 and Section 4.4] The reported 10-step MeanFlow FID_k=26.17 is better than the 20-step headline FID_k=29.73, while the text states that 20 steps is SOTA and 10 steps is 'near-SOTA.' This non-monotonicity is unexplained. No error bars, seeds, or evaluation sample counts are reported anywhere in the paper, so the pairwise differences that ground the 'state-of-the-art' qualifier in Tables 1 and 2 cannot be distinguished from noise.
  3. [Section 4.2, 'Generation Efficiency'] The FPS figures, including the 2008 FPS headline, are reported without any hardware, batch size, precision, or timing protocol. Since efficiency is half of the central claim, these numbers need a clear experimental setup to be meaningful and comparable.
  4. [Section 3.2, Eq. (6)] The MeanFlow training target is not self-contained. The target u_tgt contains v(zt,t), ∂z uθ, and ∂t uθ, but the paper does not specify how v is obtained (analytical conditional velocity or a separate network) nor how the Jacobian-vector product is computed in practice. As written, the loss depends on uθ itself, making the training objective circular and irreproducible. A pseudo-code or a derivation following the MeanFlow reference is required.
  5. [Section 4.4, 'Physical Consistency Constraint'] Table 5 shows that without PCC the training diverges to NaN; thus PCC is load-bearing for the entire few-step generation quality. However, no experiment analyzes PCC's effect at different inference step counts (e.g., S=5, 10, 20), and the weight values λ in Eq. (9) are not reported. The claim that PCC transfers from training-time single-step recovery to the multi-step Euler sampling used at inference is asserted, not demonstrated. Please provide this analysis or at least report the λ values and training curves.
minor comments (5)
  1. [Section 3.2] The sentence 'Notably, this reduces to the FM loss when r=t' is technically incorrect because Eq. (4) is undefined at r=t; the intended statement is the limit r→t.
  2. [Section 3.1] Reference [45] is cited as MERT, but [45] is CoDancers (Yang et al., ICMR 2024). The MERT paper does not appear in the bibliography; this citation needs correction.
  3. [Table 5] The 'NaN' entries should be explained (e.g., training divergence, loss overflow, or evaluation failure) so that readers understand what was observed and how the experiment was handled.
  4. [Section 3.3 and Eq. (10)] The notation t is used both as ODE time and as sequence time; this dual use makes Eq. (10) hard to follow. Please disambiguate, e.g., by using a separate symbol for the sampling step.
  5. [Figure 3] The caption claims FlowerDance generates motions that are 'natural, diverse, and complex' based on a single clip. Since this is a qualitative, subjective assessment, a more neutral description would be appropriate.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: central quality/efficiency claims are tested on held-out public benchmarks against external baselines, and the generative strategy is an adopted external method with an auxiliary training loss rather than a self-referential prediction.

full rationale

No load-bearing step reduces to its own input by construction. The MeanFlow objective (Eq. 6) is taken from external work [9] and is a self-consistency training loss, not a fitted quantity later renamed as a prediction; the Physical Consistency Constraint (Eqs. 8-9) compares a recovered motion with ground truth as a regularizer. The few-step efficiency claim is supported by an architecture-matched ablation (Table 4) against RectFlow and Diffusion. The SOTA claim rests on quantitative comparisons on held-out test splits of AIST++ and FineDance, including multiple non-self-cited baselines (FACT, MNET, Bailando, EDGE, FineNet, Lodge, DanceNet); self-citations to MEGA [49] and Match [48] are used as comparison baselines with reported numbers, not as justification of the derivation. No uniqueness theorem or ansatz is imported from the authors' prior work. The skeptic's concerns (duplicate baseline FID_k=50.00, 10-step FID_k=26.17 vs 20-step 29.73, absent error bars, ground-truth FSR worse than generated values) are empirical-validity and statistical-robustness issues, not circularity under the rubric; likewise, the NaN result without PCC in Table 5 is an ablation observation, not a definitional tautology.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The system is assembled from borrowed, cited components: the MeanFlow identity (Eq. 5-6 from Geng et al.), straight-line flow matching (Eq. 1-3 from Lipman et al.), and Mamba (Eq. 11-12 from Gu and Dao). What the paper itself supplies are the physical-consistency objective (Eq. 8-9), the BiMamba wrapper, and the soft-mask editing rule (Eq. 10). The most consequential free knobs—four loss weights, sampling steps S=20, architecture dimensions, and evaluation-protocol details—are unreported or hand-chosen, and the headline numbers depend on them. No new physical or latent entities are postulated, so invented_entities is empty.

free parameters (4)
  • Loss weights lambda_mf, lambda_rec, lambda_pos, lambda_vel = not reported (chosen to balance loss magnitudes at start of training)
    Eq. 9. The balance between the borrowed MeanFlow objective and the paper's own physical-consistency losses is hand-set; since Table 5 shows training collapses without PCC, these weights gate the central few-step-quality claim.
  • Number of sampling steps S = 20 (main results)
    Table 4: quality is non-monotonic in S (10-step FID_k 26.17 beats 20-step 29.73); the headline 2008 FPS corresponds to S=20. The choice affects both the quality and the efficiency claims.
  • Architecture hyperparameters (layers, hidden dims, SSM state size) = not reported
    BiMamba blocks, FiLM, and channel-level addition are described only at block level; no dimensions or training schedule are given, so the 63M parameter count and 2008 FPS cannot be reproduced from the paper.
  • Evaluation-protocol details (FSR threshold, FID extractor checkpoints, BAS window) = not specified in paper
    Metrics are referenced to prior work, but the exact FSR threshold and feature extractors matter: GT FSR 0.216 exceeds several generated models' FSR (Lodge 0.028), indicating threshold sensitivity that affects the physical-plausibility comparison.
assumptions (6)
  • standard math MeanFlow identity (Eq. 5): u(z_t, r, t) = v(z_t, t) − (t−r) d/dt u(z_t, r, t)
    Borrowed from Geng et al.; the training target Eq. 6, with its Jacobian-vector correction term, rests on this identity holding for the learned estimator. The paper quotes rather than re-derives it.
  • standard math Straight-line conditional flow path F_P(x,t) = (1−t)x + t·eps with velocity v = eps − x (Eq. 1)
    Standard flow matching (Lipman et al.); assumed without proof, used throughout Section 3.2.
  • ad hoc to paper Supervising recovered z0 with L2 reconstruction, velocity, and FK joint-position losses (Eq. 8) is sufficient to keep few-step trajectories on the human-motion manifold
    The paper's own ablation (Table 5, 'w/o PCC' → NaN) shows the central few-step-quality claim is carried by this auxiliary objective; its sufficiency at inference is asserted, not demonstrated.
  • domain assumption Predicting interval-averaged velocity (MeanFlow, Eq. 7) improves few-step quality over instantaneous-velocity modeling
    Practical claim about ODE discretization error vs. training objective; motivated by Geng et al. and supported only by in-paper ablations (Table 4), not by external few-step benchmarks on dance.
  • domain assumption 35-dim Librosa features (MFCC, Chroma, Peak, Beat, Envelope) at 30 FPS are a sufficient music representation for dance generation
    Follows Lodge; several modern baselines use deep embeddings (Jukebox, MERT). If the music representation is impoverished, model-vs-baseline quality comparisons may be biased.
  • domain assumption FID (kinetic/geometric), DIV, FSR, and BAS are valid proxies for dance quality, physical plausibility, and beat alignment
    Used as ground truth for the 'SOTA' claim; no error bars are reported, and GT FSR 0.216 exceeding generated models' FSR (Table 1) indicates the physical-plausibility metric is not well calibrated as presented.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FlowerDance: MeanFlow for Efficient and Refined 3D Dance Generation." pith.science (2026). https://pith.science/paper/S4CH4BEM

@misc{pith2026251121029,
  author       = {Pith},
  title        = {Pith review of: FlowerDance: MeanFlow for Efficient and Refined 3D Dance Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S4CH4BEM}},
  note         = {Machine review of arXiv:2511.21029}
}
read the original abstract

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

Figures

Figures reproduced from arXiv: 2511.21029 by the authors.

Figure 1
Figure 1. FlowerDance not only generates refined motion with physical plausibility and artistic expressiveness, but also achieves generation [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of model architecture of FlowerDance at different levels. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Qualitative Analysis on a typical Eastern Folk music clip. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Given start (yellow) and end (red) motions, FlowerDance [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. OmniDance: Multimodal Driven Dance Video Generation with Large-scale Internet Data

    cs.CV 2026-06 unverdicted novelty 7.0 of 10

    Introduces CIPE-Dance as the largest dance video dataset and OmniDance framework for unified text-music multimodal dance video generation achieving SOTA on TI2V, MI2V, and MTI2V tasks.

  2. CustomDance: Customized 3D Dance Generation with Coarse-to-Fine Human-Centered Interactive Control

    cs.HC 2026-08 conditional novelty 6.0 of 10

    CustomDance combines an MLLM-based choreographic planner, multimodal dance-phrase retrieval, and diffusion inpainting into one three-stage interactive system for user-customized 3D dance generation.

  3. DiscoForcing: A Unified Framework for Real-Time Audio-Driven Character Control with Diffusion Forcing

    cs.CV 2026-05 unverdicted novelty 6.0 of 10

    DiscoForcing introduces a causal diffusion-forcing model with a hybrid temporal schedule for stable real-time audio-to-motion generation under abrupt audio changes.

  4. PersonaGesture: Single-Reference Co-Speech Gesture Personalization for Unseen Speakers

    cs.CV 2026-05 unverdicted novelty 6.0 of 10

    A no-update diffusion pipeline personalizes co-speech gestures to an unseen speaker from one reference clip by combining zero-initialized style-memory cross-attention during denoising with length-aware latent moment c...

Reference graph

Works this paper leans on

52 extracted references · 15 linked inside Pith · cited by 4 Pith papers

  1. [1]

    Gated multimodal units for infor- mation fusion.arXiv preprint arXiv:1702.01992, 2017

    John Arevalo, Thamar Solorio, Manuel Montes-y G ´omez, and Fabio A Gonz ´alez. Gated multimodal units for infor- mation fusion.arXiv preprint arXiv:1702.01992, 2017. 5

    arXiv 2017

  2. [2]

    Teaching choreography in higher educa- tion: A process continuum model.Research in dance educa- tion, 5(1):45–67, 2004

    Jo Butterworth*. Teaching choreography in higher educa- tion: A process continuum model.Research in dance educa- tion, 5(1):45–67, 2004. 2, 3

    2004

  3. [3]

    Visual generation without guidance.arXiv preprint arXiv:2501.15420, 2025

    Huayu Chen, Kai Jiang, Kaiwen Zheng, Jianfei Chen, Hang Su, and Jun Zhu. Visual generation without guidance.arXiv preprint arXiv:2501.15420, 2025

    arXiv 2025

  4. [4]

    Gaussian mixture flow matching models.arXiv preprint arXiv:2504.05304, 2025

    Hansheng Chen, Kai Zhang, Hao Tan, Zexiang Xu, Fu- jun Luan, Leonidas Guibas, Gordon Wetzstein, and Sai Bi. Gaussian mixture flow matching models.arXiv preprint arXiv:2504.05304, 2025. 2

    arXiv 2025

  5. [5]

    Efficient video prediction via sparsely conditioned flow matching

    Aram Davtyan, Sepehr Sameni, and Paolo Favaro. Efficient video prediction via sparsely conditioned flow matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23263–23274, 2023. 2

    2023

  6. [6]

    Scaling recti- fied flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling recti- fied flow transformers for high-resolution image synthesis. InForty-first international conference on machine learning,

  7. [7]

    Cfg-zero*: Improved classifier-free guidance for flow matching models.arXiv preprint arXiv:2503.18886, 2025

    Weichen Fan, Amber Yijia Zheng, Raymond A Yeh, and Zi- wei Liu. Cfg-zero*: Improved classifier-free guidance for flow matching models.arXiv preprint arXiv:2503.18886, 2025

    arXiv 2025

  8. [8]

    Mambagesture: En- hancing co-speech gesture generation with mamba and dis- entangled multi-modality fusion

    Chencan Fu, Yabiao Wang, Jiangning Zhang, Zhengkai Jiang, Xiaofeng Mao, Jiafu Wu, Weijian Cao, Chengjie Wang, Yanhao Ge, and Yong Liu. Mambagesture: En- hancing co-speech gesture generation with mamba and dis- entangled multi-modality fusion. InProceedings of the 32nd ACM International Conference on Multimedia, pages 10794–10803, 2024. 5

    2024

Show all 52 references
  1. [9]

    Mean flows for one-step genera- tive modeling.arXiv preprint arXiv:2505.13447, 2025

    Zhengyang Geng, Mingyang Deng, Xingjian Bai, J Zico Kolter, and Kaiming He. Mean flows for one-step genera- tive modeling.arXiv preprint arXiv:2505.13447, 2025. 2, 3

    2025 arXiv

  2. [10]

    Mamba: Linear-time sequence modeling with selective state spaces.arXiv preprint arXiv:2312.00752, 2023

    Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces.arXiv preprint arXiv:2312.00752, 2023. 5

    2023 arXiv

  3. [11]

    Efficiently modeling long sequences with structured state spaces.arXiv preprint arXiv:2111.00396, 2021

    Albert Gu, Karan Goel, and Christopher R ´e. Efficiently modeling long sequences with structured state spaces.arXiv preprint arXiv:2111.00396, 2021. 5

    2021 arXiv

  4. [12]

    Generating diverse and natural 3d human motions from text

    Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152–5161, 2022. 2

    2022

  5. [13]

    Tm2t: Stochastic and tokenized modeling for the reciprocal genera- tion of 3d human motions and texts

    Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal genera- tion of 3d human motions and texts. InEuropean Conference on Computer Vision, pages 580–597. Springer, 2022. 2

    2022

  6. [14]

    V oiceflow: Efficient text-to-speech with rectified flow matching

    Yiwei Guo, Chenpeng Du, Ziyang Ma, Xie Chen, and Kai Yu. V oiceflow: Efficient text-to-speech with rectified flow matching. InICASSP 2024-2024 IEEE International Confer- ence on Acoustics, Speech and Signal Processing (ICASSP), pages 11121–11125. IEEE, 2024. 2

    2024

  7. [15]

    Springer, 1993

    Ernst Hairer, Gerhard Wanner, and Syvert P Nørsett.Solv- ing ordinary differential equations I: Nonstiff problems. Springer, 1993

    1993

  8. [16]

    Neue methoden zur approximativen integration der differentialgleichungen einer unabh ¨angigen ver¨anderlichen.Z

    Karl Heun et al. Neue methoden zur approximativen integration der differentialgleichungen einer unabh ¨angigen ver¨anderlichen.Z. Math. Phys, 45:23–38, 1900

    1900

  9. [17]

    Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020. 2

    2020

  10. [18]

    Choreography cgan: gen- erating dances with music beats using conditional generative adversarial networks.Neural Computing and Applications, 33(16):9817–9833, 2021

    Yin-Fu Huang and Wei-De Liu. Choreography cgan: gen- erating dances with music beats using conditional generative adversarial networks.Neural Computing and Applications, 33(16):9817–9833, 2021. 1, 3

    2021

  11. [19]

    Pyramidal flow matching for efficient video generative modeling.arXiv preprint arXiv:2410.05954,

    Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling.arXiv preprint arXiv:2410.05954,

  12. [20]

    Float: Generative motion latent flow matching for audio-driven talking portrait.arXiv preprint arXiv:2412.01064, 2024

    Taekyung Ki, Dongchan Min, and Gyeongsu Chae. Float: Generative motion latent flow matching for audio-driven talking portrait.arXiv preprint arXiv:2412.01064, 2024. 2, 4

    2024

  13. [21]

    A brand new dance partner: Music- conditioned pluralistic dancing controlled by multiple dance genres

    Jinwoo Kim, Heeseok Oh, Seongjean Kim, Hoseok Tong, and Sanghoon Lee. A brand new dance partner: Music- conditioned pluralistic dancing controlled by multiple dance genres. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3490– 3500, 2022. 6

    2022

  14. [22]

    V oicebox: Text-guided multilingual universal speech generation at scale.Advances in neural information processing systems, 36:14005–14034,

    Matthew Le, Apoorv Vyas, Bowen Shi, Brian Karrer, Leda Sari, Rashel Moritz, Mary Williamson, Vimal Manohar, Yossi Adi, Jay Mahadeokar, et al. V oicebox: Text-guided multilingual universal speech generation at scale.Advances in neural information processing systems, 36:14005–14034,

  15. [23]

    Perceiving subjectiv- ity in bodily movement: The case of dancers.Phenomenol- ogy and the Cognitive Sciences, 8:389–408, 2009

    Doroth ´ee Legrand and Susanne Ravn. Perceiving subjectiv- ity in bodily movement: The case of dancers.Phenomenol- ogy and the Cognitive Sciences, 8:389–408, 2009. 7

    2009

  16. [24]

    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++. InProceedings of the IEEE/CVF In- ternational Conference on Computer Vision, pages 13401– 13412, 2021. 1, 2, 3, 6

    2021

  17. [25]

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

    Ronghui Li, Junfan Zhao, Yachao Zhang, Mingyang Su, Zeping Ren, Han Zhang, Yansong Tang, and Xiu Li. Finedance: A fine-grained choreography dataset for 3d full body dance generation. InProceedings of the IEEE/CVF In- ternational Conference on Computer Vision, pages 10234– 1024...

    2023

  18. [26]

    Lodge++: High-quality and long dance gen- eration with vivid choreography patterns.arXiv preprint arXiv:2410.20389, 2024

    Ronghui Li, Hongwen Zhang, Yachao Zhang, Yuxiang Zhang, Youliang Zhang, Jie Guo, Yan Zhang, Xiu Li, and Yebin Liu. Lodge++: High-quality and long dance gen- eration with vivid choreography patterns.arXiv preprint arXiv:2410.20389, 2024. 1, 3

    2024 arXiv

  19. [27]

    Lodge: 9 A coarse to fine diffusion network for long dance generation guided by the characteristic dance primitives

    Ronghui Li, YuXiang Zhang, Yachao Zhang, Hongwen Zhang, Jie Guo, Yan Zhang, Yebin Liu, and Xiu Li. Lodge: 9 A coarse to fine diffusion network for long dance generation guided by the characteristic dance primitives. InProceed- ings of the IEEE/CVF Conference on Computer Vision...

    2024

  20. [28]

    Flow matching for generative mod- eling.arXiv preprint arXiv:2210.02747, 2022

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximil- ian Nickel, and Matt Le. Flow matching for generative mod- eling.arXiv preprint arXiv:2210.02747, 2022. 2, 3

    2022 arXiv

  21. [29]

    Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003, 2022

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003, 2022. 2, 7

    2022 arXiv

  22. [30]

    Gcdance: Genre-controlled 3d full body dance generation driven by music.arXiv preprint arXiv:2502.18309, 2025

    Xinran Liu, Xu Dong, Diptesh Kanojia, Wenwu Wang, and Zhenhua Feng. Gcdance: Genre-controlled 3d full body dance generation driven by music.arXiv preprint arXiv:2502.18309, 2025. 2, 3, 5

    2025 arXiv

  23. [31]

    Smpl: A skinned multi- person linear model

    Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multi- person linear model. InSeminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 851–866. 2023. 3

    2023

  24. [32]

    librosa: Audio and music signal analysis in python

    Brian McFee, Colin Raffel, Dawen Liang, Daniel PW Ellis, Matt McVicar, Eric Battenberg, and Oriol Nieto. librosa: Audio and music signal analysis in python. InSciPy, pages 18–24, 2015. 3

    2015

  25. [33]

    Matcha-tts: A fast tts architecture with conditional flow matching

    Shivam Mehta, Ruibo Tu, Jonas Beskow, ´Eva Sz´ekely, and Gustav Eje Henter. Matcha-tts: A fast tts architecture with conditional flow matching. InICASSP 2024-2024 IEEE In- ternational Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP), pages 11341–11345. IEEE, 2024. 2, 4

    2024

  26. [34]

    Dance studies/cultural studies.Dance Research Journal, 41(1):82–100, 2009

    Gay Morris. Dance studies/cultural studies.Dance Research Journal, 41(1):82–100, 2009. 2, 3

    2009

  27. [35]

    Stablemotion: Training motion cleanup models with un- paired corrupted data.arXiv preprint arXiv:2505.03154, 2025

    Yuxuan Mu, Hung Yu Ling, Yi Shi, Ismael Baira Ojeda, Pengcheng Xi, Chang Shu, Fabio Zinno, and Xue Bin Peng. Stablemotion: Training motion cleanup models with un- paired corrupted data.arXiv preprint arXiv:2505.03154, 2025

    2025

  28. [36]

    Film: Visual reasoning with a general conditioning layer

    Ethan Perez, Florian Strub, Harm De Vries, Vincent Du- moulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. InProceedings of the AAAI con- ference on artificial intelligence, 2018. 5

    2018

  29. [37]

    Bailando: 3d dance generation by actor-critic gpt with choreographic memory

    Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando: 3d dance generation by actor-critic gpt with choreographic memory. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11050– 1...

    2022

  30. [38]

    Bailando++: 3d dance gpt with choreographic memory.IEEE Transac- tions on Pattern Analysis and Machine Intelligence, 2023

    Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando++: 3d dance gpt with choreographic memory.IEEE Transac- tions on Pattern Analysis and Machine Intelligence, 2023. 1, 2, 3, 5

    2023

  31. [39]

    Duolando: Follower gpt with off-policy reinforcement learning for dance accompaniment.arXiv preprint arXiv:2403.18811,

    Li Siyao, Tianpei Gu, Zhitao Yang, Zhengyu Lin, Ziwei Liu, Henghui Ding, Lei Yang, and Chen Change Loy. Duolando: Follower gpt with off-policy reinforcement learning for dance accompaniment.arXiv preprint arXiv:2403.18811,

    arXiv

  32. [40]

    Deep high-resolution representation learning for human pose es- timation

    Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose es- timation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5693–5703,

  33. [41]

    Edge: Editable dance generation from music

    Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 448–458, 2023. 1, 2, 3, 5, 6, 7

    2023

  34. [42]

    Attention is all you need.Advances in Neural Information Processing Systems, 2017

    A Vaswani. Attention is all you need.Advances in Neural Information Processing Systems, 2017. 2, 5

    2017

  35. [43]

    Audiobox: Unified au- dio generation with natural language prompts.arXiv preprint arXiv:2312.15821, 2023

    Apoorv Vyas, Bowen Shi, Matthew Le, Andros Tjandra, Yi-Chiao Wu, Baishan Guo, Jiemin Zhang, Xinyue Zhang, Robert Adkins, William Ngan, et al. Audiobox: Unified au- dio generation with natural language prompts.arXiv preprint arXiv:2312.15821, 2023. 2, 4

    2023 arXiv

  36. [44]

    Mambatalk: Ef- ficient holistic gesture synthesis with selective state space models

    Zunnan Xu, Yukang Lin, Haonan Han, Sicheng Yang, Ronghui Li, Yachao Zhang, and Xiu Li. Mambatalk: Ef- ficient holistic gesture synthesis with selective state space models. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 5

    2024

  37. [45]

    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. InProceed- ings of the 2024 International Conference on Multimedia Re- trieval, pages 675–683, 2024. 3

    2024

  38. [46]

    Cohedancers: Enhanc- ing interactive group dance generation through music-driven coherence decomposition.arXiv preprint arXiv:2412.19123,

    Kaixing Yang, Xulong Tang, Haoyu Wu, Qinliang Xue, Biao Qin, Hongyan Liu, and Zhaoxin Fan. Cohedancers: Enhanc- ing interactive group dance generation through music-driven coherence decomposition.arXiv preprint arXiv:2412.19123,

    arXiv

  39. [47]

    Beatdance: A beat-based model-agnostic contrastive learning framework for music-dance retrieval

    Kaixing Yang, Xukun Zhou, Xulong Tang, Ran Diao, Hongyan Liu, Jun He, and Zhaoxin Fan. Beatdance: A beat-based model-agnostic contrastive learning framework for music-dance retrieval. InProceedings of the 2024 Inter- national Conference on Multimedia Retrieval, pages 11–19,

    2024

  40. [48]

    Matchdance: Collaborative mamba-transformer architecture matching for high-quality 3d dance synthesis.arXiv preprint arXiv:2505.14222, 2025

    Kaixing Yang, Xulong Tang, Yuxuan Hu, Jiahao Yang, Hongyan Liu, Qinnan Zhang, Jun He, and Zhaoxin Fan. Matchdance: Collaborative mamba-transformer architecture matching for high-quality 3d dance synthesis.arXiv preprint arXiv:2505.14222, 2025. 6

    2025

  41. [49]

    Megadance: Mixture-of-experts architecture for genre-aware 3d dance generation.arXiv preprint arXiv:2505.17543, 2025

    Kaixing Yang, Xulong Tang, Ziqiao Peng, Yuxuan Hu, Jun He, and Hongyan Liu. Megadance: Mixture-of-experts architecture for genre-aware 3d dance generation.arXiv preprint arXiv:2505.17543, 2025. 1, 2, 3, 5, 6, 7

    2025

  42. [50]

    Consistency flow matching: Defin- ing straight flows with velocity consistency.arXiv preprint arXiv:2407.02398, 2024

    Ling Yang, Zixiang Zhang, Zhilong Zhang, Xingchao Liu, Minkai Xu, Wentao Zhang, Chenlin Meng, Stefano Er- mon, and Bin Cui. Consistency flow matching: Defin- ing straight flows with velocity consistency.arXiv preprint arXiv:2407.02398, 2024. 2

    2024 arXiv

  43. [51]

    On the continuity of rotation representations in neural networks

    Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5745–5753,

  44. [52]

    Music2dance: Dancenet for music-driven dance generation.ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 18(2):1–21, 2022

    Wenlin Zhuang, Congyi Wang, Jinxiang Chai, Yangang Wang, Ming Shao, and Siyu Xia. Music2dance: Dancenet for music-driven dance generation.ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 18(2):1–21, 2022. 6 10

    2022

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.