Pith. sign in

REVIEW 2 major objections 4 minor 55 references

GNOCHI: Generative Neural mOdel for Close Human-Human Interactions

T0 review · 2 major / 4 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 A conditional generative model synthesizes diverse, collision-free 3D poses for one human reacting to another in close contact.

desk verdict Solid engineering paper that actually delivers controllable, low-collision close-contact poses; the augmentation realism is the soft spot but not fatal. read the letter →

arxiv 2607.10408 v1 pith:SKYTVOW5 submitted 2026-07-11 cs.CV

Gonzalo Gómez-Nogales , Marc Comino-Trinidad , Andrés Casado-Elvira , Dan Casas This is my paper · ORCID
classification cs.CV
keywords human-humaninteractionconditionalVAEcollisionresolutionposegenerationdataaugmentationSMPLclosecontact
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

Realistic close contact between two 3D humans is hard because body joints have many degrees of freedom and meshes must touch without interpenetrating. Existing methods either reconstruct observed interactions or generate motions from text, yet they rarely produce controlled, diverse close contacts that stay physically plausible. GNOCHI trains a conditional variational autoencoder on an automatically expanded set of contact poses so that, given one person's pose, it can sample many natural reaction poses. A second lightweight module then nudges the reaction joints just enough to clear residual collisions. The resulting pairs are more coherent and far less intersecting than prior generative baselines, and people rate them as plausible as real captured data.

What carries the argument

The GNOCHI pipeline: individual-level joint noise projected through VPoser, followed by SDF collision resolution to build the training set; a pose-conditioned cVAE that samples reaction parameters; and a frozen-decoder CapFix module that predicts joint offsets minimizing capsule penetration while preserving global root pose.

What would settle it

Generate reaction poses for a large held-out set of conditioning poses drawn from sources never used in training or augmentation (for example Mixamo keyframes or real multi-person mocap), then measure whether intersection volume, IoU and human Likert ratings remain statistically indistinguishable from those of real ground-truth contact pairs.

Watch

Extended reading notes

Core claim

Conditioning a variational autoencoder on one SMPL pose and training it on a large set of synthetically augmented yet collision-resolved contact pairs, then refining the output with a self-supervised capsule-based collision fixer, yields diverse, semantically consistent reaction poses that remain free of mesh interpenetration.

Load-bearing premise

The synthetic contact poses created by adding joint noise, projecting through a pose prior, and resolving collisions with signed-distance fields are realistic and unbiased enough that the model generalizes to true unseen interactions without systematic artifacts the fixer cannot correct.

Editorial extensions

If this is right

  • Artists can treat the model as a pose copilot: fix one avatar and instantly obtain multiple collision-free reaction candidates that already respect contact.
  • Existing monocular multi-person trackers can be refined by sampling reaction partners that resolve residual penetrations while preserving the tracked conditioning pose.
  • Depth- or pose-conditioned image generators receive precise two-person control signals, enabling photorealistic close-interaction synthesis without manual keyframing.
  • The same augmentation-plus-fixer recipe can be applied to enlarge other scarce contact datasets for grasping or multi-character animation.

Reading between the lines

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

  • Because the latent space is continuous and conditioned, the same architecture could later be unrolled over time to generate short contact sequences rather than static keyframes.
  • The capsule-based fixer is intentionally geometric rather than physical; adding soft-tissue or balance constraints would further reduce rare local-minima failures on extreme entanglements.
  • If the synthetic distribution proves slightly biased toward certain contact types, the same pipeline could re-weight or filter augmented samples using a lightweight social-plausibility classifier before training.
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 / 4 minor

Summary. The paper presents GNOCHI, a conditional VAE that generates a reaction SMPL pose given a conditioning pose for close human-human interactions. To address data scarcity it introduces an automated augmentation pipeline (per-joint axis-angle noise, VPoser projection, then SDF-based collision resolve) that expands Hi4D contact frames roughly 10 imes. Residual collisions at generation time are handled by a second self-supervised CapFix (CF) module that predicts local joint offsets using capsule-proxy distances. Ablations (Table 1) isolate the contributions of augmentation versus CF; a held-out contact-frame comparison to BUDDI (Table 2) shows substantially lower intersection volume/IoU; qualitative interpolations and multi-sample figures, plus a 35-participant TOST user study, support perceptual equivalence to real captures.

Significance. Close-contact generative modeling remains a genuine bottleneck for animation, tracking refinement, and controllable image synthesis. The combination of a practical, fully automatic interaction-augmentation pipeline with an explicit collision-aware decoder is a useful engineering contribution that demonstrably reduces mesh intersections relative to the strongest published baseline (BUDDI) under the same conditioning. The sequence-heldout split, external Mixamo OOD set, and perceptual equivalence test give the central claim more empirical grounding than is common in this sub-area. If the reported gains hold under broader scrutiny, the method supplies a usable 3D pose co-pilot for artists and a post-process for monocular multi-person recovery.

major comments (2)
  1. Section 3 (and the weakest-assumption concern): the realism and coverage of the synthetic contact distribution produced by axis-angle noise + VPoser + single-Newton SDF resolve is load-bearing for the claimed OOD generalization (Mixamo keyframes, BUDDI-generated conditioners). The paper already supplies three mitigating strands (sequence-heldout split, TOST user study, CF trained on independent VPoser samples), yet a quantitative characterization of residual contact-type bias or first-contact preservation after augmentation is still missing. Adding a short analysis (e.g., contact-region histograms or CCD comparison on a subset) would strengthen the claim that the gains are not partly artifactual.
  2. Table 2 / Section 5.2: the BUDDI comparison is performed only on held-out Hi4D contact frames after an in-painting adaptation. While the intersection metrics are clear, the paper should also report the same Inter/IoU numbers on the Mixamo OOD set (or another external source) so that the relative advantage is not confined to the distribution on which both models saw related data.
minor comments (4)
  1. Clarify the precise definition of the contact threshold used both for selecting the Hi4D subset and for filtering augmented samples after noise injection (Section 3.3).
  2. The capsule-tolerance D_T is described as the 95th percentile of training distances; state whether this value is frozen after the first epoch or recomputed, and report its numerical value for reproducibility.
  3. Figure 4 and the surrounding text use both “cVAE” and “cV AE”; standardize the acronym.
  4. In the conclusions the authors correctly note the absence of continuous collision detection and of an explicit self-contact term inside CF; a short quantitative statement of residual self-penetration rates on the test set would make the limitation more precise.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: standard data-driven cVAE + self-supervised refinement trained and evaluated on held-out/OOD data with independent metrics and external baselines.

full rationale

The paper's core claims are empirical performance of a conditional generative model (cVAE decoder + CF module) for pose-conditioned close interactions. Data augmentation (per-joint axis-angle noise projected via VPoser then SDF-resolved, Sec. 3) produces training samples; the model is trained with standard reconstruction + KL + capsule-based collision losses (Eqs. 7-11) whose free weights and DT threshold (95th percentile) are not used to force any reported metric by construction. Evaluation uses sequence-heldout splits of Hi4D, external Mixamo keyframes, adapted BUDDI comparisons (Tables 1-2), latent interpolations, and a 35-person TOST user study against real Hi4D ground truth. No equation reduces a claimed prediction to a fitted input; no uniqueness theorem or load-bearing self-citation is invoked; no ansatz is smuggled via prior author work. The pipeline is self-contained against external benchmarks and does not exhibit any of the six circularity patterns.

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

The central claim rests on standard body models (SMPL, VPoser), a physics-inspired but approximate collision solver, and several hand-chosen loss weights and thresholds. No new physical entities are postulated; the free parameters are ordinary ML hyperparameters and one data-derived tolerance.

free parameters (5)
  • λ_KL = 1e-3
    Kullback-Leibler weight in the cVAE loss; set empirically to 1e-3.
  • λ_rec = 7.5
    Surface reconstruction weight; set empirically to 7.5.
  • λ_col, λ_joints, λ_Δθ = 30 / 0.5 / 0.1
    Weights of the three terms in the CF self-supervised loss; set to 30, 0.5, 0.1.
  • capsule distance tolerance D_T = 95th percentile of training distances
    95th percentile of capsule distances on the training set; used to decide when the collision penalty activates.
  • per-joint noise bounds (w_x,w_y,w_z, α_M) = joint-specific (not tabulated)
    Hand-designed maximum sweep angles and axis weights for the stochastic pose augmentation.
assumptions (4)
  • domain assumption SMPL body model and 6D continuous rotation representation are adequate for close-contact interactions.
    Used throughout Sections 3–4; standard in the field but known to be imperfect for soft tissue and clothing.
  • domain assumption VPoser latent space is a faithful prior over anatomically plausible human poses.
    Invoked in Section 3.1 to project noisy joint angles back into a valid pose manifold.
  • ad hoc to paper A single Newton iteration of the SDF penalty force is sufficient to resolve interpenetrations for data augmentation.
    Stated in Section 3.2; no ablation on number of iterations or residual penetration after resolve.
  • ad hoc to paper 24 rigid capsules per body are a sufficiently accurate proxy for collision detection at training time.
    Section 4.3; the tolerance D_T is introduced precisely because capsules are loose.
invented entities (2)
  • CF (CapFix) collision-resolution network
    purpose: Predicts joint-angle offsets that remove residual inter-body capsule intersections while freezing global translation/rotation.
    Introduced in Section 4.3; trained self-supervised on the frozen cVAE decoder. No independent physical validation outside the paper's own metrics.
  • Automated supervised interaction data-augmentation pipeline (noise + VPoser + SDF resolve)
    purpose: Multiplies scarce Hi4D contact frames by ~10× while preserving contact and removing penetrations.
    Core contribution of Section 3; the resulting 61k-pose set is used to train the cVAE. Realism is supported only by the paper's own user study and qualitative figures.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GNOCHI: Generative Neural mOdel for Close Human-Human Interactions." pith.science (2026). https://pith.science/paper/SKYTVOW5

@misc{pith2026260710408,
  author       = {Pith},
  title        = {Pith review of: GNOCHI: Generative Neural mOdel for Close Human-Human Interactions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SKYTVOW5}},
  note         = {Machine review of arXiv:2607.10408}
}
read the original abstract

Creating realistic 3D human-human interactions in virtual environments is challenging due to the high degrees of freedom in the human body and the need for physically accurate poses that do not collide with each other. Traditional methods for human-human interaction are based on motion tracking or 3D body reconstruction, but lack generative capabilities. Recent generative methods enable the synthesis of individual or interacting motions via text or image input, but generally fall short in modeling close interactions. This paper introduces a novel generative model for close 3D human-human interactions using a conditional variational autoencoder (cVAE), which generates poses for one human conditioned on the pose of another, allowing for controlled and diverse interaction synthesis. To train our model, we address two underlying long-standing challenges in the field of human-human interaction: data scarcity, for which we propose an automated supervised data augmentation strategy that generates synthetic yet realistic interaction poses; and collision awareness in generative approaches, for which we propose a self-supervised loss based on a collision resolution technique using volumetric proxies to ensure physically correct interactions. We extensively evaluate the capabilities of our model, and demonstrate a wide variety of plausible and physically correct interactions, not possible to generate with current state-of-the-art methods.

Figures

Figures reproduced from arXiv: 2607.10408 by the authors.

Figure 1
Figure 1. Given an input 3D pose (i.e., the conditioning pose, in purple), our generative model infers a 3D pose of a human in close interaction (i.e., the reaction pose, in green). This enables the conditional generation of 3D humans in close interaction, which can be used fine-grain control signal for image generative methods (right). Abstract Creating realistic 3D human-human interactions in virtual environments is challen… view at source ↗
Figure 2
Figure 2. From an existing dataset [YGK∗ 23] (left), we apply individual-level noise at each body joint (Section 3.1), and jointly resolve the collisions using a physical simulator (Section 3.2) to create an augmented dataset. At train time, our model learns to infer a reacting avatar given a conditioning avatar (center top, Section 4.2). Then, a second model is trained to fix collisions between two interacting avatars (cente… view at source ↗
Figure 3
Figure 3. Data augmentation steps. We add noise to the origi￾nal poses (3a), yielding augmented poses (3b, 3c, and 3d) which may contain collisions. Our physics solver fixes these, yielding physically-correct interactions (3e, 3f, and 3g). 3.3. Final Dataset Details For each avatar pair in the Hi4D dataset, we generated 10 varia￾tions. Because the random noise introduced during pose augmenta￾tion can cause the avatars to drif… view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5 [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Our results (i.e., renders with conditioning and reaction poses) can be used as accurate driving signal to control image-to￾image models [Lab24]. This enables fine-control synthesis of pho￾torealistic images with complex human-human interactions. BUDDI by sampling dive…
Figure 7
Figure 7. Figure 7: Two examples (Interpolation 1 and Interpolation 2) of the reacting poses yielded by interpolating two latent vectors (left and right columns) for a constant conditioning pose. The conditioning poses were originally generated with BUDDI, hence were never seen at train t…
Figure 8
Figure 8. Figure 8: Examples of conditional sampling from our model and BUDDI [MYP∗ 24] for a set of conditioning pose, labeled as dance, from the Hi4D dataset. These conditioning poses belong to the in-distribution benchmark. We show two views and four different samples, organized in fou…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 6 canonical work pages

  1. [1]

    Ghosh, Anindita and Zhou, Bing and Dabral, Rishabh and Wang, Jian and Golyanik, Vladislav and Theobalt, Christian and Slusallek, Philipp and Guo, Chuan , booktitle =

  2. [2]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    MultiPly: Reconstruction of Multiple People from Monocular Video in the Wild , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  3. [3]

    Liu, Shaowei and Guo, Chuan and Zhou, Bing and Wang, Jian , title =

  4. [4]

    2025 , pages=

    Muhammad Gohar Javed and Chuan Guo and Li Cheng and Xingyu Li , booktitle=. 2025 , pages=

    2025

  5. [5]

    2024 , issue_date =

    Liang, Han and Zhang, Wenqian and Li, Wenxuan and Yu, Jingyi and Xu, Lan , title =. 2024 , issue_date =. doi:10.1007/s11263-024-02042-6 , journal =

    work page doi:10.1007/s11263-024-02042-6 2024

  6. [6]

    FreeMotion: A Unified Framework for Number-Free Text-to-Motion Synthesis

    Fan, Ke and Tang, Junshu and Cao, Weijian and Yi, Ran and Li, Moran and Gong, Jingyu and Zhang, Jiangning and Wang, Yabiao and Wang, Chengjie and Ma, Lizhuang. FreeMotion: A Unified Framework for Number-Free Text-to-Motion Synthesis. Computer Vision -- ECCV 2024. 2025

    2024

  7. [7]

    Towards Open Domain Text-Driven Synthesis of Multi-person Motions

    Shan, Mengyi and Dong, Lu and Han, Yutao and Yao, Yuan and Liu, Tao and Nwogu, Ifeoma and Qi, Guo-Jun and Hill, Mitch. Towards Open Domain Text-Driven Synthesis of Multi-person Motions. Computer Vision -- ECCV 2024. 2025

    2024

  8. [8]

    Role-aware Interaction Generation from Textual Description , year=

    Tanaka, Mikihiro and Fujiwara, Kent , booktitle=. Role-aware Interaction Generation from Textual Description , year=

Show all 55 references
  1. [9]

    2023 , volume=

    Chopin, Baptiste and Tang, Hao and Otberdout, Naima and Daoudi, Mohamed and Sebe, Nicu , journal=. 2023 , volume=

    2023

  2. [10]

    2023 , issue_date =

    Shuai, Qing and Yu, Zhiyuan and Zhou, Zhize and Fan, Lixin and Yang, Haijun and Yang, Can and Zhou, Xiaowei , title =. 2023 , issue_date =. doi:10.1145/3618336 , journal =

    2023 doi

  3. [11]

    GANerated Hands for Real-Time 3D Hand Tracking from Monocular RGB , year=

    Mueller, Franziska and Bernard, Florian and Sotnychenko, Oleksandr and Mehta, Dushyant and Sridhar, Srinath and Casas, Dan and Theobalt, Christian , booktitle=. GANerated Hands for Real-Time 3D Hand Tracking from Monocular RGB , year=

  4. [12]

    Proceedings of the 29th International Conference on Neural Information Processing Systems - Volume 2 , pages =

    Sohn, Kihyuk and Yan, Xinchen and Lee, Honglak , title =. Proceedings of the 29th International Conference on Neural Information Processing Systems - Volume 2 , pages =. 2015 , publisher =

    2015

  5. [13]

    On the Continuity of Rotation Representations in Neural Networks , year=

    Zhou, Yi and Barnes, Connelly and Lu, Jingwan and Yang, Jimei and Li, Hao , booktitle=. On the Continuity of Rotation Representations in Neural Networks , year=

  6. [14]

    and Tzionas, Dimitrios and Black, Michael J

    Pavlakos, Georgios and Choutas, Vasileios and Ghorbani, Nima and Bolkart, Timo and Osman, Ahmed A. and Tzionas, Dimitrios and Black, Michael J. , booktitle=. Expressive Body Capture: 3D Hands, Face, and Body From a Single Image , year=

  7. [15]

    Interaction Replica: Tracking Human–Object Interaction and Scene Changes From Human Motion , year=

    Guzov, Vladimir and Chibane, Julian and Marin, Riccardo and He, Yannan and Saracoglu, Yunus and Sattler, Torsten and Pons-Moll, Gerard , booktitle=. Interaction Replica: Tracking Human–Object Interaction and Scene Changes From Human Motion , year=

  8. [16]

    , booktitle=

    Hassan, Mohamed and Ghosh, Partha and Tesch, Joachim and Tzionas, Dimitrios and Black, Michael J. , booktitle=. Populating 3D Scenes by Learning Human-Scene Interaction , year=

  9. [17]

    Scaling Up Dynamic Human-Scene Interaction Modeling , year=

    Jiang, Nan and Zhang, Zhiyuan and Li, Hongjie and Ma, Xiaoxuan and Wang, Zan and Chen, Yixin and Liu, Tengyu and Zhu, Yixin and Huang, Siyuan , booktitle=. Scaling Up Dynamic Human-Scene Interaction Modeling , year=

  10. [18]

    Decoupling Human and Camera Motion from Videos in the Wild , year=

    Ye, Vickie and Pavlakos, Georgios and Malik, Jitendra and Kanazawa, Angjoo , booktitle=. Decoupling Human and Camera Motion from Videos in the Wild , year=

  11. [19]

    MultiPhys: Multi-Person Physics-Aware 3D Motion Estimation , year=

    Ugrinovic, Nicolas and Pan, Boxiao and Pavlakos, Georgios and Paschalidou, Despoina and Shen, Bokui and Sanchez-Riera, Jordi and Moreno-Noguer, Francesc and Guibas, Leonidas , booktitle=. MultiPhys: Multi-Person Physics-Aware 3D Motion Estimation , year=

  12. [20]

    Hi4D: 4D Instance Segmentation of Close Human Interaction , year=

    Yin, Yifei and Guo, Chen and Kaufmann, Manuel and Zarate, Juan Jose and Song, Jie and Hilliges, Otmar , booktitle=. Hi4D: 4D Instance Segmentation of Close Human Interaction , year=

  13. [21]

    Three-Dimensional Reconstruction of Human Interactions , year=

    Fieraru, Mihai and Zanfir, Mihai and Oneata, Elisabeta and Popa, Alin-Ionut and Olaru, Vlad and Sminchisescu, Cristian , booktitle=. Three-Dimensional Reconstruction of Human Interactions , year=

  14. [22]

    Generative Proxemics: A Prior for 3D Social Interaction from Images , year=

    Müller, Lea and Ye, Vickie and Pavlakos, Georgios and Black, Michael and Kanazawa, Angjoo , booktitle=. Generative Proxemics: A Prior for 3D Social Interaction from Images , year=

  15. [23]

    Müller, Lea and Osman, Ahmed A. A. and Tang, Siyu and Huang, Chun-Hao P. and Black, Michael J. , booktitle=. On Self-Contact and Human Pose , year=

  16. [24]

    and Laptev, Ivan and Schmid, Cordelia , booktitle=

    Hasson, Yana and Varol, Gül and Tzionas, Dimitrios and Kalevatykh, Igor and Black, Michael J. and Laptev, Ivan and Schmid, Cordelia , booktitle=. Learning Joint Reconstruction of Hands and Manipulated Objects , year=

  17. [25]

    G-HOP: Generative Hand-Object Prior for Interaction Reconstruction and Grasp Synthesis , year=

    Ye, Yufei and Gupta, Abhinav and Kitani, Kris and Tulsiani, Shubham , booktitle=. G-HOP: Generative Hand-Object Prior for Interaction Reconstruction and Grasp Synthesis , year=

  18. [26]

    , title =

    Loper, Matthew and Mahmood, Naureen and Romero, Javier and Pons-Moll, Gerard and Black, Michael J. , title =. 2015 , issue_date =. doi:10.1145/2816795.2818013 , journal =

    2015 doi

  19. [27]

    , booktitle=

    Sun, Yu and Liu, Wu and Bao, Qian and Fu, Yili and Mei, Tao and Black, Michael J. , booktitle=. Putting People in their Place: Monocular Regression of 3D People in Depth , year=

  20. [28]

    and Mei, Tao , booktitle=

    Sun, Yu and Bao, Qian and Liu, Wu and Fu, Yili and Black, Michael J. and Mei, Tao , booktitle=. Monocular, One-stage, Regression of Multiple 3D People , year=

  21. [29]

    Multi-Person Extreme Motion Prediction , year=

    Guo, Wen and Bie, Xiaoyu and Alameda-Pineda, Xavier and Moreno–Noguer, Francesc , booktitle=. Multi-Person Extreme Motion Prediction , year=

  22. [30]

    DeepMultiCap: Performance Capture of Multiple Characters Using Sparse Multiview Cameras , year=

    Zheng, Yang and Shao, Ruizhi and Zhang, Yuxiang and Yu, Tao and Zheng, Zerong and Dai, Qionghai and Liu, Yebin , booktitle=. DeepMultiCap: Performance Capture of Multiple Characters Using Sparse Multiview Cameras , year=

  23. [31]

    D-Grasp: Physically Plausible Dynamic Grasp Synthesis for Hand-Object Interactions , year=

    Christen, Sammy and Kocabas, Muhammed and Aksan, Emre and Hwangbo, Jemin and Song, Jie and Hilliges, Otmar , booktitle=. D-Grasp: Physically Plausible Dynamic Grasp Synthesis for Hand-Object Interactions , year=

  24. [32]

    2022 , isbn =

    Xie, Xianghui and Bhatnagar, Bharat Lal and Pons-Moll, Gerard , title =. 2022 , isbn =. doi:10.1007/978-3-031-20086-1_8 , booktitle =

    2022 doi

  25. [33]

    and Tzionas, Dimitrios , booktitle=

    Taheri, Omid and Choutas, Vasileios and Black, Michael J. and Tzionas, Dimitrios , booktitle=. GOAL: Generating 4D Whole-Body Motion for Hand-Object Grasping , year=

  26. [34]

    COAP: Compositional Articulated Occupancy of People , year=

    Mihajlovic, Marko and Saito, Shunsuke and Bansal, Aayush and Zollhoefer, Michael and Tang, Siyu , booktitle=. COAP: Compositional Articulated Occupancy of People , year=

  27. [35]

    Proceedings of the 35th International Conference on Neural Information Processing Systems , articleno =

    Fieraru, Mihai and Zanfir, Mihai and Szente, Teodor Alexandru and Bazavan, Eduard Gabriel and Olaru, Vlad and Sminchisescu, Cristian , title =. Proceedings of the 35th International Conference on Neural Information Processing Systems , articleno =. 2021 , isbn =

    2021

  28. [36]

    Stochastic Scene-Aware Motion Prediction , year=

    Hassan, Mohamed and Ceylan, Duygu and Villegas, Ruben and Saito, Jun and Yang, Jimei and Zhou, Yi and Black, Michael , booktitle=. Stochastic Scene-Aware Motion Prediction , year=

  29. [37]

    Inter-X: Towards Versatile Human-Human Interaction Analysis , year=

    Xu, Liang and Lv, Xintao and Yan, Yichao and Jin, Xin and Wu, Shuwen and Xu, Congsheng and Liu, Yifan and Zhou, Yizhou and Rao, Fengyun and Sheng, Xingdong and Liu, Yunhui and Zeng, Wenjun and Yang, Xiaokang , booktitle=. Inter-X: Towards Versatile Human-Human Interaction Anal...

  30. [38]

    Synthesizing Diverse Human Motions in 3D Indoor Scenes , year=

    Zhao, Kaifeng and Zhang, Yan and Wang, Shaofei and Beeler, Thabo and Tang, Siyu , booktitle=. Synthesizing Diverse Human Motions in 3D Indoor Scenes , year=

  31. [39]

    Compositional Human-Scene Interaction Synthesis with Semantic Control

    Zhao, Kaifeng and Wang, Shaofei and Zhang, Yan and Beeler, Thabo and Tang, Siyu. Compositional Human-Scene Interaction Synthesis with Semantic Control. Computer Vision -- ECCV 2022. 2022

    2022

  32. [40]

    Full-Body Articulated Human-Object Interaction , year=

    Jiang, Nan and Liu, Tengyu and Cao, Zhexuan and Cui, Jieming and Zhang, Zhiyuan and Chen, Yixin and Wang, He and Zhu, Yixin and Huang, Siyuan , booktitle=. Full-Body Articulated Human-Object Interaction , year=

  33. [41]

    IEEE transactions on visualization and computer graphics , volume=

    Optimization integrator for large time steps , author=. IEEE transactions on visualization and computer graphics , volume=. 2015 , publisher=

    2015

  34. [42]

    and Marsden, J

    Kane, C. and Marsden, J. E. and Ortiz, M. and West, M. , title =. International Journal for Numerical Methods in Engineering , volume =. doi:https://doi.org/10.1002/1097-0207(20001210)49:10<1295::AID-NME993>3.0.CO;2-W , year =

    doi

  35. [43]

    ACM Transactions on Graphics , volume=

    Resolving Collisions in Dense 3D Crowd Animations , author=. ACM Transactions on Graphics , volume=. 2024 , publisher=

    2024

  36. [44]

    , title =

    Ramon, Pablo and Romero, Cristian and Tapia, Javier and Otaduy, Miguel A. , title =. SIGGRAPH Asia Conference Papers (SA Conference Papers ’23) , year =

  37. [45]

    2024 , howpublished=

    Black Forest Labs , title=. 2024 , howpublished=

    2024

  38. [46]

    Reconstructing People, Places, and Cameras , year=

    Müller, Lea and Choi, Hongsuk and Zhang, Anthony and Yi, Brent and Malik, Jitendra and Kanazawa, Angjoo , booktitle=. Reconstructing People, Places, and Cameras , year=

  39. [47]

    Maluleke, Vongani H. and M. Synergy and Synchrony in Couple Dances , journal=. 2024 , volume=. doi:10.48550/arXiv.2409.04440 , url=

    2024 doi arXiv

  40. [48]

    Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data , year=

    Yang, Lihe and Kang, Bingyi and Huang, Zilong and Xu, Xiaogang and Feng, Jiashi and Zhao, Hengshuang , booktitle=. Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data , year=

  41. [49]

    Karen and Hertzmann, Aaron and Popovi\'

    Liu, C. Karen and Hertzmann, Aaron and Popovi\'. Composition of complex optimal multi-character motions , year =. Proceedings of the 2006 ACM SIGGRAPH/Eurographics Symposium on Computer Animation , pages =

    2006

  42. [50]

    Ho, Edmond S. L. and Komura, Taku and Tai, Chiew-Lan , title =. 2010 , isbn =. doi:10.1145/1833349.1778770 , booktitle =

    2010 doi

  43. [51]

    Proceedings of the ACM SIGGRAPH/Eurographics Symposium on Computer Animation , pages =

    Kim, Manmyung and Hwang, Youngseok and Hyun, Kyunglyul and Lee, Jehee , title =. Proceedings of the ACM SIGGRAPH/Eurographics Symposium on Computer Animation , pages =. 2012 , isbn =

    2012

  44. [52]

    and Lee, Jehee , title =

    Won, Jungdam and Lee, Kyungho and O'Sullivan, Carol and Hodgins, Jessica K. and Lee, Jehee , title =. 2014 , issue_date =. doi:10.1145/2661229.2661271 , journal =

    2014 doi

  45. [53]

    Shum, Hubert P. H. and Komura, Taku and Yamazaki, Shuntaro , title =. Proceedings of the 13th. 2007 , isbn =. doi:10.1145/1315184.1315194 , pages =

    2007 doi

  46. [54]

    and Komura, Taku and Yamazaki, Shuntaro , journal=

    Shum, Hubert P.H. and Komura, Taku and Yamazaki, Shuntaro , journal=. Simulating Multiple Character Interactions with Collaborative and Adversarial Goals , year=

  47. [55]

    2023 , isbn =

    Zhang, Yunbo and Gopinath, Deepak and Ye, Yuting and Hodgins, Jessica and Turk, Greg and Won, Jungdam , title =. 2023 , isbn =. doi:10.1145/3588432.3591491 , booktitle =

    2023 doi

Pith tools

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