Pith. sign in

REVIEW 2 major objections 5 minor 68 references

HandFlow: Fully Generative 4D Hand Recovery with Flow Matching

T0 review · 2 major / 5 minor · reviewed 2026-07-14 · grok-4.5

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

Pith's one-line read HandFlow recovers smooth, accurate 4D hand motion from monocular video by generating full MANO sequences with flow matching instead of regressing frame by frame.

desk verdict Solid methods paper: full-window rectified-flow over raw MANO with cmask and velocity blending delivers real world-space and speed gains; soft spots are genre-standard, not load-bearing. read the letter →

arxiv 2607.11221 v1 pith:4TD2QICW submitted 2026-07-13 cs.CV cs.AI

Mingxi Xu , Bowen Duan , Yi Gu , Zhengyang Shen , Renjing Xu , Yutao Yue This is my paper
classification cs.CVcs.AI
keywords 4DhandreconstructionflowmatchingMANOmonocularvideomotioncapturegenerativemodelstemporalcoherenceocclusion
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

Monocular hand video is full of occlusion, blur, and viewpoint change, so per-frame estimators jitter and deterministic temporal models often invent a single compromise pose. HandFlow treats reconstruction as conditional generation in MANO parameter space: it learns a prior over plausible hand-motion sequences and recovers one coherent trajectory by integrating a rectified-flow ODE once over an entire temporal window. A dual-stream transformer jointly attends over hand parameters and visual-skeletal conditions across that window, while confidence-aware continuous masking blends unreliable detections with learnable mask tokens so the motion prior can complete missing evidence. On DexYCB and HOT3D the method cuts world-space pose error by more than 30% versus the strongest baseline, records the lowest acceleration error among all evaluated methods, and reconstructs 150-frame sequences at about 47 fps—roughly twelve times faster than the fastest prior video method—while staying competitive on per-frame accuracy. A sympathetic reader cares because AR/VR, robotics, and interaction systems need both geometric fidelity and temporal coherence under exactly the ambiguous conditions where pure regressors fail.

What carries the argument

Rectified-flow generation of a packed MANO window, driven by a Flux-style dual-stream transformer and confidence-aware continuous masking (cmask) that interpolates observed visual and skeletal tokens with learnable mask tokens according to per-frame detection confidence.

What would settle it

On clips where detection confidence is systematically miscalibrated—hands fully out of view scored as merely occluded, or high-confidence detections that are actually wrong—if world-space and acceleration errors rise to the level of strong deterministic video baselines, the claim that generative cmask recovery drives the gains would be falsified.

Watch

Extended reading notes

Core claim

Casting monocular 4D hand reconstruction as fully generative rectified-flow matching over whole windows of MANO parameters yields trajectories that are substantially more accurate in world space and smoother over time than deterministic video regressors, while remaining competitive per frame and an order of magnitude faster at reconstruction because each window is denoised in a single short ODE integration.

Load-bearing premise

The method relies on a frozen detector’s per-frame confidence being a trustworthy continuous signal for how much to trust visual and skeletal observations versus a learned motion prior.

Editorial extensions

If this is right

  • World-space hand trajectories improve when temporal coherence is learned as a generative prior rather than enforced by deterministic aggregation.
  • Reconstruction can become a small fraction of end-to-end latency, so future SLAM speedups translate almost directly into faster full pipelines.
  • Under heavy occlusion, generative completion can keep pose error near the low-millimeter range where single-frame regressors degrade sharply.
  • Velocity-blended overlapping windows extend a single continuous trajectory to long videos without boundary jumps from independent window ODEs.

Reading between the lines

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

  • The same single-pass flow-matching window design is a natural candidate for other articulated monocular problems—full body, face-plus-hands—where observations are intermittent and multimodal.
  • Because a generic, much smaller frontend still beats several specialized video baselines, lighter detectors may be enough once a strong motion prior is in place.
  • Confidence-weighted mask tokens offer a reusable pattern for any conditional motion model that must tolerate noisy or missing detector streams.
  • As camera-pose estimation improves, generative window methods could push world-space hand tracking toward interactive rates on a single GPU.
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

2 major / 5 minor

Summary. HandFlow formulates monocular 4D hand reconstruction as conditional generation of MANO parameter sequences via rectified flow. Given frozen HaMeR visual and skeletal features, a Flux-style dual-stream transformer denoises an entire temporal window of shape, pose, and translation through a few-step ODE; confidence-aware continuous masking (cmask) interpolates unreliable observations with learnable mask tokens, and velocity-blended overlapping windows extend inference to long videos. On DexYCB (s0) and a held-out HOT3D split the method reports SOTA world-space accuracy (WA-MPJPE 16.17 mm vs. 25.24 UniHand / 27.37 HaWoR), lowest acceleration error (4.18 m/s²), competitive per-frame pose, and ~12× faster reconstruction than prior video methods (~47 fps on 150 frames), with ablations on conditioning, masking, architecture, overlap, ODE steps, noise, and frontend generality.

Significance. If the reported gains hold under the stated protocol, the paper is a clear empirical advance for monocular 4D hand recovery: generative sequence modeling with few-step rectified flow delivers large world-space and smoothness improvements while inverting the usual latency bottleneck (reconstruction becomes a small fraction of end-to-end time). Strengths include multi-metric evaluation (camera/world, occlusion bins, Accel, RTE), systematic ablations (Tables 4, 7; Appendices C–F), noise-sensitivity and ODE-step sweeps, a frontend-generality check with DINOv3+MediaPipe, and public code/models. The dual-stream design, cmask, and velocity blending are concrete, reusable engineering contributions within the MANO+SLAM evaluation regime standard to the field.

major comments (2)
  1. §4.1 / Table 2: UniHand is the closest generative baseline and the second-best world-space method, yet its HOT3D numbers are quoted from the original paper while its code and the paper’s custom HOT3D val split are undocumented. Because HandFlow’s strongest claim is a >30% WA-MPJPE reduction versus this baseline (16.17 vs 25.24), the comparison is not fully reproducible under a shared split and pipeline. Either re-evaluate UniHand on the released split with a public implementation, or demote the UniHand comparison to secondary status and anchor the SOTA claim primarily against re-run baselines (HaWoR, Dyn-HaMR, WiLoR+SLAM).
  2. §3.2 Eq. (5) and Limitations G: cmask treats HaMeR confidence as a continuous reliability signal that confounds partial occlusion with full FOV exit. Ablations (Table 4) show that disabling confidence weighting inflates MPJPE by ~86% while PA-MPJPE is nearly unchanged, so the mechanism is load-bearing for global translation. The paper should quantify FOV-exit vs occlusion failure rates on HOT3D/DexYCB (or add an explicit out-of-view flag) so that the generative-completion claim is scoped to cases where some visual evidence remains.
minor comments (5)
  1. §4.5 / Table 4: Absolute metrics on the merged DexYCB+HOT3D set are not comparable to Tables 1–2; state this more prominently in the table caption and main text to avoid misreading.
  2. Appendix B: Auxiliary loss weights (L_β=1.0, L_vel/acc/reproj=0.5, L_j3d=0.05) and the ¯t scaling are given without sensitivity analysis; a short note on whether results are stable under modest reweighting would help.
  3. Fig. 1 and Fig. 6: Color legend (green GT, blue HandFlow, orange HaWoR, red HaMeR) is clear in captions but should also appear once in the main figure panels for print readability.
  4. §3.1: The choice of 3 Euler steps is justified by Appendix C; a one-sentence pointer in the main text to the plateau at steps ≥2 would help readers who skip the appendix.
  5. Related Work: DuoMo and FMPose3D are appropriately distinguished; a brief note on why latent VAE compression (UniHand) is avoided beyond “preserving capacity” would clarify the design trade-off.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical SOTA methods paper with held-out benchmarks and standard losses, not definitional predictions.

full rationale

HandFlow is a standard empirical computer-vision methods contribution. It formulates monocular 4D hand recovery as conditional rectified-flow generation of MANO parameters (Eqs. 1–3), conditions on frozen HaMeR visual/skeletal features with continuous masking (Eq. 5), and trains with a composite flow-matching + geometric/temporal loss. All reported numbers (RA-MPJPE, WA-MPJPE, Accel, runtime) are measured on held-out DexYCB s0 and HOT3D val splits against re-run or quoted external baselines; the metrics are not algebraic rearrangements of the training objective or of any fitted scalar. Design knobs (mask ratio p=0.2, 3 Euler steps, dual-stream architecture, velocity blending) are ablated on the same held-out data and do not define the evaluation quantities by construction. Citations (rectified flow, Flux, HaMeR, MANO, SLAM) are to independent prior work; no uniqueness theorem or load-bearing self-citation forces the central claim. Limitations (cmask conflating FOV exit with occlusion, MANO expressiveness, intrinsics) are acknowledged but do not create circularity. The derivation chain is therefore self-contained against external benchmarks.

Assumptions & free parameters 5 free parameters · 4 assumptions · 2 invented entities

As an empirical CV methods paper, HandFlow rests on standard generative-modeling and hand-modeling assumptions plus a handful of design hyperparameters chosen by ablation. No new physical entities are postulated; the free parameters are architectural and training knobs that affect the reported numbers but are disclosed and ablated. The load-bearing domain assumptions are the adequacy of MANO, the reliability of HaMeR confidence for continuous masking, and the straight-path rectified-flow inductive bias for few-step sampling.

free parameters (5)
  • training-time random mask ratio p = 0.2
    Chosen as p=0.2 after ablation; trades per-frame accuracy for temporal coherence and is not derived from first principles.
  • ODE integration steps (Euler) = 3
    Fixed at 3 after a 1–10 step sweep; operating point on the accuracy plateau, not theoretically forced.
  • auxiliary loss weights (L_β, L_vel, L_acc, L_reproj, L_j3d) = 1.0 / 0.5 / 0.5 / 0.5 / 0.05
    Hand-set coefficients (1.0, 0.5, 0.5, 0.5, 0.05) scaled by mean timestep; affect the composite objective that produces the reported metrics.
  • window length T and overlap o = T=16, o=2
    T=16 and o=2 selected for the main experiments; control temporal context and boundary blending.
  • architecture size (D, dual/single blocks, heads) = D=512, 8+16 blocks
    D=512, 8 dual-stream + 16 single-stream blocks, 16 heads—standard capacity choices that determine capacity and VRAM.
assumptions (4)
  • domain assumption The MANO parametric model (shape β∈R^10, pose θ∈R^48, translation τ∈R^3) is a sufficiently expressive and accurate representation of real hand geometry and kinematics for the evaluation metrics used.
    All generation and losses operate in MANO parameter space; limitations section explicitly notes accuracy is bounded by MANO expressiveness and annotation quality.
  • domain assumption Rectified flow’s linear interpolation path yields a velocity field that can be accurately integrated with few Euler steps for sequence-level MANO generation.
    Core sampling procedure (§3.1); supported by the ODE-step sweep but still an inductive bias imported from Liu et al. / Lipman et al.
  • ad hoc to paper HaMeR detection confidence m_i ∈ [0,1] is a meaningful continuous reliability signal for interpolating observed condition tokens with learnable mask tokens.
    Defines cmask (Eq. 5); the paper’s own limitations admit it conflates occlusion with out-of-view cases.
  • domain assumption Standard flow-matching training (logit-normal time sampling, velocity L2 loss) plus geometric/temporal auxiliary losses produces a useful conditional generative prior over hand motion.
    Training objective in §3.1 and §3.4; conventional in the flow-matching literature.
invented entities (2)
  • cmask (confidence-aware continuous masking)
    purpose: Softly replace unreliable visual/skeletal tokens with learnable mask tokens according to per-frame detection confidence so the generative prior can complete missing observations.
    Introduced in §3.2 as a paper-specific mechanism; ablated but has no independent existence outside this architecture.
  • velocity-blended overlapping-window inference
    purpose: Fuse multi-window ODE trajectories during integration (rather than post-hoc averaging) to keep long-video hand trajectories continuous.
    Described in §3.5 and Appendix F; a procedural invention of the paper, not a previously established object.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HandFlow: Fully Generative 4D Hand Recovery with Flow Matching." pith.science (2026). https://pith.science/paper/4TD2QICW

@misc{pith2026260711221,
  author       = {Pith},
  title        = {Pith review of: HandFlow: Fully Generative 4D Hand Recovery with Flow Matching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4TD2QICW}},
  note         = {Machine review of arXiv:2607.11221}
}
read the original abstract

Accurate monocular 4D hand reconstruction remains challenging. Per-frame discriminative regressors lack temporal context and often produce jittery predictions. Temporal models improve consistency by aggregating information across frames, but they are typically deterministic regressors, making them vulnerable to ambiguous observations caused by occlusion and motion blur. Generative modeling offers a natural alternative by learning a prior over plausible hand motion sequences, enabling coherent hand-state recovery when visual evidence is incomplete or unreliable. Motivated by this observation, we present HandFlow, a fully generative flow-matching framework for temporally coherent 3D hand pose and shape estimation from monocular video. Given visual and skeletal observations, HandFlow denoises an entire temporal window of MANO parameters through a single ODE integration. To support this, we use a Flux-style dual-stream transformer that attends across the full sequence to capture long-range dependencies without autoregressive decoding, and a confidence-aware continuous masking mechanism that blends observed features with learnable mask tokens to handle noisy or missing observations. Experiments on DexYCB and HOT3D show that HandFlow achieves state-of-the-art performance, with particularly large gains in world-space accuracy and temporal smoothness. It reduces world-space pose error by over 30% compared with the strongest baseline and achieves the lowest acceleration error among all evaluated methods, while remaining competitive in per-frame pose accuracy. Moreover, on a single GPU HandFlow reconstructs a 150-frame sequence at 47 fps, about 12x faster than the fastest prior video-based method, with reconstruction itself accounting for only a small fraction of the end-to-end latency.

Figures

Figures reproduced from arXiv: 2607.11221 by the authors.

Figure 1
Figure 1. Given monocular RGB videos, HandFlow reconstructs temporally coherent 4D hand motion through a fully generative flow-matching framework. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of HandFlow. A frozen HaMeR [Pavlakos et al [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Comparison on heavily occluded DexYCB frames: GT vs. HaMeR, HaWoR, and HandFlow. Each method is shown in four columns (from left to right: [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: HandFlow tracking on heavily occluded clips from DexYCB and HOT3D (eight sampled frames per sequence, overlaid on input views), with plausible [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: HOT3D sequences rendered for HaMeR, HaWoR, HandFlow, and GT from a shared 3D viewpoint. HandFlow best matches GT in completeness and [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: For each sequence, four sampled frames show wrist acceleration arrows (length encodes magnitude), with a time-series plot of acceleration magnitude [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: ODE steps vs. MPJPE. Five MPJPE variants (M, PA-M, RA-M, W-M, WA-M) across 1–10 Euler integration steps. Key observations. (1) Step 1 collapses under Euler truncation error (MPJPE ≈ 43 mm), as a single-step integration cannot resolve the continuous flow trajectory. (2)…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

68 extracted references · 3 canonical work pages

  1. [1]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Yu, Zhengdi and Zafeiriou, Stefanos and Birdal, Tolga , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2025 , doi =

    2025

  2. [2]

    and Tulsiani, Shubham , title =

    Ye, Yufei and Feng, Yao and Taheri, Omid and Feng, Haiwen and Black, Michael J. and Tulsiani, Shubham , title =. Thirteenth International Conference on. 2026 , url =

    2026

  3. [3]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Pavlakos, Georgios and Shan, Dandan and Radosavovic, Ilija and Kanazawa, Angjoo and Fouhey, David and Malik, Jitendra , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2024 , doi =

    2024

  4. [4]

    , title =

    Romero, Javier and Tzionas, Dimitrios and Black, Michael J. , title =. 2017 , doi =

    2017

  5. [5]

    2026 , doi =

    Lin, Dixuan and Zhang, Yuxiang and Li, Mengcheng and Jing, Wei and Yan, Qi and Wang, Qianying and Liu, Yebin and Zhang, Hongwen , title =. 2026 , doi =

    2026

  6. [6]

    , title =

    Fu, Qichen and Liu, Xingyu and Xu, Ran and Niebles, Juan Carlos and Kitani, Kris M. , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =. 2023 , doi =

    2023

  7. [7]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Potamias, Rolandos Alexandros and Zhang, Jinglei and Deng, Jiankang and Zafeiriou, Stefanos , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2025 , doi =

    2025

  8. [8]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Zhang, Jinglei and Deng, Jiankang and Ma, Chao and Potamias, Rolandos Alexandros , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

Show all 68 references
  1. [9]

    Bimanual

    Prakash, Aditya and Forsyth, David and Gupta, Saurabh , month = oct, year =. Bimanual. doi:10.48550/arXiv.2510.06145 , abstract =

    doi

  2. [10]

    arXiv preprint arXiv:1906.08172 , author =

    Mediapipe:. arXiv preprint arXiv:1906.08172 , author =

    1906 arXiv

  3. [11]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

    Zhen, Haoyu and Sun, Qiao and Zhang, Hongxin and Li, Junyan and Zhou, Siyuan and Du, Yilun and Gan, Chuang , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =. 2025 , doi =

    2025

  4. [12]

    Woo, Taeyun and Park, Wonjung and Jeong, Woohyun and Park, Jinah , title =. Comput. Graph. , volume =. 2023 , doi =

    2023

  5. [13]

    Huang, Lin and Lin, Chung-Ching and Lin, Kevin and Liang, Lin and Wang, Lijuan and Yuan, Junsong and Liu, Zicheng , month = jun, year =. Neural. 2023. doi:10.1109/CVPR52729.2023.00866 , abstract =

    2023 doi

  6. [14]

    Computer

    Moon, Gyeongsik and Lee, Kyoung Mu , editor =. Computer. 2020 , pages =. doi:10.1007/978-3-030-58571-6_44 , abstract =

    2020 doi

  7. [15]

    Proceedings of the 20th

    Ali, Muhammad and Robertini, Nadia and Stricker, Didier , year =. Proceedings of the 20th. doi:10.5220/0013107300003912 , abstract =

    doi

  8. [16]

    Wang, Shuaibing and Wang, Shunli and Kuang, HaoPeng and Li, Fang and Qian, Ziyun and Li, Mingcheng and Zhang, Lihua , month = nov, year =. A. 2022. doi:10.1109/CCIS57298.2022.10016310 , abstract =

    2022 doi

  9. [17]

    and Liu, Changliu and Shi, Guanya , title =

    He, Tairan and Luo, Zhengyi and He, Xialin and Xiao, Wenli and Zhang, Chong and Zhang, Weinan and Kitani, Kris M. and Liu, Changliu and Shi, Guanya , title =. Proceedings of Machine Learning Research , volume =. 2025 , url =

    2025

  10. [18]

    The Fourteenth International Conference on Learning Representations , year =

    Liu, Zhizheng and Lin, Joe and Wu, Wayne and Zhou, Bolei , title =. The Fourteenth International Conference on Learning Representations , year =

  11. [19]

    The Fourteenth International Conference on Learning Representations , year =

    Zhou, Kaichen and Wang, Yuhan and Chen, Grace and Beaudouin, Gaspard and Zhan, Fangneng and Liang, Paul Pu and Wang, Mengyu , title =. The Fourteenth International Conference on Learning Representations , year =

  12. [20]

    Computer Vision -- ECCV 2024 , pages =

    Prakash, Aditya and Tu, Ruisen and Chang, Matthew and Gupta, Saurabh , title =. Computer Vision -- ECCV 2024 , pages =. 2025 , publisher =

    2024

  13. [21]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Choi, Hongsuk and Moon, Gyeongsik and Chang, Ju Yong and Lee, Kyoung Mu , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2021 , doi =

    2021

  14. [22]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

    Saleem, Muhammad Usama and Pinyoanuntapong, Ekkasit and Patel, Mayur Jagdishbhai and Xue, Hongfei and Helmy, Ahmed and Das, Srijan and Wang, Pu , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =. 2025 , abstract =

    2025

  15. [23]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Chen, Xingyu and Song, Zhuheng and Jiang, Xiaoke and Hu, Yaoqing and Yu, Junzhi and Zhang, Lei , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2025 , abstract =

    2025

  16. [24]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Cheng, Wencan and Tang, Hao and Van Gool, Luc and Ko, Jong Hwan , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2024 , abstract =

    2024

  17. [25]

    doi:10.48550/arXiv.2601.11514 , abstract =

    Siddiqui, Yawar and Frost, Duncan and Aroudj, Samir and Avetisyan, Armen and Howard-Jenkins, Henry and DeTone, Daniel and Moulon, Pierre and Wu, Qirui and Li, Zhengqin and Straub, Julian and Newcombe, Richard and Engel, Jakob , month = jan, year =. doi:10.48550/arXiv.2601.1151...

    doi

  18. [26]

    The Fourteenth International Conference on Learning Representations , year =

    Sun, Zhihao and Wu, Tong and Tu, Ruirui and Dong, Daoguo and Wu, Zuxuan , title =. The Fourteenth International Conference on Learning Representations , year =

  19. [27]

    and Van Wyk, Karl and Iqbal, Umar and Birchfield, Stan and Kautz, Jan and Fox, Dieter , title =

    Chao, Yu-Wei and Yang, Wei and Xiang, Yu and Molchanov, Pavlo and Handa, Ankur and Tremblay, Jonathan and Narang, Yashraj S. and Van Wyk, Karl and Iqbal, Umar and Birchfield, Stan and Kautz, Jan and Fox, Dieter , title =. Proceedings of the IEEE/CVF Conference on Computer Visi...

    2021

  20. [28]

    Zhang, Jiahui and Li, Yuelei and Chen, Anpei and Xu, Muyu and Liu, Kunhao and Wang, Jianyuan and Long, Xiao-Xiao and Liang, Hanxue and Xu, Zexiang and Su, Hao and Theobalt, Christian and Rupprecht, Christian and Vedaldi, Andrea and Zhou, Kaichen and Pfister, Hanspeter and Lian...

    2025

  21. [29]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Ye, Vickie and Pavlakos, Georgios and Malik, Jitendra and Kanazawa, Angjoo , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2023 , abstract =

    2023

  22. [30]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Chen, Rui and Zhang, Jianfeng and Liang, Yixun and Luo, Guan and Li, Weiyu and Liu, Jiarui and Li, Xiu and Long, Xiaoxiao and Feng, Jiashi and Tan, Ping , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2025 , abstract =

    2025

  23. [31]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Banerjee, Prithviraj and Shkodrani, Sindi and Moulon, Pierre and Hampali, Shreyas and Han, Shangchen and Zhang, Fan and Zhang, Linguang and Fountain, Jade and Miller, Edward and Basol, Selen and Newcombe, Richard and Wang, Robert and Engel, Jakob Julian and Hodan, Tomas , titl...

    2025

  24. [32]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

    Tian, Rui and Dai, Qi and Bao, Jianmin and Qiu, Kai and Yang, Yifan and Luo, Chong and Wu, Zuxuan and Jiang, Yu-Gang , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =. 2025 , abstract =

    2025

  25. [33]

    2023 , doi =

    Zhang, Biao and Tang, Jiapeng and Nie. 2023 , doi =

    2023

  26. [34]

    Wan, Team and Wang, Ang and Ai, Baole and Wen, Bin and Mao, Chaojie and Xie, Chen-Wei and Chen, Di and Yu, Feiwu and Zhao, Haiming and Yang, Jianxiao and Zeng, Jianyuan and Wang, Jiayu and Zhang, Jingfeng and Zhou, Jingren and Wang, Jinkai and Chen, Jixuan and Zhu, Kai and Zha...

    doi arXiv

  27. [35]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Xu, Zhongcong and Zhang, Jianfeng and Liew, Jun Hao and Yan, Hanshu and Liu, Jia-Wei and Zhang, Chenxu and Feng, Jiashi and Shou, Mike Zheng , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2024 , abstract =

    2024

  28. [36]

    , title =

    Rempe, Davis and Birdal, Tolga and Hertzmann, Aaron and Yang, Jimei and Sridhar, Srinath and Guibas, Leonidas J. , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =. 2021 , abstract =

    2021

  29. [37]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

    Zhang, Youliang and Li, Ronghui and Zhang, Yachao and Pan, Liang and Wang, Jingbo and Liu, Yebin and Li, Xiu , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =. 2025 , abstract =

    2025

  30. [38]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

    Chen, Shoufa and Sun, Peize and Song, Yibing and Luo, Ping , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =. 2023 , abstract =

    2023

  31. [39]

    doi:10.48550/arXiv.2601.10632 , abstract =

    Zhao, Chengfeng and Shu, Jiazhi and Zhao, Yubo and Huang, Tianyu and Lu, Jiahao and Gu, Zekai and Ren, Chengwei and Dou, Zhiyang and Shuai, Qing and Liu, Yuan , month = apr, year =. doi:10.48550/arXiv.2601.10632 , abstract =

    doi arXiv

  32. [40]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    He, Yannan and Tiwari, Garvita and Birdal, Tolga and Lenssen, Jan Eric and Pons-Moll, Gerard , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2024 , abstract =

    2024

  33. [41]

    and Li, Zhenyu and Zhao, Yang and Peng, Sida and Guo, Hengkai and Zhou, Xiaowei and Shi, Guang and Feng, Jiashi and Kang, Bingyi , title =

    Lin, Haotong and Chen, Sili and Liew, Jun Hao and Chen, Donny Y. and Li, Zhenyu and Zhao, Yang and Peng, Sida and Guo, Hengkai and Zhou, Xiaowei and Shi, Guang and Feng, Jiashi and Kang, Bingyi , title =. The Fourteenth International Conference on Learning Representations , year =

  34. [42]

    doi:10.48550/arXiv.2508.10934 , abstract =

    Huang, Jiahui and Zhou, Qunjie and Rabeti, Hesam and Korovko, Aleksandr and Ling, Huan and Ren, Xuanchi and Shen, Tianchang and Gao, Jun and Slepichev, Dmitry and Lin, Chen-Hsuan and Ren, Jiawei and Xie, Kevin and Biswas, Joydeep and Leal-Taixe, Laura and Fidler, Sanja , month...

    doi arXiv

  35. [43]

    doi:10.48550/arXiv.2603.03265 , abstract =

    Wang, Yufu and Ng, Evonne and Shin, Soyong and Khirodkar, Rawal and Dong, Yuan and Su, Zhaoen and Park, Jinhyung and Kitani, Kris and Richard, Alexander and Prada, Fabian and Zollhofer, Michael , month = mar, year =. doi:10.48550/arXiv.2603.03265 , abstract =

    doi

  36. [44]

    doi:10.48550/arXiv.2601.01222 , abstract =

    Li, Mengfei and Li, Peng and Zhang, Zheng and Lu, Jiahao and Zhao, Chengfeng and Xue, Wei and Liu, Qifeng and Peng, Sida and Zhang, Wenxiao and Luo, Wenhan and Liu, Yuan and Guo, Yike , month = jan, year =. doi:10.48550/arXiv.2601.01222 , abstract =

    doi

  37. [45]

    , booktitle =

    Tevet, Guy and Raab, Sigal and Gordon, Brian and Shafir, Yoni and Cohen-Or, Daniel and Bermano, Amit H. , booktitle =. Human Motion Diffusion Model , url =. 2023 , note =

    2023

  38. [46]

    2024 , url =

    Flux , author =. 2024 , url =

    2024

  39. [47]

    RoFormer: Enhanced Transformer with Rotary Position Embedding , url =

    Su, Jianlin and Yu, Lu and Pan, Shengfeng and Murtadha, Ahmed and Wen, Bo and Liu, Yunfeng , journal =. RoFormer: Enhanced Transformer with Rotary Position Embedding , url =. 2024 , note =

    2024

  40. [48]

    Scalable Diffusion Models with Transformers , url =

    Peebles, William and Xie, Saining , booktitle =. Scalable Diffusion Models with Transformers , url =. 2023 , note =

    2023

  41. [49]

    doi:10.48550/arXiv.2602.05755 , abstract =

    Wang, Ti and Yu, Xiaohang and Mathis, Mackenzie Weygandt , month = feb, year =. doi:10.48550/arXiv.2602.05755 , abstract =

    doi

  42. [50]

    Spurr, Adrian and Song, Jie and Park, Seonwook and Hilliges, Otmar , year =. Self-. European

  43. [51]

    Lin, Kevin and Wang, Lijuan and Liu, Zicheng , year =. IEEE/

  44. [52]

    Chen, Yufei and others , year =. IEEE/

  45. [53]

    Park, Jihyun and Oh, Youngtaek and Moon, Gyeongsik and Choi, Hongsuk and Lee, Kyoung Mu , year =. IEEE/

  46. [54]

    Chen, Yujin and Tu, Zhigang and Ge, Liuhao and Zhang, Dejun and Chen, Ruizhi and Yuan, Junsong , year =. IEEE/

  47. [55]

    , year =

    Kocabas, Muhammed and Athanasiou, Nikos and Black, Michael J. , year =. IEEE/

  48. [56]

    Valassakis, Eugene and Garcia-Hernando, Guillermo , booktitle =

  49. [57]

    , booktitle =

    Duran, Enes and Kocabas, Muhammed and Choutas, Vasileios and Fan, Zicong and Black, Michael J. , booktitle =

  50. [58]

    The Eleventh International Conference on Learning Representations , year =

    Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow , author =. The Eleventh International Conference on Learning Representations , year =

  51. [59]

    The Eleventh International Conference on Learning Representations , year =

    Flow Matching for Generative Modeling , author =. The Eleventh International Conference on Learning Representations , year =

  52. [60]

    ACM Transactions on Graphics (ToG) , volume=

    Rgb2hands: real-time tracking of 3d hand interactions from monocular rgb video , author=. ACM Transactions on Graphics (ToG) , volume=. 2020 , publisher=

    2020

  53. [61]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Hoi4d: A 4d egocentric dataset for category-level human-object interaction , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  54. [62]

    2024 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Synh2r: Synthesizing hand-object motions for learning human-to-robot handovers , author=. 2024 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2024 , organization=

    2024

  55. [63]

    arXiv preprint arXiv:2511.09484 , year=

    Spider: Scalable physics-informed dexterous retargeting , author=. arXiv preprint arXiv:2511.09484 , year=

  56. [64]

    Communications of the ACM , volume=

    Nerf: Representing scenes as neural radiance fields for view synthesis , author=. Communications of the ACM , volume=. 2021 , publisher=

    2021

  57. [65]

    European Conference on Computer Vision , pages=

    Multi-hmr: Multi-person whole-body human mesh recovery in a single shot , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  58. [66]

    Proceedings of the SIGGRAPH Asia 2025 Conference Papers , pages=

    PhysHMR: Learning Humanoid Control Policies from Vision for Physically Plausible Human Motion Reconstruction , author=. Proceedings of the SIGGRAPH Asia 2025 Conference Papers , pages=

    2025

  59. [67]

    European Conference on Computer Vision , pages=

    Cliff: Carrying location information in full frames into human pose and shape estimation , author=. European Conference on Computer Vision , pages=. 2022 , organization=

    2022

  60. [68]

    doi:10.48550/arXiv.2508.10104 , author =

    doi arXiv

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.