Pith. sign in

REVIEW 3 major objections 5 minor 39 references

A frozen body-hand kinematic prior plus lightweight adapters completes hand motion that stays plausible and controllable

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-08 19:23 UTC pith:DG6EA5RW

load-bearing objection Clean prior-then-adapt systems paper for hand completion; useful if the frozen prior still holds body-hand coupling once semantic adapters fire. the 3 major comments →

arxiv 2607.05938 v1 pith:DG6EA5RW submitted 2026-07-07 cs.GR cs.RO

Prior-First, Condition-Second: Scalable and Controllable Hand Motion Completion

Mingyi Shi , Xuelin Chen , Taku Komura This is my paper
classification cs.GR cs.RO
keywords hand motion completionbody-hand kinematic priorautoregressive motion generationlightweight adapterssemantic controlreal-time animationlow-resource learningmotion synthesis
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Synthesizing hand motion that matches full-body dynamics and semantic labels is hard because hands have many degrees of freedom and labeled data are scarce. This paper argues that the right order of operations is prior-first, condition-second: first learn a generic, streaming autoregressive body-hand kinematic prior from large unlabeled motion, capturing the intrinsic coordination between body and hands; then freeze that prior and inject control only through lightweight, semantically layered adapters. The adapters place conditioning at the right kinematic levels so that self-supervised attributes or a few hours of text labels can steer the hands without forcing the model to relearn kinematics. If the claim holds, animators and interactive systems get real-time, kinematically coherent hand completion that stays robust in low-resource and cross-dataset settings and supports interactive authoring—without the brittleness of end-to-end conditioned models trained on limited labeled data.

Core claim

A streaming autoregressive body-hand kinematic prior learned from large unlabeled motion, kept frozen, plus lightweight semantically layered adapters that inject control at appropriate kinematic levels, yields more plausible, robust, and controllable body-conditioned hand motion completion than end-to-end conditioned baselines, especially under limited labels and across datasets, while remaining real-time.

What carries the argument

The prior-first, condition-second pipeline: a frozen streaming autoregressive body-hand kinematic prior that maintains mechanical body-hand coupling, with semantically layered adapters that inject attribute or text control only at the right kinematic levels without relearning the prior.

Load-bearing premise

That freezing a generic body-hand kinematic prior and injecting control only through lightweight adapters at chosen kinematic levels is enough for semantic control, without relearning kinematics or breaking body-hand coupling when labeled data are scarce.

What would settle it

Train the same prior end-to-end with conditioning instead of freezing it and adding adapters; if kinematic error, cross-dataset robustness, and low-label text control match or beat the frozen-plus-adapters setup on the paper's benchmarks, the prior-first claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Hand completion stays kinematically coherent even when body motion is novel or out-of-distribution relative to labeled sets.
  • A few hours of text labels suffice for weakly supervised text-driven hand control without full joint retraining.
  • Real-time streaming inference supports interactive authoring in production animation pipelines.
  • Self-supervised attribute control can be layered at different kinematic levels without destroying body-hand coupling.
  • Cross-dataset transfer improves because the bulk of kinematic structure is learned once from unlabeled motion.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same prior-first pattern may transfer to other high-DoF, sparsely labeled body parts (face, feet) where kinematics dominate and semantics are scarce.
  • Interactive authoring could treat the adapters as live handles, letting artists dial attributes while the frozen prior keeps motion legal.
  • If unlabeled body-hand corpora grow, the prior could absorb new styles without touching existing control adapters.
  • Weak supervision budgets might shrink further if self-supervised attributes already cover most of the semantic axes needed for text control.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The manuscript proposes a prior-first, condition-second framework for body-conditioned hand motion completion. A streaming autoregressive body-hand kinematic prior is first trained on large-scale unlabeled motion to capture intrinsic body-hand coordination and mechanical coupling. Semantic control is then added via lightweight, semantically-layered adapters on the frozen prior, supporting self-supervised attribute control and weakly supervised text-driven control with only a few hours of labeled data. The authors claim improved kinematic plausibility, robustness, and controllability relative to end-to-end conditioned baselines, especially in low-resource and cross-dataset regimes, together with real-time inference and an interactive authoring workflow for production animation.

Significance. If the empirical claims hold, the work offers a modular and practical alternative to end-to-end conditioned hand motion models. Decoupling a reusable kinematic prior from task-specific control interfaces is valuable for animation pipelines where labeled semantic data are scarce and streaming real-time generation is required. Credit is due for the streaming autoregressive prior design, the explicit focus on low-supervision controllability, the cross-dataset robustness emphasis, and the interactive authoring demonstration. These elements, if validated, constitute a useful design pattern for controllable body-hand synthesis under limited labels.

major comments (3)
  1. [Methods (prior + semantically-layered adapters); abstract claims on coupling] The central claim that freezing the body-hand kinematic prior and injecting control only via lightweight semantically-layered adapters preserves mechanical body-hand coupling is load-bearing for the reported gains in kinematic plausibility and robustness versus end-to-end baselines. The manuscript should provide direct evidence—e.g., body-hand joint correlation or relative-pose consistency metrics with vs. without adapters, under attribute and text control, and in cross-dataset settings—that higher-layer adapters do not dilute or override the prior’s learned coordination. Without such measurements, the advantage of the frozen-prior design remains incompletely substantiated relative to the weakest design assumption.
  2. [Experiments / low-resource and weakly supervised text control] The low-resource and “few hours of labeled data” claims are central to the condition-second contribution. Please report exact labeled data volumes (hours/frames) for each adapter setting, the corresponding training budgets and data scales for end-to-end baselines, and a controlled ablation that varies labeled-data scale. Without matched budgets and a scale curve, the claimed efficiency and robustness advantages over end-to-end conditioning cannot be fairly assessed.
  3. [Evaluation metrics and main result tables] Kinematic plausibility is multi-faceted. Clarify which metrics (hand-body relative pose error, joint-limit violations, interpenetration, temporal jerk/acceleration consistency, or analogous coordination scores) drive the “improved kinematic plausibility” claim, and whether they are evaluated under free semantic control as well as pure body-conditioned completion. If primary tables emphasize reconstruction-style errors under body conditioning alone, they do not fully address coupling preservation when adapters inject strong semantic signals.
minor comments (5)
  1. [Adapter architecture subsection] Define adapter placement and “kinematic levels” more precisely (which layers/joints receive which semantic signals) and state adapter rank/capacity relative to the frozen prior so that “lightweight” is reproducible.
  2. [Related Work] Related work should more explicitly position the design against frozen-backbone + adapter literature in motion and vision/language, and against prior body-hand or hand-only completion models, to clarify what is new beyond the prior-first framing.
  3. [Real-time inference / authoring workflow] For the interactive authoring and real-time claims, report latency, streaming context length, and hardware so that production applicability can be judged independently of qualitative demos.
  4. [Baselines and experimental setup] Ensure all baselines are described with architecture, training data, and conditioning interface parity; any mismatch should be stated explicitly in the experimental protocol.
  5. [Throughout / figures] Minor presentation: expand acronyms on first use, check figure axis labels and caption self-containment, and verify that the homepage link and any supplementary videos are stable for review.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a careful and constructive review. The three major comments correctly identify load-bearing claims of the prior-first, condition-second design: (i) that freezing the body–hand kinematic prior and injecting control only via lightweight semantically-layered adapters preserves mechanical coupling; (ii) that the low-resource / few-hours labeled-data claims require matched budgets and a scale curve; and (iii) that “improved kinematic plausibility” must be tied to explicit coordination metrics under free semantic control, not only reconstruction under body conditioning. We address each point below with concrete revisions: new coupling-preservation metrics (with vs. without adapters, under attribute and text control, and cross-dataset), exact labeled volumes and matched end-to-end budgets plus a labeled-data scale ablation, and a clarified metric suite with primary-table updates that evaluate coordination under free semantic control. We believe these changes fully substantiate the design claims and make the empirical case fair and transparent.

read point-by-point responses

  1. Referee: The central claim that freezing the body-hand kinematic prior and injecting control only via lightweight semantically-layered adapters preserves mechanical body-hand coupling is load-bearing. Provide direct evidence—e.g., body-hand joint correlation or relative-pose consistency metrics with vs. without adapters, under attribute and text control, and in cross-dataset settings—that higher-layer adapters do not dilute or override the prior’s learned coordination.

    Authors: We agree this is load-bearing and that the original manuscript under-specified direct coupling evidence. We will add a dedicated “Coupling Preservation” analysis. Concretely: (1) body–hand joint correlation matrices and Pearson/Spearman scores between body root/upper-limb velocities and hand joint angles, comparing frozen prior alone vs. prior+attribute adapters vs. prior+text adapters; (2) relative-pose consistency (wrist-to-body and finger-to-wrist error under body-driven kinematics) with and without adapters; (3) the same suite under free attribute and free text control (not only body-conditioned completion); and (4) cross-dataset transfer of these scores. We will also report a “coupling dilution” delta (adapter-on minus prior-only) to show higher-layer adapters do not override the prior. These measurements will appear in a new table/figure and be referenced from the abstract and claims. This directly substantiates the frozen-prior advantage over end-to-end baselines. revision: yes

  2. Referee: The low-resource and “few hours of labeled data” claims are central. Report exact labeled data volumes (hours/frames) for each adapter setting, the corresponding training budgets and data scales for end-to-end baselines, and a controlled ablation that varies labeled-data scale. Without matched budgets and a scale curve, efficiency and robustness advantages cannot be fairly assessed.

    Authors: We agree the original wording was insufficiently precise. We will report exact labeled volumes (hours and frames) for every adapter setting (self-supervised attributes and weakly supervised text), and the matched training budgets (data scale, steps, compute) for all end-to-end conditioned baselines. We will add a controlled labeled-data scale ablation (e.g., 0.5 h / 1 h / few hours / full available labels) for both our adapters and the end-to-end baselines under identical optimization budgets, with a scale curve for kinematic and control metrics. This makes the efficiency and robustness claims assessable on equal footing and will replace the informal “few hours” phrasing with quantitative statements in abstract, method, and experiments. revision: yes

  3. Referee: Kinematic plausibility is multi-faceted. Clarify which metrics (hand-body relative pose error, joint-limit violations, interpenetration, temporal jerk/acceleration consistency, or analogous coordination scores) drive the “improved kinematic plausibility” claim, and whether they are evaluated under free semantic control as well as pure body-conditioned completion. If primary tables emphasize reconstruction-style errors under body conditioning alone, they do not fully address coupling preservation when adapters inject strong semantic signals.

    Authors: We agree. We will explicitly define the kinematic-plausibility suite: hand–body relative pose error, joint-limit violation rate, self-/body–hand interpenetration, temporal jerk and acceleration consistency, and the new body–hand coordination/correlation scores from Comment 1. We will state which of these drive the main claim and report them under (a) pure body-conditioned completion and (b) free attribute and free text control. Primary result tables will be revised so that reconstruction-style errors under body conditioning alone are no longer the sole support for the plausibility claim; coordination and constraint metrics under free semantic control will be first-class. The abstract and experimental narrative will be updated accordingly so that “improved kinematic plausibility” is unambiguously tied to these measurements. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical prior-then-adapter ML system evaluated against external baselines

full rationale

This is a standard empirical graphics/ML systems paper. The body-hand kinematic prior is trained on large-scale unlabeled motion; lightweight semantically-layered adapters are then trained on limited labeled or self-supervised attributes while the prior is frozen. Reported gains in kinematic plausibility, robustness, controllability, low-resource, and cross-dataset settings are measured by comparison to end-to-end conditioned baselines on held-out data and external datasets, not by defining the target quantities via fitted constants or by renaming a known empirical pattern. There is no self-definitional loop (X defined in terms of Y then claimed to derive Y), no fitted parameter re-labeled as an independent prediction, no load-bearing uniqueness theorem imported from the same authors, and no ansatz smuggled in via self-citation that forces the central claim. Residual methodological risks (whether adapters preserve body-hand coupling under semantic control, possible train/eval leakage) are correctness or generalization concerns, not circularity of the derivation chain. The framework is self-contained against external benchmarks; score 0 is the honest finding.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 0 invented entities

Central claim rests on standard ML training assumptions, the sufficiency of a frozen kinematic prior plus light adapters for semantic control, and empirical evaluation protocols. No new physical entities. Free parameters are the usual model and training hyperparameters (not enumerated in the abstract). Domain assumptions include that unlabeled large-scale motion captures transferable body-hand coupling and that limited labeled data suffices for adapter-based control.

free parameters (1)
  • model/training hyperparameters (architecture, losses, adapter ranks/placement, data scales)
    Abstract does not list fitted values; any deep motion model depends on many tuned hyperparameters that affect reported gains.
axioms (4)
  • domain assumption Large-scale unlabeled body-hand motion encodes a transferable kinematic prior sufficient for coherent real-time hand completion from body dynamics.
    Core premise of the prior-first stage; if coupling is dataset-specific or underdetermined, the frozen prior fails.
  • ad hoc to paper Lightweight adapters on a frozen prior can inject semantic control without relearning kinematic structure.
    Design choice that separates the paper from end-to-end baselines; load-bearing for low-resource controllability claims.
  • domain assumption Self-supervised attribute signals and a few hours of text labels are adequate supervision for controllable adapters.
    Stated low-supervision regime; if labels are too weak or misaligned, controllability claims fail.
  • standard math Standard autoregressive/streaming motion modeling and kinematic coupling constraints are valid for real-time generation.
    Background methods assumed rather than re-derived.

pith-pipeline@v0.9.1-grok · 6355 in / 2159 out tokens · 27115 ms · 2026-07-08T19:23:54.331084+00:00 · methodology

0 comments
read the original abstract

Synthesizing hand motion that matches the full body motion and the semantic labels is a difficult task due to their high degrees of freedom and the lack of semantic labels. To cope with this issue, we propose a prior-first, condition-second framework for body-conditioned hand motion completion. Our framework first learns a generic body-hand kinematic prior from large-scale unstructured and unlabeled motion data, capturing the intrinsic coordination between global body dynamics and hand articulation. Semantic control is then introduced through lightweight adaptation on top of the frozen prior, avoiding the need to relearn kinematic structure for each control interface. Our framework centers on a streaming, autoregressive body-hand prior that generates coherent, kinematically consistent hand motion from body dynamics in real time, using structured kinematic modeling to maintain mechanical body-hand coupling. To enable practical controllability under limited supervision, we introduce semantically-layered adapters that inject conditioning signals at appropriate kinematic levels, supporting both self-supervised attribute control and weakly supervised text-driven control with only a few hours of labeled data. Extensive evaluations demonstrate that our framework improves kinematic plausibility, robustness, and controllability compared to end-to-end conditioned baselines, particularly in low-resource and cross-dataset settings. We further showcase real-time inference and an interactive authoring workflow, highlighting the applicability to production animation pipelines. Homepage: https://AIGAnimation.github.io/HandPrior/

Figures

Figures reproduced from arXiv: 2607.05938 by Mingyi Shi, Taku Komura, Xuelin Chen.

Figure 1
Figure 1. Figure 1: Our method robustly completes high-fidelity hand motion conditioned on global body dynamics and optional semantic controls. Instead of training a fully end-to-end conditioned generator, we constrain the outputs to a physically plausible kinematic manifold, resulting in coherent articulation and strong robustness across diverse inputs. Abstract Synthesizing hand motion that matches the full body motion and … view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our diffusion block in the autoregressive prior. We project the body motion window and a 10-frame history buffer (P=10) into condition tokens, and the noisy hand motion into latent tokens. The sequence is passed through the Transformer backbone to generate an output window of N frames (the prediction window of length L in the text). Tokens along the root-to-wrist chain are aggregated via kinema… view at source ↗
Figure 3
Figure 3. Figure 3: Same body motion input can generate distinct yet plau￾sible hand motion by given different random seeds. labeled data. This confirms that the adapter’s role is strictly lim￾ited to manifold navigation (selecting valid trajectories) rather than manifold construction. In the conditional stage, our adapters steer this diversity toward the user-specified control signal by aligning the generated sample with the… view at source ↗
Figure 4
Figure 4. Figure 4: Runtime screenshot of the Blender add-on. Left to right: our generation, identity hands, original input. Metrics. We propose a comprehensive evaluation protocol cov￾ering three dimensions: (1) Statistical metrics: FID (Fréchet In￾ception Distance, computed in a learned hand-motion embedding space via a pretrained autoencoder) to measure the distributional distance to the ground truth, Diversity to measure … view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of our prior with baselines on the test split of Dataset A. Our method produces stable wrist and plausible finger articulation, whereas BOTH2Hands exhibits unstable wrist behavior and MDM occasionally generates unnatural joint rotations [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Attribute-driven control: varying the fist tightness level while keeping body motion fixed. Text: Two hands are flatten Text: Two hands are dribbling [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 9
Figure 9. Figure 9: Data Efficiency for Adapter. We compare different scales of labeled data from Dataset B (horizontal axis in minutes), and report MPJPE and Text–Motion Consistency after the injection. Module Effectiveness. We ablate KCCA and semantically￾layered injection ( [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

39 extracted references · 39 canonical work pages · 4 internal anchors

  1. [1]

    https://cascadeur.com/, 2026

    Cascadeur : Cascadeur. https://cascadeur.com/, 2026. Accessed: 2026-01-20

    work page 2026

  2. [2]

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

    Christen S., Hampali S., Sener F., Remelli E., Hodan T., Sauser E., Ma S., Tekin B. : Diffh2o: Diffusion-based synthesis of hand-object interactions from textual descriptions. In SIGGRAPH Asia 2024 Conference Papers (2024), pp. 1--11

    work page 2024

  3. [3]

    2024 , publisher =

    Chen R., Shi M., Huang S., Tan P., Komura T., Chen X. : Taming diffusion probabilistic models for character control. In ACM SIGGRAPH 2024 Conference Papers (New York, NY, USA, 2024), SIGGRAPH '24, Association for Computing Machinery. URL: https://doi.org/10.1145/3641519.3657440, https://doi.org/10.1145/3641519.3657440 doi:10.1145/3641519.3657440

    work page doi:10.1145/3641519.3657440 2024

  4. [4]

    J., Taheri O

    Duran E., Athanasiou N., Kocabas M., Black M. J., Taheri O. : Fusion: Full-body unified motion prior for body and hands via diffusion. arXiv preprint arXiv:2601.03959 (2026)

    work page arXiv 2026

  5. [5]

    : Expressive forecasting of 3d whole-body human motions

    Ding P., Cui Q., Wang H., Zhang M., Liu M., Wang D. : Expressive forecasting of 3d whole-body human motions. In Proceedings of the AAAI Conference on Artificial Intelligence (2024), vol. 38, pp. 1537--1545

    work page 2024

  6. [6]

    PianoMotion10M: Dataset and Benchmark for Hand Motion Generation in Piano Performance

    Gan Q., Wang S., Wu S., Zhu J. : Pianomotion10m: Dataset and benchmark for hand motion generation in piano performance. arXiv preprint arXiv:2406.09326 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  7. [7]

    E., Alexanderson S., Beskow J

    Henter G. E., Alexanderson S., Beskow J. : Moglow: probabilistic and controllable motion synthesis using normalising flows. ACM Transactions on Graphics 39, 6 (Nov. 2020), 1–14. URL: http://dx.doi.org/10.1145/3414685.3417836, https://doi.org/10.1145/3414685.3417836 doi:10.1145/3414685.3417836

    work page doi:10.1145/3414685.3417836 2020

  8. [8]

    : Interact: A large-scale dataset of dynamic, expressive and interactive activities between two people in daily scenarios

    Ho L., Huang Y., Qin D., Shi M., Tse W., Liu W., Yamagishi J., Komura T. : Interact: A large-scale dataset of dynamic, expressive and interactive activities between two people in daily scenarios. Proceedings of the ACM on Computer Graphics and Interactive Techniques 8, 4 (2025), 1--27

    work page 2025

  9. [9]

    The Invisible EgoHand: 3D Hand Forecasting through EgoBody Pose Estimation

    Hatano M., Zhu Z., Saito H., Damen D. : The invisible egohand: 3d hand forecasting through egobody pose estimation. arXiv preprint arXiv:2504.08654 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  10. [10]

    : Data-driven finger motion synthesis for gesturing characters

    J \"o rg S., Hodgins J., Safonova A. : Data-driven finger motion synthesis for gesturing characters. ACM Transactions on Graphics (TOG) 31, 6 (2012), 1--7

    work page 2012

  11. [11]

    G., Pai D

    Kry P. G., Pai D. K. : Interaction capture and synthesis. ACM Transactions on Graphics (TOG) 25, 3 (2006), 872--880

    work page 2006

  12. [12]

    Liu C. K. : Synthesis of interactive hand manipulation. In Proceedings of the 2008 ACM SIGGRAPH/Eurographics Symposium on Computer Animation (2008), pp. 163--171

    work page 2008

  13. [13]

    Liu C. K. : Dextrous manipulation from a grasping pose. In ACM SIGGRAPH 2009 papers. 2009, pp. 1--6

    work page 2009

  14. [14]

    Y., Zinno F., Cheng G., van de Panne M

    Ling H. Y., Zinno F., Cheng G., van de Panne M. : Character controllers using motion vaes. ACM Trans. Graph. 39, 4 (2020)

    work page 2020

  15. [15]

    : Beat: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis

    Liu H., Zhu Z., Iwamoto N., Peng Y., Li Z., Zhou Y., Bozkurt E., Zheng B. : Beat: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. In European conference on computer vision (2022), Springer, pp. 612--630

    work page 2022

  16. [16]

    : Automatic splicing for hand and body animations

    Majkowska A., Zordan V., Faloutsos P. : Automatic splicing for hand and body animations. In ACM SIGGRAPH 2006 Sketches. 2006, pp. 32--es

    work page 2006

  17. [17]

    : Body2hands: Learning to infer 3d hands from conversational gesture body dynamics

    Ng E., Ginosar S., Darrell T., Joo H. : Body2hands: Learning to infer 3d hands from conversational gesture body dynamics. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2021), pp. 11865--11874

    work page 2021

  18. [18]

    S., Zordan V

    Pollard N. S., Zordan V. B. : Physically based grasping control from example. In Proceedings of the 2005 ACM SIGGRAPH/Eurographics symposium on Computer animation (2005), pp. 311--318

    work page 2005

  19. [19]

    : Diverse 3d hand gesture prediction from body dynamics by bilateral hand disentanglement

    Qi X., Liu C., Sun M., Li L., Fan C., Yu X. : Diverse 3d hand gesture prediction from body dynamics by bilateral hand disentanglement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2023), pp. 4616--4626

    work page 2023

  20. [20]

    Rempe D., Birdal T., Hertzmann A., Yang J., Sridhar S., Guibas L. J. : Humor: 3d human motion model for robust pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision (2021), pp. 11488--11499

    work page 2021

  21. [21]

    H., Cohen-Or D

    Raab S., Gat I., Sala N., Tevet G., Shalev-Arkushin R., Fried O., Bermano A. H., Cohen-Or D. : Monkey see, monkey do: Harnessing self-attention in motion diffusion for zero-shot motion transfer. In SIGGRAPH Asia 2024 Conference Papers (2024), pp. 1--13

    work page 2024

  22. [22]

    Sawdayee H., Guo C., Tevet G., Zhou B., Wang J., Bermano A. H. : Dance like a chicken: Low-rank stylization for human motion diffusion. arXiv preprint arXiv:2503.19557 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  23. [23]

    MotionPersona: Characteristics-aware Locomotion Control

    Shi M., Liu W., Mei J., Tse W., Chen R., Chen X., Komura T. : Motionpersona: Characteristics-aware locomotion control. arXiv preprint arXiv:2506.00173 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  24. [24]

    Shafir Y., Tevet G., Kapon R., Bermano A. H. : Human motion diffusion as a generative prior. In The Twelfth International Conference on Learning Representations (2024)

    work page 2024

  25. [25]

    Shi Y., Wang J., Jiang X., Lin B., Dai B., Peng X. B. : Interactive character control with auto-regressive motion diffusion models, July 2024. URL: https://doi.org/10.1145/3658140, https://doi.org/10.1145/3658140 doi:10.1145/3658140

    work page doi:10.1145/3658140 2024

  26. [26]

    E., Sarafianos N., Tung T., Pons-Moll G

    Tiwari G., Anti \'c D., Lenssen J. E., Sarafianos N., Tung T., Pons-Moll G. : Pose-ndf: Modeling human pose manifolds with neural distance fields. In European Conference on Computer Vision (2022), Springer, pp. 572--589

    work page 2022

  27. [27]

    J., Tzionas D

    Taheri O., Choutas V., Black M. J., Tzionas D. : Goal: Generating 4d whole-body motion for hand-object grasping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022), pp. 13263--13273

    work page 2022

  28. [28]

    Team T. H. D. D. H. : Hy-motion 1.0: Scaling flow matching models for text-to-motion generation. arXiv preprint arXiv:2512.23464 (2025)

    work page arXiv 2025

  29. [29]

    J., Tzionas D

    Taheri O., Ghorbani N., Black M. J., Tzionas D. : Grab: A dataset of whole-body human grasping of objects. In European conference on computer vision (2020), Springer, pp. 581--600

    work page 2020

  30. [30]

    Tevet G., Raab S., Gordon B., Shafir Y., Cohen-Or D., Bermano A. H. : Human motion diffusion model

    work page

  31. [31]

    : Prompting future driven diffusion model for hand motion prediction

    Tang B., Zhang K., Luo W., Liu W., Li H. : Prompting future driven diffusion model for hand motion prediction. In European Conference on Computer Vision (2024), Springer, pp. 169--186

    work page 2024

  32. [32]

    : Marker-removal networks to collect precise 3d hand data for rgb-based estimation and its application in piano

    Wu E., Nishioka H., Furuya S., Koike H. : Marker-removal networks to collect precise 3d hand data for rgb-based estimation and its application in piano. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (2023), pp. 2977--2986

    work page 2023

  33. [33]

    Wang R., Xu P., Shi H., Schumann E., Liu C. K. : F \"u relise: Capturing and physically synthesizing hand motion of piano performance. In SIGGRAPH Asia 2024 Conference Papers (2024), pp. 1--11

    work page 2024

  34. [34]

    Ye Y., Liu C. K. : Synthesis of detailed hand manipulations using contact sampling. ACM Transactions on Graphics (ToG) 31, 4 (2012), 1--10

    work page 2012

  35. [35]

    : On the continuity of rotation representations in neural networks

    Zhou Y., Barnes C., Jingwan L., Jimei Y., Hao L. : On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (June 2019), CVPR'19

    work page 2019

  36. [36]

    L., Lenssen J

    Zhou K., Bhatnagar B. L., Lenssen J. E., Pons-Moll G. : Toch: Spatio-temporal object-to-hand correspondence for motion refinement. In European Conference on Computer Vision (2022), Springer, pp. 1--19

    work page 2022

  37. [37]

    : Both2hands: Inferring 3d hands from both text prompts and body dynamics

    Zhang W., Huang M., Zhou Y., Zhang J., Yu J., Wang J., Xu L. : Both2hands: Inferring 3d hands from both text prompts and body dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024), pp. 2393--2404

    work page 2024

  38. [38]

    : Manipnet: neural manipulation synthesis with a hand-object spatial representation

    Zhang H., Ye Y., Shiratori T., Komura T. : Manipnet: neural manipulation synthesis with a hand-object spatial representation. ACM Transactions on Graphics (ToG) 40, 4 (2021), 1--14

    work page 2021

  39. [39]

    : Cams: Canonicalized manipulation spaces for category-level functional hand-object manipulation synthesis

    Zheng J., Zheng Q., Fang L., Liu Y., Yi L. : Cams: Canonicalized manipulation spaces for category-level functional hand-object manipulation synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023), pp. 585--594

    work page 2023