Pith. sign in

REVIEW 3 major objections 6 minor 53 references

Behave Your Motion: Habit-preserved Cross-category Animal Motion Transfer

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Animal motion transfer can preserve species-specific habits rather than only re-targeting skeletons, and text descriptions extend the transfer to species whose motion has never been seen.

desk verdict A plausible habit-preserving motion transfer pipeline with a useful new dataset, but the zero-shot unseen-species claim rests on untested text-embedding retrieval. read the letter →

arxiv 2507.07394 v1 pith:Z5HA3O7D submitted 2025-07-10 cs.CV cs.AI

classification cs.CVcs.AI
keywords motiontransferanimalhabitpreservationVQ-VAEzero-shotlargelanguagemodelgenerationquadrupedanimation
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 tries to establish that animal motion can be transferred from one quadruped species to another while keeping the target species' characteristic habits, such as the different joint bending of a moose with long limbs versus a dog with a compact body. It proposes a VQ-VAE-based generator conditioned on a per-category habit latent, learned through a normalizing-flow prior, so that the output action matches the source while the joint articulation matches the target. To reach species never seen in training, the paper supplements the habit latent with text embeddings from a large language model describing each animal's size, speed, limb length, and force patterns, then retrieves the nearest observed habit in that embedding space. The authors introduce a skeleton-bound quadruped dataset and report FID, Intra-FID, 1-NNA, diversity, and joint-error metrics showing substantial gains over direct transfer and style-transfer baselines.

What carries the argument

The central mechanism is the habit latent $z_c$, a per-category random variable modeled with a normalizing-flow prior $p(z)=p_w(w)\,|\det \partial F_\alpha/\partial w|^{-1}$, learned by a separate transformer-based encoder $q_\varphi(z|m_c)$ for each category. The motion decoder is conditioned on $z_c$ along with the LLM text embedding $g_c$, and for unseen categories the habit-retrieval step selects the observed category with minimum Euclidean distance in the text-embedding space and reuses its habit latent as a proxy. The VQ-VAE codebook, trained with exponential moving average and codebook reset, provides the discrete motion representation that lets the source action and target habit be combined at inference.

What would settle it

Hold out entire species during training, run text-only habit retrieval for each held-out species, and compare the transferred motion to ground-truth target motion with FID and 1-NNA; if the scores are not clearly better than retrieving a random observed habit, the zero-shot generalization claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that cross-category animal motion transfer should be formulated as transferring the source action while preserving the target species' habitual behavior, an objective distinct from skeletal retargeting and style transfer. This is achieved by conditioning a vector-quantized motion autoencoder on a category-specific habit latent together with an LLM-derived text embedding: during inference, the source motion is encoded into discrete codes, the target category supplies the habit latent either from a per-category normalizing-flow prior or, for unseen species, via nearest-neighbor retrieval in the text-embedding space, and the decoder produces motion that retains the source action but adopts the target's natural articulation. The authors validate the approach on a new DeformingThings4D-skl dataset and on AnimalML3D, with quantitative metrics favoring the method over direct and style-transfer baselines.

Load-bearing premise

The load-bearing premise is that nearest-neighbor retrieval in the text-embedding space returns an observed species whose habit latent is biologically and behaviorally appropriate for an unseen species, and the paper provides no quantitative held-out test of this premise.

Editorial extensions

If this is right

  • The same source motion will produce different, species-appropriate joint articulation when transferred to different target quadrupeds while keeping the action recognizable.
  • Motion transfer to a species with no motion data becomes possible from a text description alone, via nearest-neighbor habit retrieval.
  • The new DeformingThings4D-skl dataset supplies skeleton-bound quadruped motion with habit descriptions, supporting future motion generation and transfer tasks.
  • Quantitative gains over direct transfer and style transfer on FID, Intra-FID, downstream-task FID, and MPJPE support habit conditioning as an axis separate from style.

Reading between the lines

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

  • The zero-shot claim rests on the geometry of the text-embedding space holding that a nearest observed species is behaviorally similar to the unseen species; this reader would want a held-out-species quantitative test, since the paper only shows qualitative examples.
  • If nearest-neighbor habit retrieval works, the same mechanism could be extended to blend habit latents from multiple neighboring species rather than picking one, potentially producing intermediate gaits.
  • The paper's framing suggests human motion habits (for example, body-type or gender differences) as a natural next domain, but the authors note that current human style datasets encode exaggerated actions rather than subtle habitual cues.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes a VQ-VAE-based framework for cross-category quadruped motion transfer that aims to preserve the target species' habitual motion characteristics. It introduces a category-specific habit encoder modeled with normalizing flows, an LLM-based text encoder that supplies semantic category knowledge, and a habit-retrieval mechanism for transferring motion to previously unobserved species. The authors also introduce DeformingThings4D-skl, a skeletal animal motion dataset with habit descriptions. Reported experiments compare the method with direct transfer and a style-transfer baseline, showing improvements in FID, Intra-FID, downstream transfer FID, diversity, 1-NNA, and MPJPE on both DeformingThings4D-skl and AnimalML3D.

Significance. If validated, the paper would make a useful contribution by defining and attacking a new task: habit-preserved cross-category animal motion transfer. The dataset extension with skeletal bindings and habitual descriptions is a concrete asset for the community. The architectural idea of separating a category-specific habit latent from a generic motion VQ-VAE and bridging to unseen categories through an LLM text embedding is plausible and is backed by substantial quantitative gains on the observed categories. However, the central zero-shot claim currently rests on qualitative examples only, and the evaluation design has gaps (limited baselines, no variance reporting) that prevent the results from fully supporting the claimed superiority.

major comments (3)
  1. [§3.3.2, Eq. (11)] The paper's headline claim of transfer to 'previously unobserved species' rests on BGE nearest-neighbor habit retrieval, but no quantitative held-out evaluation is provided. Tables 1-3 evaluate only the 17 training categories, and Figure 4 is qualitative. This is load-bearing because the retrieval step is the only mechanism that supplies a habit latent for an unseen category, and the assumption that Euclidean proximity in a web-text embedding space implies kinematic similarity is untested. I request a held-out category experiment (e.g., remove several categories during training, compute downstream FID/Intra-FID/MPJPE for transfers into those categories using the Eq. (11) proxy, and compare with an oracle habit latent trained on the held-out category) or a corresponding revision of the claims.
  2. [§4.1.2, Tables 1-3] The reported downstream-task FID is computed against the same target-category data that was used to train the category-specific habit encoder and to derive the target text embedding that conditions generation. This makes the metric, at least in part, a measure of reproduction of the conditioning distribution rather than an independent test of habit preservation. Please clarify whether the metric is computed on held-out motion sequences of the same categories or on the training set, and report the metric separately for source-target category pairs that were not jointly seen during training.
  3. [§4.1.4, Tables 1-2] The quantitative comparison is limited to two baselines (direct transfer and Aberman et al. style transfer) with no error bars, confidence intervals, or significance tests, and no animal-motion-aware baseline is included. Since the paper claims 'substantial improvements' and 'superiority,' the authors should report variance across seeds or bootstrap resamples and add at least one recent quadruped motion-generation/retargeting baseline to demonstrate that the gains are not an artifact of metric choice.
minor comments (6)
  1. [§1, last paragraph] The word 'anction-level' should read 'action-level.'
  2. [§4.2, Tables 1-2] The misspelling 'Abermen' appears in the table text; it should be 'Aberman.'
  3. [§3.2.1 and Eq. (4)] The symbol \hat f is first defined as a sampled codebook index, but the losses \mathcal{L}_{emb} and \mathcal{L}_{com} use \hat f as if it were a codebook vector; please disambiguate the notation, for example by writing e_{\hat f} for the selected code, and make Eq. (1)'s decoder input notation consistent.
  4. [§4.1.1] The description of DeformingThings4D-skl does not specify how the train/validation split is made; please state whether categories overlap between the splits.
  5. [§4.4.2, Table 4] The text reports alpha=0.5 as the best trade-off, but alpha=0.75 achieves a lower FID (0.037 vs. 0.053) with only a slightly higher downstream FID (0.017 vs. 0.012); please clarify the selection criterion used to identify the best hyperparameter.
  6. [§4.1.1] The paper does not state whether DeformingThings4D-skl and the implementation will be released; please include a data/code availability statement.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the core transfer pipeline is not equivalent to its inputs by construction; the zero-shot generalization claim is under-validated but that is a correctness concern, not circularity.

full rationale

The paper's derivation chain does not reduce to its own inputs. The habit latent z_c is learned from target-category motion data, and the downstream FID/Intra-FID metrics compare generated motions to held-out target-category distributions; this is a standard generative-model evaluation, not a fitted parameter renamed as a prediction. The cross-category transfer composes a source motion's content latent with a target category's habit latent during inference, and the training objective reconstructs source motions with their own category, so the transfer evaluation is not statistically forced by the training loss. The only potentially circular-looking element is Eq. 11, where an unseen category's habit latent is set equal to the nearest observed category's latent in BGE text-embedding space. That is an unvalidated assumption about the semantic structure of text embeddings, and the paper provides only qualitative evidence for zero-shot generalization. However, an unsupported assumption is not circularity: Eq. 11 does not redefine the output as the input, nor does it smuggle the conclusion into the premise. There are no load-bearing self-citations; the VQ-VAE backbone and quantization strategies are credited to external prior work without author overlap. The introduction of 'habitual behaviors' is a modeling and evaluation choice rather than a renaming that makes the result true by definition. Overall, the central transfer mechanism has independent empirical content, and the zero-shot claim, while insufficiently evaluated, is not circular.

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

The central claim rests on a few standard generative-modeling axioms and dataset-construction assumptions. The most fragile premise is the LLM text embedding space being a reliable proxy for animal habits, since it enables zero-shot transfer and is not quantitatively validated. The hand-set hyperparameters (alpha, beta, codebook size, temperature) are standard choices, but the temperature value is unreported.

free parameters (4)
  • velocity loss coefficient alpha = 0.5
    Balances position and velocity terms in the reconstruction loss; selected via ablation in Section 4.4.2.
  • commitment loss coefficient beta = 0.02
    Standard VQ-VAE commitment weight, set in Section 3.2.2.
  • codebook size = 512x512
    VQ-VAE codebook dimensions chosen in Section 4.1.3.
  • quantization temperature tau = not reported
    Temperature in Equation 3 for soft code sampling; the paper does not disclose its value.
assumptions (4)
  • domain assumption VQ-VAE discrete codes can represent quadruped motion sequences well enough for reconstruction and transfer.
    The method builds on prior work [44] and assumes the learned codebook captures the motion manifold; no formal guarantee is given.
  • domain assumption A normalizing-flow prior per category can model the distribution of habit latents.
    Used in Section 3.3.1 to define the prior p(z); assumes the flow bijector is flexible enough for the true habit distribution.
  • domain assumption GPT-4o behavioral descriptions and BGE embeddings place animals with similar habits close in text space.
    This is central to Equation 11 habit retrieval; no validation of the embedding structure is provided beyond the qualitative Figure 3.
  • domain assumption The horse-skeleton reference with 21 joints supports faithful rigging of all 21 categories in DeformingThings4D-skl.
    Section 4.1.1; if the skeleton topology does not match the anatomy of some species, the habit latents are computed on a distorted representation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Behave Your Motion: Habit-preserved Cross-category Animal Motion Transfer." pith.science (2026). https://pith.science/paper/Z5HA3O7D

@misc{pith2026250707394,
  author       = {Pith},
  title        = {Pith review of: Behave Your Motion: Habit-preserved Cross-category Animal Motion Transfer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z5HA3O7D}},
  note         = {Machine review of arXiv:2507.07394}
}
read the original abstract

Animal motion embodies species-specific behavioral habits, making the transfer of motion across categories a critical yet complex task for applications in animation and virtual reality. Existing motion transfer methods, primarily focused on human motion, emphasize skeletal alignment (motion retargeting) or stylistic consistency (motion style transfer), often neglecting the preservation of distinct habitual behaviors in animals. To bridge this gap, we propose a novel habit-preserved motion transfer framework for cross-category animal motion. Built upon a generative framework, our model introduces a habit-preservation module with category-specific habit encoder, allowing it to learn motion priors that capture distinctive habitual characteristics. Furthermore, we integrate a large language model (LLM) to facilitate the motion transfer to previously unobserved species. To evaluate the effectiveness of our approach, we introduce the DeformingThings4D-skl dataset, a quadruped dataset with skeletal bindings, and conduct extensive experiments and quantitative analyses, which validate the superiority of our proposed model.

Figures

Figures reproduced from arXiv: 2507.07394 by the authors.

Figure 1
Figure 1. Visual results of motion transfer methods. The same "walking back" motion should appear differently across species: [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The framework of our habit-preserved cross category motion transfer method. Motion is either reconstructed (during [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. 2D Category Motion Information Embeddings. An [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visualization of motion transfer results at the skeleton and mesh levels. The [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 34 canonical work pages

  1. [1]

    Kfir Aberman, Peizhuo Li, Dani Lischinski, Olga Sorkine-Hornung, Daniel Cohen- Or, and Baoquan Chen. 2020. Skeleton-aware networks for deep motion retarget- ing. ACM Transactions on Graphics (TOG) 39, 4 (2020), 62–1

  2. [2]

    Kfir Aberman, Yijia Weng, Dani Lischinski, Daniel Cohen-Or, and Baoquan Chen

  3. [3]

    Tenglong Ao, Zeyi Zhang, and Libin Liu. 2023. Gesturediffuclip: Gesture diffusion model with clip latents. ACM Transactions on Graphics (TOG) 42, 4 (2023), 1–18

  4. [4]

    Ilya Baran and Jovan Popović. 2007. Automatic rigging and animation of 3d characters. ACM Transactions on graphics (TOG) 26, 3 (2007), 72–es

  5. [5]

    Benjamin Biggs, Thomas Roddick, Andrew Fitzgibbon, and Roberto Cipolla. 2019. Creatures great and smal: Recovering the shape and motion of animals from video. In Computer Vision–ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part V 14 . Springer, 3–19

  6. [6]

    Xi Chen, Diederik P Kingma, Tim Salimans, Yan Duan, Prafulla Dhariwal, John Schulman, Ilya Sutskever, and Pieter Abbeel. 2016. Variational lossy autoencoder. arXiv preprint arXiv:1611.02731 (2016)

  7. [7]

    Enric Corona, Albert Pumarola, Guillem Alenya, and Francesc Moreno-Noguer

  8. [8]

    Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. 2023. Mofusion: A framework for denoising-diffusion-based motion synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 9760–9770

Show all 53 references
  1. [9]

    In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Context-aware human motion prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 6992–7001

  2. [10]

    Michael Gleicher. 1998. Retargetting motion to new characters. In Proceedings of the 25th annual conference on Computer graphics and interactive techniques . 33–42

  3. [11]

    Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. 2016. Density estimation using real nvp. arXiv preprint arXiv:1605.08803 (2016)

  4. [12]

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

  5. [13]

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

  6. [14]

    Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. 2023. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 16750–16761

  7. [15]

    Daniel Holden, Jun Saito, and Taku Komura. 2016. A deep learning framework for character motion synthesis and editing. ACM Transactions on Graphics (TOG) 35, 4 (2016), 1–11

  8. [16]

    Boeun Kim, Jungho Kim, Hyung Jin Chang, and Jin Young Choi. 2024. MoST: Motion Style Transformer between Diverse Action Contents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 1705–1714

  9. [17]

    Deok-Kyeong Jang, Soomin Park, and Sung-Hee Lee. 2022. Motion puzzle: Arbi- trary motion style transfer by body part. ACM Transactions on Graphics (TOG) 41, 3 (2022), 1–16

  10. [18]

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

  11. [19]

    Diederik P Kingma. 2013. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)

  12. [20]

    Jehee Lee and Sung Yong Shin. 1999. A hierarchical approach to interactive motion editing for human-like figures. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques . 39–48

  13. [21]

    Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming-Hsuan Yang, and Jan Kautz. 2019. Dancing to music. Advances in neural information processing systems 32 (2019)

  14. [22]

    Shujie Li, Lei Wang, Wei Jia, Yang Zhao, and Liping Zheng. 2022. An iterative solution for improving the generalization ability of unsupervised skeleton motion retargeting. Computers & Graphics 104 (2022), 129–139

  15. [23]

    Buyu Li, Yongchi Zhao, Shi Zhelun, and Lu Sheng. 2022. Danceformer: Music con- ditioned 3d dance generation with parametric motion transformer. InProceedings of the AAAI Conference on Artificial Intelligence , Vol. 36. 1272–1279

  16. [24]

    David Lopez-Paz and Maxime Oquab. 2016. Revisiting classifier two-sample tests. arXiv preprint arXiv:1610.06545 (2016)

  17. [25]

    Yang Li, Hikari Takehara, Takafumi Taketomi, Bo Zheng, and Matthias Nießner

  18. [26]

    OpenAI. 2024. GPT-4o. https://openai.com/index/hello-gpt-4o/ Accessed on October 25, 2024

  19. [27]

    Alec Radford. 2018. Improving language understanding by generative pre- training. (2018)

  20. [28]

    Takeru Miyato and Masanori Koyama. 2018. cGANs with projection discriminator. arXiv preprint arXiv:1802.05637 (2018)

  21. [29]

    Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. 2022. Bailando: 3d dance generation by actor- critic gpt with choreographic memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 11050–11059

  22. [30]

    Wenfeng Song, Xingliang Jin, Shuai Li, Chenglizhao Chen, Aimin Hao, Xia Hou, Ning Li, and Hong Qin. 2024. Arbitrary Motion Style Transfer with Multi- condition Motion Latent Diffusion Model. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition...

  23. [31]

    Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. 2019. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems 32 (2019)

  24. [32]

    Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or

  25. [34]

    Seyoon Tak and Hyeong-Seok Ko. 2005. A physically-based motion retargeting filter. ACM Transactions on Graphics (ToG) 24, 1 (2005), 98–117

  26. [35]

    A Vaswani. 2017. Attention is all you need. Advances in Neural Information Processing Systems (2017)

  27. [36]

    Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang

  28. [37]

    Yu-Hui Wen, Zhipeng Yang, Hongbo Fu, Lin Gao, Yanan Sun, and Yong-Jin Liu. 2021. Autoregressive stylized motion synthesis with generative flow. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 13612–13621

  29. [38]

    Aaron Van Den Oord, Oriol Vinyals, et al. 2017. Neural discrete representation learning. Advances in neural information processing systems 30 (2017)

  30. [39]

    Shitao Xiao, Zheng Liu, Peitian Zhang, and Xingrun Xing. 2023. LM-Cocktail: Re- silient Tuning of Language Models via Model Merging. arXiv:2311.13534 [cs.CL]

  31. [40]

    Zhangsihao Yang, Mingyuan Zhou, Mengyi Shan, Bingbing Wen, Ziwei Xuan, Mitch Hill, Junjie Bai, Guo-Jun Qi, and Yalin Wang. 2024. OmniMotionGPT: Animal Motion Generation with Limited Data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVP...

  32. [41]

    Advances in Neural Information Processing Systems 35 (2022), 14959–14971

    Humanise: Language-conditioned human motion generation in 3d scenes. Advances in Neural Information Processing Systems 35 (2022), 14959–14971

  33. [42]

    Hao Zhang, Di Chang, Fang Li, Mohammad Soleymani, and Narendra Ahuja

  34. [43]

    Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff. 2023. C-Pack: Packaged Resources To Advance General Chinese Embedding. arXiv:2309.07597 [cs.CL]

  35. [44]

    Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. 2023. Generating human motion from textual descriptions with discrete representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recogniti...

  36. [45]

    Peitian Zhang, Shitao Xiao, Zheng Liu, Zhicheng Dou, and Jian-Yun Nie. 2023. Retrieve Anything To Augment Large Language Models. arXiv:2310.07554 [cs.IR]

  37. [46]

    Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. 2023. Generating holistic 3d human motion from speech. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 469–480

  38. [47]

    Zixiang Zhou and Baoyuan Wang. 2023. Ude: A unified driving engine for human motion generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 5632–5641

  39. [48]

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

  40. [49]

    Jiaxu Zhang, Junwu Weng, Di Kang, Fang Zhao, Shaoli Huang, Xuefei Zhe, Linchao Bao, Ying Shan, Jue Wang, and Zhigang Tu. 2023. Skinned motion retargeting with residual perception of motion semantics & geometry. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and...

  41. [52]

    Zeyu Zhang, Yiran Wang, Biao Wu, Shuo Chen, Zhiyuan Zhang, Shiya Huang, Wenbo Zhang, Meng Fang, Ling Chen, and Yang Zhao. 2024. Motion Avatar: Generate Human and Animal Avatars with Arbitrary Motion. arXiv preprint arXiv:2405.11286 (2024)

  42. [2020]

    ACM Transactions on Graphics (TOG) 39, 4 (2020), 64–1

    Unpaired motion style transfer from video to animation. ACM Transactions on Graphics (TOG) 39, 4 (2020), 64–1

  43. [2021]

    In Proceedings of the IEEE/CVF International Conference on Computer Vision

    4DComplete: Non-rigid motion estimation beyond the observable surface. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 12706– 12716

  44. [2022]

    In European Conference on Computer Vision

    Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision . Springer, 358–374

  45. [2024]

    arXiv:2405.14017 [cs.CV]

    MagicPose4D: Crafting Articulated Models with Appearance and Motion Control. arXiv:2405.14017 [cs.CV]

Pith tools

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