Pith. sign in

REVIEW 3 major objections 5 minor 4 cited by

MulSMo: Multimodal Stylized Motion Generation by Bidirectional Control Flow

T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read MulSMo claims that letting the style signal and content signal adjust each other during generation, rather than letting style only influence content, produces stylized motion that keeps the style's dynamics while still following the…

desk verdict A solid, well-engineered stylized motion generation paper with a genuine architectural novelty and a real, fixable evaluation concern. read the letter →

arxiv 2412.09901 v2 pith:TROCD2YK submitted 2024-12-13 cs.CV

Zhe Li , Yisheng He , Lei Zhong , Weichao Shen , Qi Zuo , Lingteng Qiu , Zilong Dong , Laurence Tianruo Yang
show 1 more author
Weihao Yuan
This is my paper · ORCID
classification cs.CV
keywords humanmotiongenerationstylizeddiffusionbidirectionalcontrolflowcontrastivelearningVAEmultimodalstyle
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

Stylized motion generation has a built-in tension: the content prompt says what the person does, and the style says how, but when style and content diverge, one-way injection of style into a frozen content generator tends to flatten the style's motion dynamics. This paper proposes that the style representation should also adapt to the content, through a bidirectional control flow between the style encoder and the content encoder, and reports that this reduces style-content conflict and better preserves the style. It further aligns style embeddings from motion, text, and image with contrastive learning, so a single label word or a picture can stand in for a reference motion clip. A motion-aligned temporal VAE is introduced as the latent diffusion base, improving the underlying text-to-motion generator. If the claims hold, stylized motion generation becomes more controllable and more faithful to styles that are defined by how the body moves, not just how it poses.

What carries the argument

The central object is the bidirectional control flow itself: paired fusion modules placed at every encoder block of the denoiser, one sending style features into the content stream and one sending content features back into the style stream, with both initialized as zero linear layers so training starts from the frozen generation model. Alongside it, a contrastive learning scheme with a small MLP adaptor aligns CLIP text and image embeddings with style-motion embeddings, which is what converts a label word or a picture into a usable style control. The motion-aligned temporal VAE uses a temporal gating mechanism to suppress noise in attention and aligns its latent tokens with a pretrained motion transformer, giving the diffusion process a more temporally faithful latent representation.

What would settle it

Ask human raters to match generated motions to style labels for dynamics-heavy styles such as 'flapping' versus 'aeroplane' under identical content prompts, comparing MulSMo with a one-way control baseline; if raters do not select MulSMo's outputs as more faithful to the intended label, the claim that bidirectional flow preserves style dynamics is refuted.

Watch

Extended reading notes

Core claim

The paper's central claim is that the conflict between style and content in stylized motion generation is not just a matter of fusion strength but a structural flaw in the direction of control. Existing ControlNet-style methods let the style stream influence the content stream only, so when the style motion's dynamics diverge from the content text, the model tends to drop the conflicting dynamics. MulSMo instead wires zero-initialized linear fusion blocks in both directions at every encoder layer, so the style stream is continuously reshaped by the content stream while the content stream is shaped by the style. The paper argues this co-adaptation preserves the temporal dynamics that distinguish nearby styles such as 'flapping' and 'aeroplane'. The same framework is extended to text and image style inputs by a contrastive alignment between CLIP embeddings and style-motion embeddings, and to a stronger latent diffusion base by a motion-aligned temporal VAE.

Load-bearing premise

The entire style signal, both for guidance and for the reported metric, is defined by a pretrained style classifier trained on 100STYLE labels; if that classifier does not agree with how humans perceive style, the model may be optimizing toward a proxy rather than toward real style transfer.

Editorial extensions

If this is right

  • Styles that differ mainly in dynamics, such as 'flapping' versus 'aeroplane', can be told apart in generation instead of collapsing to similar motions.
  • A single word or an image can act as the style input, so users no longer need a reference motion capture clip to specify a style.
  • The motion-aligned temporal VAE improves text-to-motion generation even with a CLIP text encoder, not only with the stronger LaMP encoder.
  • Because the generation network stays frozen, the style network can be trained once and then reused across content prompts and style modalities without per-style fine-tuning.
  • The same bidirectional mechanism carries over to motion style transfer on datasets unseen during training, such as Xia, indicating that the learned control flow generalizes.

Reading between the lines

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

  • The bidirectional feedback idea is a general architectural principle: any conditional diffusion model that injects a control stream into a frozen generator could benefit from letting the control stream be reshaped by the content stream, especially when the control signal and the content can conflict.
  • The contrastive alignment between CLIP embeddings and motion embeddings is not limited to text and images; an audio or video style encoder that shares the same latent space could be added without retraining the motion generation network.
  • Because the style guidance and the style evaluation both rely on the same 100STYLE-trained classifier, a natural next test is to measure the model's style accuracy with independent human raters or an out-of-distribution style classifier; that would separate genuine style transfer from optimization against the proxy.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 MulSMo, a stylized human motion generation method that combines three components: a bidirectional control flow between the style and content encoders of a ControlNet-style architecture, contrastive learning to align motion, text, and image style embeddings, and a motion-aligned temporal VAE (MaTLD) for latent diffusion. Experiments on 100STYLE, HumanML3D, KIT-ML, and Xia datasets report improvements over SMooDi and other baselines in Style Recognition Accuracy (SRA), FID, R-precision, and foot-skating ratio, as well as qualitative results for text- and image-guided style control. The central claim is that the bidirectional feedback adjusts the style toward the content, alleviating style-content conflict and better preserving style dynamics.

Significance. If the claimed results hold, the bidirectional control flow is a conceptually simple and potentially useful architectural modification for stylized motion generation, and the multimodal style control (motion/text/image) is practically valuable. The paper ships pseudocode and commits to public code, includes ablations that isolate the bidirectional encoder flow, provides a user study, and reports results on multiple datasets. However, the main quantitative evaluation relies on a style classifier that is also used as the feature extractor for guidance, which is a correctness risk that needs to be resolved before the central claim can be fully accepted. The lack of error bars on the primary stylized-generation tables is an additional weakness.

major comments (3)
  1. [Section 3.5, Eq. (5); supplementary Section C] The classifier-based style guidance uses a style feature extractor f obtained by training a style classifier on 100STYLE labels and removing its final layer, while the Style Recognition Accuracy (SRA) metric is computed with a pretrained style classifier on the same 100STYLE label set. Since Eq. (5) explicitly minimizes the L1 distance between f(ˆx0) and f(s) at every denoising step, the reported SRA improvements (e.g., Table 1: 78.029 vs. 72.418) may partly reflect optimization against the evaluation metric rather than perceptually faithful style transfer. The paper provides no validation that f's style judgments agree with human perception; the user study in supplementary Section H compares only whole systems (Ours vs. SMooDi) and does not assess the classifier's accuracy. Please address this by (i) using a held-out or independently trained style classifier for SRA, (ii) validating the classifier against human style annotations, or (iii) demonstrating that guidance using a different feature extractor still yields SRA gains.
  2. [Tables 1–3 and Table 5] The central quantitative claims are reported without error bars or significance tests. For example, Table 2 shows an SRA improvement of 1.089 percentage points between BiFlow+MLD (70.238) and the full model (71.327), and Table 3 shows a difference of 0.092 points (67.870 vs. 67.962); without variance estimates or significance tests, these differences may be within sampling noise. Table 4 already reports means with standard deviations for the text-to-motion task, so please extend this protocol to the stylized generation and style transfer tables.
  3. [Section 4.3, Table 5] The ablation of control-flow architectures is not a fully controlled comparison. Row (a) (unidirectional) uses the optimal guidance scale and style guidance scale from SMooDi, while rows (b)–(d) use the same fixed scales. To support the claim that bidirectional encoder fusion alone causes the observed improvement, either optimize the guidance scales for architecture (a) under the same protocol or add an additional row for (b) using (a)'s scales. Otherwise the difference may be attributed to guidance-scale tuning rather than the architectural change.
minor comments (5)
  1. [Section 4.3] The text refers to 'Table 4' when reporting the ablation results; the corresponding table is numbered Table 5.
  2. [Section 3.6, Eq. (9)] The loss is defined as Lall = Lstd + λprLpr, but the text states 'λpr and λcyc are hyperparameters' without ever defining or using λcyc. Either remove λcyc or introduce the corresponding cyclic loss.
  3. [Equation (5)] The gradient term is written as +τ ∇zt G(zt,t,s), and later in the text and supplementary τ is set to negative values (e.g., τ = −0.2, τ = −0.4). The sign convention should be stated explicitly to avoid confusion about whether the guidance moves toward or away from the style embedding.
  4. [Supplementary Section B] The evaluation excludes the ACT group of style labels from SRA computation 'following SMooDi'. This filtering is not mentioned in the main evaluation section; please state it there for transparency, as it affects the interpretation of the reported SRA numbers.
  5. [Supplementary Section H] The user study says each participant is 'provided with 24 video segments' but does not clarify whether these are 24 pairwise comparisons (Ours vs. SMooDi) or something else. Please specify the exact number of trials and how preferences were aggregated.

Circularity Check

1 steps flagged · score 5.0 of 10

The style classifier used for classifier-based guidance (Eq. 5) is also used to compute the SRA metric, so reported style-reflection gains partly measure optimization against the evaluation classifier.

  1. self definitional [Section 3.5 (Eq. 5) and Supplementary Section C (Evaluation Metrics)]
    "G(zt, t,s) =|f ( ˆx0) − f (s)|, ... The style feature extractor is obtained by training a style classifier on the 100STYLE dataset [26] and removing its final layer. ... We compute the SRA for the generated motion with a pretrainde style classifier."

    Equation (5) defines the classifier-based style guidance as the gradient of the L1 distance between f(ˆx0) and f(s), where Section 3.5 states that f is 'obtained by training a style classifier on the 100STYLE dataset and removing its final layer.' Supplementary Section C states that 'SRA for the generated motion' is computed 'with a pretrainde style classifier.' The paper describes only this one 100STYLE-trained style classifier, so the same network (or a network trained on the same labels) supplies both the optimization target and the evaluation metric.

full rationale

Score 5 reflects a partial circularity in the headline SRA metric, not a fully circular derivation. The bidirectional control flow itself is an architectural contribution with independent evidence: the Table 5 ablation compares unidirectional and bidirectional flows under comparable guidance settings, and the paper reports independent metrics (FID, R-precision, CRA, foot-skating ratio) plus a user study that does not rely on the SRA classifier. However, the main style-reflection evidence (SRA in Tables 1-3) is not independent of the method's own classifier-based guidance: Eq. (5) steers generated latents toward f(s), and Supplementary C computes SRA with a pretrained style classifier; the paper describes a single 100STYLE-trained style classifier for both roles. This makes the absolute SRA improvements, and the advantage over baselines that do not use this classifier guidance (MLD+Aberman, MLD+Motion Puzzle), partly a metric-optimization artifact. No load-bearing self-citation chain was found: LaMP [24] is a same-group pretrained model used as a component, but the supplementary CLIP-based MaTLD results show the VAE contribution is not solely due to LaMP; SMooDi [50] is used as a baseline and protocol source, not as proof of the method's claims.

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

The central claim rests on the validity of the style classifier (used for both guidance and evaluation), the retargeting pipeline, the pseudo text labels, and the contrastive alignment of modalities. No new physical entities are introduced.

free parameters (5)
  • classifier-free content guidance weight w_c = 15
    Selected by sweeping w_c from 10 to 16 in supplementary Figure 10 to balance content preservation and style accuracy.
  • classifier-free style guidance weight w_s = 1.6
    Selected by sweeping w_s from 0.5 to 4.5 in supplementary Figure 9; higher w_s improves SRA but degrades content metrics.
  • classifier-based style guidance weight tau = -0.2 for stylized generation, -0.4 for style transfer
    Set manually; tau=-0.2 used in main experiments, tau=-0.4 for motion style transfer (supplementary Section F).
  • content prior preservation loss weight lambda_pr = not reported
    Equation 9 defines L_all = L_std + lambda_pr * L_pr, but lambda_pr is not given in the paper.
  • InfoNCE temperature tau_t = not reported
    Equation 1b uses tau_t as temperature, value not specified.
assumptions (4)
  • domain assumption The style classifier trained on 100STYLE labels is a valid style oracle for both evaluation and guidance.
    Used in Section 3.5 for classifier guidance and supplementary Section C for SRA; if this classifier is biased, both metric and guidance are compromised.
  • domain assumption Motion sequences from 100STYLE can be retargeted to the HumanML3D SMPL skeleton without losing stylistic dynamics.
    Supplementary Section B describes retargeting; the whole pipeline depends on style being preserved through retargeting and the HumanML3D root-velocity representation.
  • domain assumption Pseudo text descriptions generated by MotionGPT for 100STYLE style motions are reliable content labels for training.
    Supplementary Section B states 'use MotionGPT [18] to generate pseudo text descriptions'; the content prior preservation loss and evaluation rely on these pseudo labels.
  • ad hoc to paper Contrastive learning on the 100STYLE dataset alone can align text/image embeddings with motion style embeddings sufficiently for multimodal control.
    Section 3.4 assumes CLIP text/image features projected by a simple MLP adaptor align with style motion embeddings via InfoNCE on 100STYLE; generalization to unseen styles is only shown qualitatively.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MulSMo: Multimodal Stylized Motion Generation by Bidirectional Control Flow." pith.science (2026). https://pith.science/paper/TROCD2YK

@misc{pith2026241209901,
  author       = {Pith},
  title        = {Pith review of: MulSMo: Multimodal Stylized Motion Generation by Bidirectional Control Flow},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TROCD2YK}},
  note         = {Machine review of arXiv:2412.09901}
}
read the original abstract

Generating motion sequences conforming to a target style while adhering to the given content prompts requires accommodating both the content and style. In existing methods, the information usually only flows from style to content, which may cause conflict between the style and content, harming the integration. Differently, in this work we build a bidirectional control flow between the style and the content, also adjusting the style towards the content, in which case the style-content collision is alleviated and the dynamics of the style is better preserved in the integration. Moreover, we extend the stylized motion generation from one modality, i.e. the style motion, to multiple modalities including texts and images through contrastive learning, leading to flexible style control on the motion generation. Extensive experiments demonstrate that our method significantly outperforms previous methods across different datasets, while also enabling multimodal signals control. The code of our method will be made publicly available.

Figures

Figures reproduced from arXiv: 2412.09901 by the authors.

Figure 1
Figure 1. MulSMo enables multimodal signals to control the styl￾ized motion generation. style of the generated motion. AdaIn is an effective mech￾anism for enforcing global styles, yet it inherently has lim￾itations when it comes to capturing local details of styles. LoRA has been successfully utilized to incorporate the tar￾get style into the generated content, but it requires training a separate LoRA model for each style, w… view at source ↗
Figure 2
Figure 2. Overview of MulSMo. Our approach can take in various modalities as style signals, such as style motion sequences, texts, and [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Motion-aligned Temporal VAE. they rely on motion sequences to extract content informa￾tion, while stylized motion generation [3, 50] can generate motions according to given texts while conforming to a tar￾get style. Inspired by the ControlNet, SMooDi [50] enables stylized motion generation from content text descriptions and style motion sequences by introducing two kinds of style guidance to steer the pretrained tex… view at source ↗
Figures from the paper (8 more)
Figure 5
Figure 5. Figure 5: Contrastive learning for enabling multi-modality styl [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: Qualitative results of using text or image as the style control signal. [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: Qualitative results of stylized motion generation on the 100STYLE dataset. The style input is a style motion while the content [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: User study on the stylized motion generation task. [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Varying SRA, FID, R-Precision, and MM Dist under different values of classifier-free style guidance [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: Varying SRA, FID, R-Precision, and MM Dist under different values of classifier-free content guidance [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: More qualitative results of stylized motion generation on the 100STYLE dataset. The style name is annotated on the top-left of [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]
Figure 12
Figure 12. Figure 12: Image control stylized motion generation of styles not present in the training data, validating the generalization ability of our [PITH_FULL_IMAGE:figures/full_fig_p016_12.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. Absolute Coordinates Make Motion Generation Easy

    cs.CV 2025-05 conditional novelty 7.0 of 10

    Using absolute 3D joint coordinates with a plain Transformer and velocity-prediction diffusion outperforms the standard local-relative motion representation, improving fidelity and enabling direct control.

  2. Rethinking Diffusion for Text-Driven Human Motion Generation: Redundant Representations, Evaluation, and Masked Autoregression

    cs.CV 2024-11 conditional novelty 7.0 of 10

    A masked-autoregressive diffusion model trained on a compact essential-feature latent space claims state-of-the-art text-to-motion generation under a new essential-dimension evaluation protocol.

  3. ClusterStyle: Modeling Intra-Style Diversity with Prototypical Clustering for Stylized Motion Generation

    cs.CV 2025-12 conditional novelty 6.0 of 10

    ClusterStyle clusters each motion style into global and local prototypes and conditions a latent diffusion model on them, improving stylized motion generation fidelity and enabling controllable within-style diversity ...

  4. Motion Generation: A Survey of Generative Approaches and Benchmarks

    cs.CV 2025-07 unverdicted novelty 3.0 of 10

    A structured survey that categorizes recent motion generation methods by underlying generative approach and compiles datasets, metrics, and statistical trends.

Reference graph

Works this paper leans on

59 extracted references · 42 canonical work pages · cited by 4 Pith papers

  1. [1]

    Unpaired motion style transfer from video to animation

    Kfir Aberman, Yijia Weng, Dani Lischinski, Daniel Cohen- Or, and Baoquan Chen. Unpaired motion style transfer from video to animation. ACM Transactions on Graphics, 2020. 3, 7

    work page 2020

  2. [2]

    Lead: Latent realignment for human motion diffu- sion

    Nefeli Andreou, Xi Wang, Victoria Fern ´andez Abrevaya, Marie-Paule Cani, Yiorgos Chrysanthou, and Vicky Kalo- geiton. Lead: Latent realignment for human motion diffu- sion. arXiv preprint arXiv:2410.14508, 2024. 3

    arXiv 2024

  3. [3]

    Gesturediffuclip: Gesture diffusion model with clip latents.ACM Transactions on Graphics, 2023

    Tenglong Ao, Zeyi Zhang, and Libin Liu. Gesturediffuclip: Gesture diffusion model with clip latents.ACM Transactions on Graphics, 2023. 3, 5

    work page 2023

  4. [4]

    Gener- ating human motion in 3d scenes from text descriptions

    Zhi Cen, Huaijin Pi, Sida Peng, Zehong Shen, Minghui Yang, Shuai Zhu, Hujun Bao, and Xiaowei Zhou. Gener- ating human motion in 3d scenes from text descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 1855–1866, 2024. 3

    2024

  5. [5]

    Motionclr: Motion generation and training-free edit- ing via understanding attention mechanisms

    Ling-Hao Chen, Wenxun Dai, Xuan Ju, Shunlin Lu, and Lei Zhang. Motionclr: Motion generation and training-free edit- ing via understanding attention mechanisms. arXiv preprint arXiv:2410.18977, 2024. 3

    arXiv 2024

  6. [6]

    Executing your commands via motion diffusion in latent space

    Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition, 2023. 3, 4, 8, 11

    work page 2023

  7. [7]

    Diffh2o: Diffusion-based synthesis of hand-object interactions from textual descriptions

    Sammy Christen, Shreyas Hampali, Fadime Sener, Edoardo Remelli, Tomas Hodan, Eric Sauser, Shugao Ma, and Bugra Tekin. Diffh2o: Diffusion-based synthesis of hand-object interactions from textual descriptions. In SIGGRAPH Asia 2024 Conference Papers, 2024. 3

    work page 2024

  8. [8]

    Motionlcm: Real-time control- lable motion generation via latent consistency model

    Wenxun Dai, Ling-Hao Chen, Jingbo Wang, Jinpeng Liu, Bo Dai, and Yansong Tang. Motionlcm: Real-time control- lable motion generation via latent consistency model. arXiv preprint arXiv:2404.19759, 2024. 3

    arXiv 2024

Show all 59 references
  1. [9]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Informa- tion Processing Systems, 2021. 6, 12

    2021

  2. [10]

    Im- age style transfer using convolutional neural networks

    Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Im- age style transfer using convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016. 3

    2016

  3. [11]

    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. In Proceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition, 2022. 3, 7, 8, 11

    2022

  4. [12]

    Momask: Generative masked model- ing of 3d human motions

    Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked model- ing of 3d human motions. In Proceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition, 2024. 3

    2024

  5. [13]

    Generative human motion stylization in latent space

    Chuan Guo, Yuxuan Mu, Xinxin Zuo, Peng Dai, Youliang Yan, Juwei Lu, and Li Cheng. Generative human motion stylization in latent space. arXiv preprint arXiv:2401.13505,

    arXiv

  6. [14]

    Momentum contrast for unsupervised visual rep- resentation learning

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual rep- resentation learning. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 9729–9738, 2020. 4

    2020

  7. [15]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1, 3, 5

    2021 arXiv

  8. [16]

    Arbitrary style transfer in real-time with adaptive instance normalization

    Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceed- ings of the International Conference on Computer Vision ,

  9. [17]

    Mo- tion puzzle: Arbitrary motion style transfer by body part

    Deok-Kyeong Jang, Soomin Park, and Sung-Hee Lee. Mo- tion puzzle: Arbitrary motion style transfer by body part. ACM Transactions on Graphics, 2022. 2, 3, 7, 11

    2022

  10. [18]

    Motiongpt: Human motion as a foreign lan- guage

    Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign lan- guage. Advances in Neural Information Processing Systems, 36:20067–20079, 2023. 11

    2023

  11. [19]

    Customizing text-to-image models with a single image pair

    Maxwell Jones, Sheng-Yu Wang, Nupur Kumari, David Bau, and Jun-Yan Zhu. Customizing text-to-image models with a single image pair. arXiv preprint arXiv:2405.01536, 2024. 3

    2024 arXiv

  12. [20]

    Gmd: Controllable human motion synthesis via guided diffusion models

    Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Gmd: Controllable human motion synthesis via guided diffusion models. In Proceed- ings of the International Conference on Computer Vision ,

  13. [21]

    Most: Motion style transformer between diverse ac- tion contents

    Boeun Kim, Jungho Kim, Hyung Jin Chang, and Jin Young Choi. Most: Motion style transformer between diverse ac- tion contents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 1705– 1714, 2024. 1, 3

    2024

  14. [22]

    Perceptual characteristics by motion style category

    Hye Ji Kim and Sung-Hee Lee. Perceptual characteristics by motion style category. In Eurographics (Short Papers),

  15. [23]

    Auto-encoding vari- ational bayes, 2013

    Diederik P Kingma, Max Welling, et al. Auto-encoding vari- ational bayes, 2013. 4

    2013

  16. [24]

    Lamp: Language-motion pretraining for motion generation, retrieval, and captioning

    Zhe Li, Weihao Yuan, Yisheng He, Lingteng Qiu, Shen- hao Zhu, Xiaodong Gu, Weichao Shen, Yuan Dong, Zi- long Dong, and Laurence T Yang. Lamp: Language-motion pretraining for motion generation, retrieval, and captioning. arXiv preprint arXiv:2410.07093, 2024. 3, 4, 7, 13

    arXiv

  17. [25]

    Decoupled weight decay regularization

    I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 7, 11

    2017 arXiv

  18. [26]

    Ian Mason, Sebastian Starke, and Taku Komura. Real-time style modelling of human locomotion via feature-wise trans- formations and local motion phases.Proceedings of the ACM on Computer Graphics and Interactive Techniques, 2022. 3, 6, 7, 11

    2022

  19. [27]

    Soomin Park, Deok-Kyeong Jang, and Sung-Hee Lee. Di- verse motion stylization for multiple style domains via spatial-temporal graph-based generative model.Proceedings of the ACM on Computer Graphics and Interactive Tech- niques, 2021. 3

    2021

  20. [28]

    Hoi-diff: Text-driven synthe- sis of 3d human-object interactions using diffusion models

    Xiaogang Peng, Yiming Xie, Zizhao Wu, Varun Jampani, Deqing Sun, and Huaizu Jiang. Hoi-diff: Text-driven synthe- sis of 3d human-object interactions using diffusion models. arXiv preprint arXiv:2312.06553, 2023. 3

    2023 arXiv

  21. [29]

    The KIT motion-language dataset

    Matthias Plappert, Christian Mandery, and Tamim Asfour. The KIT motion-language dataset. Big Data, 4(4):236–252,

  22. [30]

    Smcd: High realism motion style transfer via mamba-based diffusion

    Ziyun Qian, Zeyu Xiao, Zhenyi Wu, Dingkang Yang, Mingcheng Li, Shunli Wang, Shuaibing Wang, Dongliang Kou, and Lihua Zhang. Smcd: High realism motion style transfer via mamba-based diffusion. arXiv preprint arXiv:2405.02844, 2024. 1, 3

    2024 arXiv

  23. [31]

    Single motion diffusion

    Sigal Raab, Inbal Leibovitch, Guy Tevet, Moab Arar, Amit H Bermano, and Daniel Cohen-Or. Single motion diffusion. arXiv preprint arXiv:2302.05905, 2023. 3

    2023 arXiv

  24. [32]

    Learn- ing transferable visual models from natural language super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. In Proceedings of the International Conference ...

    2021

  25. [33]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recogni- tion, 2022. 3

    2022

  26. [34]

    Ziplora: Any subject in any style by effectively merging loras

    Viraj Shah, Nataniel Ruiz, Forrester Cole, Erika Lu, Svetlana Lazebnik, Yuanzhen Li, and Varun Jampani. Ziplora: Any subject in any style by effectively merging loras. In Pro- ceedings of the European Conference on Computer Vision , pages 422–438. Springer, 2025. 3

    2025

  27. [35]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 12

    2010 arXiv

  28. [36]

    Arbitrary mo- tion style transfer with multi-condition motion latent diffu- sion model

    Wenfeng Song, Xingliang Jin, Shuai Li, Chenglizhao Chen, Aimin Hao, Xia Hou, Ning Li, and Hong Qin. Arbitrary mo- tion style transfer with multi-condition motion latent diffu- sion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pa...

  29. [37]

    Rsmt: Real-time stylized motion transition for characters

    Xiangjun Tang, Linjun Wu, He Wang, Bo Hu, Xu Gong, Yuchen Liao, Songnan Li, Qilong Kou, and Xiaogang Jin. Rsmt: Real-time stylized motion transition for characters. In ACM SIGGRAPH 2023 Conference Proceedings , pages 1– 10, 2023

    2023

  30. [38]

    Style-erd: Responsive and coherent online motion style transfer

    Tianxin Tao, Xiaohang Zhan, Zhongquan Chen, and Michiel van de Panne. Style-erd: Responsive and coherent online motion style transfer. Arxiv, 2022. 3

    2022

  31. [39]

    Human motion diffusion model

    Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In Proceedings of the International Confer- ence on Learning Representations, 2023. 3

    2023

  32. [40]

    Tlcontrol: Trajectory and language control for human motion synthesis

    Weilin Wan, Zhiyang Dou, Taku Komura, Wenping Wang, Dinesh Jayaraman, and Lingjie Liu. Tlcontrol: Trajectory and language control for human motion synthesis. arXiv preprint arXiv:2311.17135, 2023. 3

    2023 arXiv

  33. [41]

    Styleadapter: A unified stylized image generation model

    Zhouxia Wang, Xintao Wang, Liangbin Xie, Zhongang Qi, Ying Shan, Wenping Wang, and Ping Luo. Styleadapter: A unified stylized image generation model. International Jour- nal of Computer Vision, 2024. 3

    2024

  34. [42]

    Autoregressive stylized motion syn- thesis with generative flow

    Yu-Hui Wen, Zhipeng Yang, Hongbo Fu, Lin Gao, Yanan Sun, and Yong-Jin Liu. Autoregressive stylized motion syn- thesis with generative flow. InProceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition, 2021. 3

    2021

  35. [43]

    Thor: Text to human-object inter- action diffusion via relation intervention

    Qianyang Wu, Ye Shi, Xiaoshui Huang, Jingyi Yu, Lan Xu, and Jingya Wang. Thor: Text to human-object inter- action diffusion via relation intervention. arXiv preprint arXiv:2403.11208, 2024. 3

    2024 arXiv

  36. [44]

    Realtime style transfer for unlabeled heterogeneous human motion

    Shihong Xia, Congyi Wang, Jinxiang Chai, and Jessica Hod- gins. Realtime style transfer for unlabeled heterogeneous human motion. ACM Transactions on Graphics, 34(4):1–10,

  37. [45]

    Omnicontrol: Control any joint at any time for human motion generation

    Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. Omnicontrol: Control any joint at any time for human motion generation. In Proceedings of the Interna- tional Conference on Learning Representations, 2024. 3

    2024

  38. [46]

    Hierarchical style-based networks for motion synthesis

    Jingwei Xu, Huazhe Xu, Bingbing Ni, Xiaokang Yang, Xi- aolong Wang, and Trevor Darrell. Hierarchical style-based networks for motion synthesis. In Proceedings of the Euro- pean Conference on Computer Vision, 2020. 3

    2020

  39. [47]

    Generating human interaction motions in scenes with text control

    Hongwei Yi, Justus Thies, Michael J Black, Xue Bin Peng, and Davis Rempe. Generating human interaction motions in scenes with text control. In European Conference on Com- puter Vision, pages 246–263. Springer, 2025. 3

    2025

  40. [48]

    Freedom: Training-free energy-guided condi- tional diffusion model

    Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang. Freedom: Training-free energy-guided condi- tional diffusion model. In Proceedings of the IEEE/CVF In- ternational Conference on Computer Vision , pages 23174– 23184, 2023. 6

    2023

  41. [49]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the International Conference on Computer Vision, 2023. 1, 3, 4

    2023

  42. [50]

    Smoodi: Stylized motion diffusion model

    Lei Zhong, Yiming Xie, Varun Jampani, Deqing Sun, and Huaizu Jiang. Smoodi: Stylized motion diffusion model. In Proceedings of the European Conference on Computer Vi- sion, 2024. 2, 3, 4, 5, 6, 7, 11, 12, 13 A. Introduction This is the supplementary material, which is divided ...

    2024

  43. [51]

    The details of dataset are shown in Section B

  44. [52]

    The evaluation metrics are shown in Section C

  45. [53]

    The baselines we use in the experiment are shown in Section D

  46. [54]

    We provide the pseudo code, which is presented in Algorithm 1

  47. [55]

    The details of motion style transfer task are elaborated in Section F

  48. [56]

    The inference time cost is provided in Section G

  49. [57]

    We report a user study of our method in Section H

  50. [58]

    More ablation studies are conducted in Section I

  51. [59]

    Runsforward

    More visualization results are presented in Section J, including additional qualitative results and some cases to validate the generalization ability. B. Datasets. We utilize the HumanML3D dataset [11] as the motion con- tent dataset and the 100STYLE dataset [26] as the motion...

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.