Pith. sign in

REVIEW 4 major objections 7 minor 40 references

Dance recalibration for dance coherency with recurrent convolution block

T0 review · 4 major / 7 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read Adding a recurrent recalibration block to the coarse dance decoder of Lodge improves long-dance coherency on the FineDance dataset, raising beat alignment by about 4% over the baseline.

desk verdict A plausible but under-evidenced incremental extension of Lodge: the recurrent recalibration block is sensible, but the coherency claim is unsupported because no ablation or direct consistency metric is provided. read the letter →

arxiv 2502.01190 v1 pith:ZZTGTB35 submitted 2025-02-03 cs.LG cs.AI

Seungho Eum , Ihjoon Cho , Junghyeon Kim This is my paper · ORCID
classification cs.LGcs.AI
keywords dancegenerationcoarse-to-finediffusionrecurrentneuralnetworkcoherencymusic-conditionedFinedatasetbeatalignment
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 claims that long AI-generated dance sequences become noticeably more coherent when the coarse dance stage of the Lodge model receives a recurrent recalibration pass before fine motion synthesis. The proposed model, R-Lodge, adds a stack of Dance Recalibration Blocks to the coarse decoder, so that each coarse motion primitive carries information from the previous primitives. On the FineDance benchmark, the authors report about a 4% improvement in beat alignment and about 6% better FID over Lodge, while keeping runtime essentially unchanged. If the results hold, the work suggests that a lightweight recurrence injected into the coarse stage can fix awkward transitions in long dance generation without rebuilding the diffusion backbone.

What carries the argument

Dance Recalibration (DR) is the central mechanism: a recurrent sequential representation applied only to the 126 joint-rotation channels of the 139-channel coarse dance representation. Each Dance Recalibration Block forms $GRD_i = P(C(GD_i, GRD_{i-1} + G(T)))$, where $C$ concatenates the current global-decoder output with the previously recalibrated primitive plus Gaussian noise, and the pooling block $P$ runs layer normalization followed by three 1D-convolution blocks with activations and batch normalization. This injects a bounded amount of history into every coarse primitive so that the coarse sequence follows the global diffusion distribution while remaining sequentially connected.

What would settle it

Repeatedly regenerate the same test set with different random seeds for both Lodge and R-Lodge and compute confidence intervals for BAS, FID, and FSR; the central claim fails if the roughly 4% BAS gap and roughly 6% FID gap fall within the run-to-run variance. A complementary check is evaluating on a second dataset such as AIST++.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the inconsistency of Lodge's long dances stems from its coarse dance primitives being generated independently, without sequential linkage. R-Lodge addresses this by treating each coarse primitive as a recurrent state: the global diffusion outputs are passed through Dance Recalibration Blocks that concatenate the previous recalibrated primitive with the current one, add Gaussian noise to preserve variety, and pool the pair through layer normalization and 1D convolutions. The authors report state-of-the-art beat alignment (BAS 0.2501 vs Lodge's 0.2397), improved motion quality (FID$_k$ 42.76 vs 45.56, FID$_g$ 32.17 vs 34.29), and essentially unchanged runtime (31.17s vs 30.93s), concluding that the added recurrence makes generated dances smoother and less awkward while preserving the coarse-to-fine generation framework.

Load-bearing premise

The load-bearing premise is that the reported gains over Lodge are real and not noise, since the comparison in Table 1 is a single run without error bars or significance tests and the paper states that no ablation studies were conducted.

Editorial extensions

If this is right

  • Long dance generation can gain coherence at the coarse level without replacing the diffusion backbone, only by adding a recurrent connection between coarse primitives.
  • R-Lodge's BAS gain suggests that beat alignment is improved by smoothing coarse choreographic patterns, not just by local motion refinement.
  • Because runtime increases only from 30.93s to 31.17s for a 1024-frame dance, the recurrent recalibration is cheap enough for practical use.
  • Improving coarse-stage consistency reduces awkward transitions, a known failure mode in coarse-to-fine long dance models.

Reading between the lines

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

  • We infer that the same recurrent recalibration could be applied to other coarse-to-fine generation tasks, such as long video or gesture synthesis, wherever intermediate representations are generated piecewise.
  • The reported drop in diversity (Div$_k$ from 6.75 to 5.79, Div$_g$ from 5.64 to 5.24) suggests the method trades variety for smoothness; a genre-aware conditioning or adaptive noise schedule could recover some diversity.
  • A direct test of the mechanism would be to vary the noise threshold $T$ in the recalibration update; the paper omits ablation due to time constraints, so the optimal noise strength remains an open question.
  • We infer that the claimed 4% BAS improvement may partly reflect the single-run evaluation; replications with multiple seeds are needed before the improvement is taken as stable.
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

4 major / 7 minor

Summary. The paper proposes R-Lodge, an extension of the Lodge coarse-to-fine long dance generation model. A Dance Recalibration module, implemented with N recurrent convolution blocks, is inserted after the coarse dance decoder so that each coarse motion representation is pooled with information from the previous representation, with Gaussian noise added to avoid overly simple outputs. The model is evaluated on FineDance with 1024-frame sequences, and Table 1 reports BAS, FID, FSR, diversity, and runtime. The authors claim improved dance coherency, state-of-the-art coherency on FineDance, and smoother, less awkward motions. The paper explicitly states that no ablation studies were conducted and that evaluation is limited to a single dataset.

Significance. If the reported gains are real, R-Lodge is a simple and computationally cheap modification of a strong baseline: BAS improves from 0.2397 to 0.2501, FIDk from 45.56 to 42.76, FIDg from 34.29 to 32.17, and runtime is nearly unchanged. The use of an external benchmark and comparisons with several published methods are strengths. However, the current evidence is too thin to establish the central claim: there are no error bars or significance tests, no ablation isolating the recurrent block, no direct measure of temporal smoothness or transition consistency, and the diversity metrics drop substantially. The paper would be significant if these gaps were filled, but as it stands it reads more like a preliminary report than a fully supported contribution.

major comments (4)
  1. [§4.1, Table 1] The central claim that R-Lodge improves dance coherency is not directly measured. Table 1 reports BAS, FID, FSR, and diversity, but none of these is a temporal smoothness or transition-consistency metric. BAS is a beat-alignment score, not a measure of awkwardness or sequential consistency, and all generated methods in Table 1 score above the ground-truth BAS of 0.2120, which the paper does not explain. The statements in the contributions and conclusion that movements are 'not awkward' and 'smoother' are therefore unsupported. The authors should add direct coherency metrics such as acceleration continuity, joint-velocity jerk, foot-contact transition statistics, or a controlled user study, and they should report multiple runs with error bars or significance tests.
  2. [§3.1, Table 1] The attribution of the observed gains to Dance Recalibration is untested. The authors state in Section 3.1 that 'due to time constraints, we could not conduct various ablation studies,' and Table 1 reports a single run per method. Consequently, the BAS/FID differences between Lodge and R-Lodge could be due to noise, the added Gaussian noise, the pooling block, or the recurrent structure. At minimum, the authors need an N=0 ablation (i.e., the pooling/noise mechanism without recurrence), ablations over the number of blocks N and the noise threshold, and multiple seeds to establish that the recurrent component is responsible for the improvement.
  3. [§4.1, Table 1] The substantial drop in diversity is a major concern for the coherency claim. Div_k falls from 6.75 to 5.79 and Div_g from 5.64 to 5.24. This pattern is consistent with temporal smoothing or averaging, which could make motions appear more 'consistent' simply by reducing variation. The paper should analyze this trade-off directly and show that the measured improvements are not artifacts of reduced motion variation. Since the method adds Gaussian noise specifically to increase variety, the observed diversity decrease deserves explicit discussion.
  4. [§3.1, Eq. (1), §3.2] The method is not reproducible as specified. The number N of Dance Recalibration Blocks, the Gaussian noise threshold G(Threshold), and the convolution/pooling architecture are not reported, and no code is provided. These hyperparameters are central to the proposed mechanism, so the paper should state their values and, ideally, study their influence. Without this information, a reader cannot implement R-Lodge or verify the reported numbers.
minor comments (7)
  1. [§4.1] The text says 'the FID scores increased by an average of about 6%,' but Table 1 shows that both FIDk and FIDg decrease (45.56 to 42.76 and 34.29 to 32.17). The wording should say 'decreased.'
  2. [Table 1 caption] The caption mentions 'Wins represent the ratio of victories Lodge (DDPM) achieved in the user study,' but no wins or ratio values appear in the table. Either report the user-study results or remove the sentence.
  3. [References] References [17] and [18] are identical, and MNET is cited as [18]. The citation list needs correction; in particular, [18] does not appear to point to the actual MNET paper.
  4. [Eq. (2)] The notation 'Cn i=1(P(C(GDi, GRDi−1 + G(T hreshold))), P(GD0))' is malformed; the aggregation operator needs to be defined clearly.
  5. [§3.1, Eq. (1)] The indexing around Eq. (1) is inconsistent: the text says the pooling input is 'GRDi with added G and GDi+1,' while Eq. (1) reads GRDi = P(C(GDi, GRDi−1 + G(Threshold))). The indices should be unified throughout the description.
  6. [§3.2] The term 'Pooling Block' is used for a stack of 1D convolution layers with layer/batch normalization, which is not pooling in the conventional sense; please clarify the terminology.
  7. [Throughout] There are typos such as 'FineDataset' in Section 4.1 and 'rotaion' in Section 3.1; a careful proofreading pass is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: R-Lodge's recurrent recalibration is evaluated against external FineDance baselines, and no metric or equation reduces to the method's own construction.

full rationale

The paper's derivation chain is architectural and empirical, not definitional. R-Lodge inserts a recurrent Dance Recalibration process into Lodge's coarse dance decoder, and its claims are tested on the FineDance dataset against published baselines (FACT, MNET, Bailando, EDGE, Lodge) using standard metrics (FID, FSR, diversity, BAS, run time). No fitted parameter is used to define the outcome metric, and the reported BAS gain from 0.2397 to 0.2501 is evaluated by an external beat-alignment score rather than by the presence of the recurrent block itself. The equations in Section 3 describe how the recurrent block is constructed, not a prediction that reduces by construction to its own input. The paper contains no self-citations, invokes no uniqueness theorem from the authors' prior work, and does not present an ansatz as externally derived. The absence of ablations, error bars, and a direct consistency metric weakens the support for the central claim, but that is an evidence-strength limitation, not circular reasoning. Therefore, the honest finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method depends on several unstated choices: the noise threshold, the number of blocks N, and the pooling block's internal dimensions. It also assumes BAS captures coherency and that the 126 rotation channels are the right place to add recurrence. No new physical entities are introduced.

free parameters (3)
  • Gaussian noise threshold G(Threshold) = not reported
    Added to previous coarse dance representation to prevent overly smooth motions; no ablation was run to set or validate it (Section 3.1, Eq. 1).
  • Number of Dance Recalibration Blocks N = not reported
    The method says N blocks are used corresponding to the coarse sequence length, but no value is given and no sensitivity study is presented (Section 3.1).
  • Pooling block convolution hyperparameters = not reported
    Three 1D convolution blocks with layer norm and batch norm are described without kernel sizes, channels, or strides (Section 3.2).
assumptions (4)
  • domain assumption The coarse dance representation's 126 joint-rotation channels alone carry the temporal consistency information needed for dance coherency; the foot, root, and rotation channels are excluded from recurrence.
    Section 3.1 states all DR operations are performed on the 126-dim channels because they 'directly impact the dance motion,' without a derivation or ablation showing the other 13 channels are irrelevant.
  • domain assumption The coarse dance sequence in Lodge is short enough that a plain recurrent block will not suffer gradient vanishing.
    Section 3.1 asserts the sequence length is 'not long enough to cause this issue,' justifying the choice of RNN over LSTM/GRU; no gradient analysis or experiment is given.
  • ad hoc to paper Adding Gaussian noise to the previous coarse representation prevents overly simple dances while preserving useful recurrence.
    Section 3.1 introduces this heuristic without derivation or ablation, and the noise threshold is unspecified.
  • domain assumption Beat Alignment Score (BAS) is a valid proxy for dance consistency and coherency.
    The central claim of improved consistency is supported mainly by BAS in Table 1; no user study or dedicated coherence metric is reported, and the user study numbers referenced by the table caption are absent.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dance recalibration for dance coherency with recurrent convolution block." pith.science (2026). https://pith.science/paper/ZZTGTB35

@misc{pith2026250201190,
  author       = {Pith},
  title        = {Pith review of: Dance recalibration for dance coherency with recurrent convolution block},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZZTGTB35}},
  note         = {Machine review of arXiv:2502.01190}
}
abstract

With the recent advancements in generative AI such as GAN, Diffusion, and VAE, the use of generative AI for dance generation has seen significant progress and received considerable interest. In this study, We propose R-Lodge, an enhanced version of Lodge. R-Lodge incorporates Recurrent Sequential Representation Learning named Dance Recalibration to original coarse-to-fine long dance generation model. R-Lodge utilizes Dance Recalibration method using $N$ Dance Recalibration Block to address the lack of consistency in the coarse dance representation of the Lodge model. By utilizing this method, each generated dance motion incorporates a bit of information from the previous dance motions. We evaluate R-Lodge on FineDance dataset and the results show that R-Lodge enhances the consistency of the whole generated dance motions.

Figures

Figures reproduced from arXiv: 2502.01190 by the authors.

Figure 1
Figure 1. overall procedure of Pooling processing by our Poo [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Overall of the Dance Recalibration Block Structur [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The structure of the dance recalibration block [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: overall procedure of Pooling processing by our Poo [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 31 canonical work pages

  1. [1]

    Fpn-gan: multi-class small object detection in remote sensing images

    Tanvir Ahmad, Xiaona Chen, Ali Syed Saqlain, and Yinglon g Ma. Fpn-gan: multi-class small object detection in remote sensing images. In 2021 IEEE 6th International Conference on Cloud Computing a nd Big Data Analytics (ICCCBDA), pages 478–482. IEEE, 2021

    work page 2021

  2. [2]

    Text2action: Generative adversarial synthesis from language to action

    Hyemin Ahn, Timothy Ha, Y unho Choi, Hwiyeon Y oo, and Song hwai Oh. Text2action: Generative adversarial synthesis from language to action. In 2018 IEEE International Conference on Robotics and Automat ion (ICRA), pages 5915–5920. IEEE, 2018

    work page 2018

  3. [3]

    Language2 pose: Natural language grounded pose forecasting

    Chaitanya Ahuja and Louis-Philippe Morency. Language2 pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV) , pages 719–728. IEEE, 2019

    work page 2019

  4. [4]

    Vivit: A video vision transformer

    Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun , Mario Lu ˇci´c, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on co mputer vision, pages 6836–6846, 2021

    work page 2021

  5. [5]

    Teach: Temporal action composition for 3d humans

    Nikos Athanasiou, Mathis Petrovich, Michael J Black, an d Gül V arol. Teach: Temporal action composition for 3d humans. In 2022 International Conference on 3D Vision (3DV) , pages 414–423. IEEE, 2022

    work page 2022

  6. [6]

    Executing your commands via motion diffusion in latent space

    Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Ch en, and Gang Y u. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision a nd Pattern Recognition, pages 18000–18010, 2023

    work page 2023

  7. [7]

    Generative ad- versarial graph convolutional networks for human action sy nthesis

    Bruno Degardin, Joao Neves, V asco Lopes, Joao Brito, Ehs an Y aghoubi, and Hugo Proença. Generative ad- versarial graph convolutional networks for human action sy nthesis. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages 1150–1159, 2022

    work page 2022

  8. [8]

    Gate-variants of gated recur rent unit (gru) neural networks

    Rahul Dey and Fathi M Salem. Gate-variants of gated recur rent unit (gru) neural networks. In 2017 IEEE 60th international midwest symposium on circuits and systems (M WSCAS), pages 1597–1600. IEEE, 2017. 7 Running Title for Header

    work page 2017

Show all 40 references
  1. [9]

    Multiscale vision transformers

    Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Y anghao Li, Zh icheng Y an, Jitendra Malik, and Christoph Feicht- enhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF international conference on co mputer vision, pages 6824–6835, 2021

    2021

  2. [10]

    Long video generation with time-agnostic vqgan and time-se nsitive transformer

    Songwei Ge, Thomas Hayes, Harry Y ang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and time-se nsitive transformer. In European Conference on Computer Vision, pages 102–118. Springer, 2022

    2022

  3. [11]

    Synthesis of com- positional animations from textual descriptions

    Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christia n Theobalt, and Philipp Slusallek. Synthesis of com- positional animations from textual descriptions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1396–1406, 2021

    2021

  4. [12]

    Tm2t: Stoc hastic and tokenized modeling for the recipro- cal generation of 3d human motions and texts

    Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stoc hastic and tokenized modeling for the recipro- cal generation of 3d human motions and texts. In European Conference on Computer Vision , pages 580–597. Springer, 2022

    2022

  5. [13]

    Denoising di ffusion probabilistic models

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

    2020

  6. [14]

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

    Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai , Lei Y ang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535 , 2022

    2022 arXiv

  7. [15]

    Cogvideo: Large-scale pretraining for text-to-video generation via transformers

    Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868 , 2022

    2022 arXiv

  8. [16]

    Dance revolution: Long-term dance generation with music via curriculum learning

    Ruozi Huang, Huang Hu, Wei Wu, Kei Sawada, Mi Zhang, and D axin Jiang. Dance revolution: Long-term dance generation with music via curriculum learning. arXiv preprint arXiv:2006.06119 , 2020

    2006 arXiv

  9. [18]

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

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

    2022

  10. [19]

    Afi-gan: Improving fea ture interpolation of feature pyramid networks via adversarial training for object detection

    Seong-Ho Lee and Seung-Hwan Bae. Afi-gan: Improving fea ture interpolation of feature pyramid networks via adversarial training for object detection. Pattern Recognition, 138:109365, 2023

    2023

  11. [20]

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

    Ronghui Li, Y uxiang Zhang, Y achao Zhang, Hongwen Zhang, Jie Guo, Y an Zhang, Y ebin Liu, and Xiu Li. Lodge: A coarse to fine diffusion network for long dance generation g uided by the characteristic dance primitives. In IEEE/CVF Conf. on Computer Vision and Pattern Recognitio...

    2024

  12. [21]

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

    Ronghui Li, Junfan Zhao, Y achao Zhang, Mingyang Su, Zep ing Ren, Han Zhang, Y ansong Tang, and Xiu Li. Finedance: A fine-grained choreography dataset for 3d full b ody dance generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 10234–10...

    2023

  13. [22]

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

    Ruilong Li, Shan Y ang, David A Ross, and Angjoo Kanazawa . Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Co mputer Vision, pages 13401–13412, 2021

    2021

  14. [23]

    Mvitv2: Improved multiscale vision transformers for classification and detection

    Y anghao Li, Chao-Y uan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feicht- enhofer. Mvitv2: Improved multiscale vision transformers for classification and detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogn...

    2022

  15. [24]

    Video swin transformer

    Ze Liu, Jia Ning, Y ue Cao, Yixuan Wei, Zheng Zhang, Steph en Lin, and Han Hu. Video swin transformer. In Proceedings of the IEEE/CVF conference on computer vision a nd pattern recognition, pages 3202–3211, 2022

    2022

  16. [25]

    Latte: Latent diffusion transformer for video generation

    Xin Ma, Y aohui Wang, Gengyun Jia, Xinyuan Chen, Ziwei Li u, Y uan-Fang Li, Cunjian Chen, and Y u Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048 , 2024

    2024 arXiv

  17. [26]

    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. In SciPy, pages 18–24, 2015

    2015

  18. [27]

    Learn2dance: Learning statistical music-to-dance mappings for choreography synthesis

    Ferda Ofli, Engin Erzin, Yücel Y emez, and A Murat Tekalp. Learn2dance: Learning statistical music-to-dance mappings for choreography synthesis. IEEE Transactions on Multimedia, 14(3):747–759, 2011

    2011

  19. [28]

    Temos : Generating diverse human motions from textual descriptions

    Mathis Petrovich, Michael J Black, and Gül V arol. Temos : Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, pages 480–497. Springer, 2022

    2022

  20. [29]

    Long short-term memory recurrent neural network architectures for large scale acoustic modeling

    H Sak. Long short-term memory recurrent neural network architectures for large scale acoustic modeling. In Interspeech 2014. INTERSPEECH, 2014. 8 Running Title for Header

    2014

  21. [30]

    Recurrent neural networks (rnns): A ge ntle introduction and overview

    Robin M Schmidt. Recurrent neural networks (rnns): A ge ntle introduction and overview. arXiv preprint arXiv:1912.05911, 2019

    1912 arXiv

  22. [31]

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

    Li Siyao, Weijiang Y u, Tianpei Gu, Chunze Lin, Quan Wang , Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando: 3d dance generation by actor-critic gpt with chor eographic memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 110...

    2022

  23. [32]

    Denoisi ng diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoisi ng diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020

    2010 arXiv

  24. [33]

    Motionclip: Exposing human motion generation to clip space

    Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022

    2022

  25. [34]

    Human motion diffusion model

    Guy Tevet, Sigal Raab, Brian Gordon, Y onatan Shafir, Dan iel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916 , 2022

    2022 arXiv

  26. [35]

    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 Re cognition, pages 448–458, 2023

    2023

  27. [36]

    Transflower: probabilistic autoregressive dance generation with multimodal attention

    Guillermo V alle-Pérez, Gustav Eje Henter, Jonas Besko w, 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

    2021

  28. [37]

    Videogpt: Video generation using vq-vae and transformers

    Wilson Y an, Y unzhi Zhang, Pieter Abbeel, and Aravind Sr inivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157 , 2021

    2021 arXiv

  29. [38]

    Magvit: Masked gener ative video transformer

    Lijun Y u, Y ong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming- Hsuan Y ang, Y uan Hao, Irfan Essa, et al. Magvit: Masked gener ative video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni tio...

    2023

  30. [39]

    T2m-gpt: Generating human motion from textual desc riptions with discrete representations

    Jianrong Zhang, Y angsong Zhang, Xiaodong Cun, Shaoli H uang, Y ong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual desc riptions with discrete representations. arXiv preprint arXiv:2301.06052, 2023

    2023 arXiv

  31. [40]

    Dance dance generation: Motion transfer for internet videos

    Yipin Zhou, Zhaowen Wang, Chen Fang, Trung Bui, and Tama ra Berg. Dance dance generation: Motion transfer for internet videos. In Proceedings of the IEEE/CVF International Conference on Co mputer Vision W orkshops, pages 0–0, 2019

    2019

  32. [41]

    Music2dance: Dancenet for music-driven dance generation

    Wenlin Zhuang, Congyi Wang, Jinxiang Chai, Y angang Wan g, 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. 9

    2022

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.