Pith. sign in

REVIEW 3 major objections 5 minor 34 references

Forwardrobe: Garment-Aware Gaussian Avatars from a Single Image

T0 review · 3 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read From a single photo, Forwardrobe reconstructs the observed garment as an independent, motion-conditioned 3D Gaussian asset that can be reanimated and transferred to another avatar without per-garment optimization.

desk verdict A useful garment-separation contribution for single-image avatars, but the 'independent 3D garment asset' claim is under-evidenced: the pseudo-label supervision hinges on hallucinated back geometry, and no 3D garment metric or backbone-overlap disclosure is provided. read the letter →

arxiv 2607.29106 v1 pith:FV55DPH6 submitted 2026-07-31 cs.CV

Daisheng Jin , Shuyun Wang , Ying He This is my paper · ORCID
classification cs.CV
keywords garment-awareavatars3DGaussianSplattingsingle-imagereconstructionvirtualtry-ongarmenttransferlooseclothinganimationpose-conditioneddeformationfeed-forwardhumanavatar
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to show that loose clothing—the hardest part of single-image avatar reconstruction—can be handled properly if the garment is explicitly carved out in canonical Gaussian space and given its own geometry, skinning, and motion model. The method separates clothing from the body, closes the gaps where a skirt should span the legs, reassigns skinning weights from a continuous field, and adds pose-conditioned deformation and shading residuals. If correct, one photo yields a reusable 3D garment asset—animatable, editable, and transferable to other bodies—without any per-garment optimization. That would move virtual try-on from 2D image synthesis to true 3D replacement.

What carries the argument

The load-bearing object is the garment layer in canonical Gaussian space: a subset of Gaussians with their own offsets, skinning weights, spherical-harmonic appearance, and a residual motion model. Two mechanisms carry the argument: the continuity initialization (seam closing plus diffused skinning-weight reassignment for skirt-like garments) and the Garment Dynamic Module (a transformer–MLP adapter that predicts position, scale, rotation, SH, and shading residuals from motion, image, and language features). The composition step propagates garment displacements to nearby body Gaussians to keep collar and waist boundaries coherent.

What would settle it

Take a single image of a person from the front where the back of a loose garment is not visible, reconstruct with Forwardrobe, then render the avatar from behind and compare against a ground-truth back view. If the predicted back matches the ground truth, the assumption holds; if the back looks like the backbone's hallucination rather than the garment, the classifier's supervision is the failure point.

Watch

Extended reading notes

Core claim

The central claim is that garment decoupling can be done entirely in feed-forward fashion from a single image. Starting from a pre-built static Gaussian avatar, a labeling module classifies each Gaussian as garment or body using masks rendered from the input view plus front/back renders of the canonical avatar. A visual language model reading the image provides the category, and for skirt-like garments the inner-thigh Gaussians are pulled toward the body mid-plane and skinning weights are re-sampled from a diffused field so the garment deforms as one connected piece. A dynamic module then predicts bounded geometry, opacity, and appearance residuals conditioned on pose and motion, while a sha

Load-bearing premise

The whole pipeline rests on the static backbone's canonical geometry being good enough that front/back renders of it yield correct garment masks; if the unobserved back of the body is wrong, the garment layer inherits that error and every later module—seam closing, skinning, dynamics, try-on—is built on it.

Editorial extensions

If this is right

  • Garment reconstruction quality, especially for skirts and dresses, improves over monolithic single-image avatars.
  • The separated garment is an independent asset: it can be re-colored, re-animated, and transferred to a different body in a shared canonical template.
  • 3D virtual try-on becomes a direct asset swap rather than a per-garment optimization or 2D image synthesis.
  • Because the pipeline is feed-forward, inference takes about 10 seconds per avatar on one GPU and renders frames in 65–145 ms.
  • Stronger static backbones can be dropped in with minimal changes, so gains scale with the base reconstruction model.

Reading between the lines

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

  • A natural next step the paper leaves unexplored is feeding the extracted garment asset into a physics simulator to add true cloth dynamics, which would extend the pose-conditioned model beyond learned residuals.
  • The category signal is a single VLM-derived word; a finer taxonomy (e.g., sleeve length, hem asymmetry) could improve continuity for garments that are neither clearly skirt-like nor pants-like.
  • Because garment-membership supervision comes from rendered views of the static avatar, the method's ceiling is tied to that backbone's hallucinated back/side geometry; a test on garments with heavily occluded backs would expose the limit.
  • The darkening-only shading model deliberately separates lighting from intrinsic color; allowing per-garment albedo change would make try-on more flexible but would blur that separation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. Forwardrobe proposes a feed-forward pipeline that, from a single clothed-person image, builds an animatable 3D Gaussian avatar with an explicit garment layer. It initializes from a pretrained static avatar (LHM), predicts per-Gaussian garment membership using image cues and masks rendered from front/back canonical views, uses VLM-derived garment descriptions to choose category-aware seam-closing and skinning re-initialization for skirt-like garments, and adds a pose-conditioned dynamic module predicting geometry, SH appearance, and bounded darkening shading residuals. Training combines image reconstruction, garment-focused losses, and regularization; evaluation is on UBC Fashion and NeuMan against LHM, IDOL, PERSONA, and DynaAvatar, with full-image and garment-region PSNR/SSIM/LPIPS. The authors claim improved garment reconstruction and animation quality, and the first feed-forward reconstruction of an independent, transferable, motion-conditioned 3D Gaussian garment asset supporting editing and cross-avatar virtual try-on.

Significance. If the central claims hold, this is a useful step toward editable single-image avatars: the explicit separation of a garment Gaussian asset, category-aware continuity initialization, and pose-conditioned dynamics are technically coherent and address a real limitation of monolithic avatar reconstructions. The VLM-guided switch between skirt-like and pants-like handling is an interesting use of semantic priors. The paper is clearly written and the experimental protocol is mostly transparent, including held-out test identities and garment-masked metrics. However, no code or data are released, and the main evidence for the central 'independent garment asset' claim is image-space rendering quality rather than any direct 3D measurement of the extracted garment layer. The comparative and ablative claims also lack statistical support, which tempers the strength of the conclusions.

major comments (3)
  1. [Sec. 4.2, Table 2 and Sec. 4.3, Table 3] The garment membership classifier f_label is trained with pseudo-labels obtained by rendering the LHM-initialized canonical avatar from front and back views and applying the same segmentation model. Because the single input image does not constrain back/side geometry, these labels can encode the static backbone's hallucination. A systematic membership error I_g propagates to seam closing (Eq. 7), diffused skinning (Eq. 9), dynamic residuals (Eq. 11), shading, and cross-avatar try-on. Table 2 reports only image-space PSNR/SSIM/LPIPS on the full image and a 2D garment-masked region; misclassified body-colored Gaussians can render correctly from the evaluation viewpoint, so the reported numbers do not verify that the extracted Gaussian set corresponds to the true 3D garment. Please add a direct 3D evaluation, e.g., multi-view or scan-based garment overlap/Chamfer distance for the extracted
  2. [Sec. 4.3, Table 3] All quantitative claims rest on point estimates. Table 2 reports no error bars or significance tests, and Table 3 reports ablations on four NeuMan plus four UBC Fashion subjects with no variance. Observed gaps are often small: e.g., Table 3 SSIM 0.925 vs 0.921 and LPIPS 0.065 vs 0.066; in Table 2, several full-image PSNR differences are within 0.3 dB. Without per-subject breakdowns or paired significance tests, the central claim that Forwardrobe 'improves garment reconstruction and animation quality' is not statistically supported. Please provide error bars, per-subject scatter, or paired tests, and clarify how many independent subjects underlie each number.
  3. [Sec. 2.2, Table 1] The first contribution states that Forwardrobe is 'the first feed-forward framework to reconstruct the observed garment ... as an independent, motion-conditioned 3D Gaussian asset and recompose it with another avatar without per-garment optimization.' However, Sec. 2.2 describes LayerAvatar [23] as learning a 'feed-forward diffusion model for generating component-disentangled Gaussian avatars' with component transfer, and Table 1 omits LayerAvatar entirely. Even if 'generation' versus 'reconstruction of the observed image' is a meaningful distinction, it is not argued, and the priority claim needs qualification. Please include LayerAvatar in the comparison table, add a direct discussion of the exact differences, and, if feasible, an experimental comparison or at least a qualitative positioning on the same examples.
minor comments (5)
  1. [Sec. 4.1] The sentence 'All evaluation identities and frames are held out from training' appears directly after the UBC Fashion protocol description. Since NeuMan is also used for training, please clarify whether the same held-out statement applies to the NeuMan evaluation split and how the official split avoids identity leakage.
  2. [Fig. 4] The caption uses 'LHMPose' as a method label; this appears to be a typo for 'LHM' and should be corrected.
  3. [Sec. 4.4] The discussion of virtual try-on in Sec. 4.4 states that target garment primitives are removed and source garment inserted 'through their shared canonical correspondence.' The procedure for establishing correspondence between different subjects' canonical SMPL-X meshes is not fully specified; please provide details or a reference.
  4. [Sec. 3.4, Eq. (19)] The reliability mask R_t is described verbally but its construction is not fully defined. Since it directly controls which pixels supervise the shading branch, please specify how 'uncertain boundaries, dark regions, and high-gradient textures' are detected.
  5. [General] The paper does not include a limitations section. Given the reliance on hallucinated back geometry and VLM classifications, a short discussion of failure modes would be valuable.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the central garment-asset prediction is trained on external supervision and evaluated on held-out identities; cited self-works are not load-bearing.

full rationale

The paper's core claim—single-image feed-forward reconstruction of an independent, animatable garment Gaussian asset—is not definitionally tied to its inputs. The garment membership classifier (Eqs. 2–4) is trained using external segmentation supervision [27] applied to rendered canonical views; this is a pseudo-labeling strategy, not a case of the target metric being fit by construction. Evaluation is on held-out identities and frames, and the reported garment-region metrics are computed against ground-truth frames, so the central prediction is not a renamed training loss. The static backbone LHM is an external pretrained model used as initialization, not a self-citation; the contributions are the additional labeling, continuity, dynamics, and composition modules. The only self-citations ([4], [24]) appear in related work and are not load-bearing: no uniqueness theorem, no ansatz smuggled via self-citation, and no equation reduces to its own input. The pseudo-label loop identified by the reader is a real data-supervision limitation (if the backbone hallucinates occluded geometry, labels may be wrong), but it is a correctness risk, not circularity, because the final predictions are not equivalent to those labels by construction.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central method rests on the static LHM initialization, SMPL-X canonicalization, VLM category labels, and several hand-set thresholds and loss weights. No new physical entities are introduced; the 'garment asset' is a software representation. The heaviest hidden costs are the unstated reliability of the front/back rendered masks and the lack of reported values for λ and shading bounds.

free parameters (6)
  • garment membership threshold η_g = 0.5
    Chosen threshold in Eq. (4) partitions Gaussians into garment/body; no sensitivity analysis provided.
  • seam-closing blend factor γ_i = unspecified (decreases with distance to seam)
    Eq. (7) redistributes inner-thigh Gaussians toward the mid-plane; exact schedule is not defined and is a hand-designed heuristic.
  • skinning field sample count K = 13
    Eq. (9) samples a continuous skinning field K times per Gaussian; chosen without ablation.
  • residual scales λ_x, λ_SH, λ_w, λ_sh and shading bounds gmin, smin, smax, b_sh = not reported
    Eqs. (12)-(15) and the loss composition use scaling factors that control dynamic offsets and appearance; values are not given in the paper.
  • loss weights λ_lap, λ_edge, λ_temp, λ_spr, λ_geom, λ_anchor = not reported
    Eq. (18) defines six regularization terms with unstated weights; these trade off deformation coherence versus reconstruction quality.
  • Gaussian count and training iterations = 40K primitives, 50K iterations
    Implementation choices in Sec. 4.1 that affect quality but are not justified by ablations.
assumptions (5)
  • domain assumption SMPL-X surface sampling points q_i and LBS skinning weights w^0_i from the static reconstruction model are reliable anchors for garment/body canonical space.
    Sec. 3.2 defines canonical centers as q_i + Δx^0_i on the SMPL-X mesh; if the static model's SMPL-X fit is wrong, the garment layer geometry inherits the error.
  • ad hoc to paper Front/back rendered masks of the canonical avatar are valid multi-view supervision for garment membership.
    Sec. 3.2 states that because the input-view mask can be unreliable in self-occluded regions, canonical views are rendered and segmented with the same model; this assumes the static backbone's unseen geometry is correct.
  • ad hoc to paper VLM-derived category τ correctly classifies the garment as skirt-like versus pants-like.
    Eq. (8) switches continuity initialization on τ; a misclassification would activate or suppress seam closing wrongly.
  • domain assumption Short SMPL-X motion windows M_t encode enough motion cues (velocity/acceleration) to infer non-rigid garment dynamics from a single image.
    Sec. 3.3 uses a motion encoder over a short motion window; single-image training cannot observe true garment dynamics, so priors must supply them.
  • ad hoc to paper Pseudo-target shading S*_t derived from target/base luminance ratio is a valid supervision signal for shading.
    Eq. (19) constructs a darken-only pseudo-ground-truth from target frames; this bakes in the assumption that all appearance change is multiplicative darkening (g_i,t ∈ [g_min,1]).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Forwardrobe: Garment-Aware Gaussian Avatars from a Single Image." pith.science (2026). https://pith.science/paper/FV55DPH6

@misc{pith2026260729106,
  author       = {Pith},
  title        = {Pith review of: Forwardrobe: Garment-Aware Gaussian Avatars from a Single Image},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FV55DPH6}},
  note         = {Machine review of arXiv:2607.29106}
}
read the original abstract

Reconstructing animatable 3D human avatars from a single image remains particularly challenging for loose garments, whose geometry and motion cannot be adequately represented by body-aligned topology and skinning. We present Forwardrobe, a feed-forward framework for reconstructing garment-aware Gaussian avatars from a single image. Forwardrobe explicitly separates clothing from the body in canonical Gaussian space and equips the garment layer with continuity-aware geometry and skinning initialization, pose-conditioned non-rigid deformation, and appearance adaptation. These designs improve garment reconstruction and visual quality during animation, particularly for skirts and dresses. The separated garment layer additionally forms an independently controllable 3D asset, enabling garment editing, transfer, and 3D virtual try-on. Experiments demonstrate improved garment reconstruction quality and greater flexibility in garment manipulation compared with existing single-image avatar reconstruction methods.

Figures

Figures reproduced from arXiv: 2607.29106 by the authors.

Figure 1
Figure 1. Forwardrobe reconstructs garment-aware Gaussian avatars from single-view images, improving the continuity and animation of loose garments. The explicitly separated garment layer forms an independently controllable 3D asset, supporting animation, appearance editing, and 3D virtual try-on across different avatars. ABSTRACT Reconstructing animatable 3D human avatars from a single image remains particularly challenging … view at source ↗
Figure 2
Figure 2. Overview of Forwardrobe. 1) Given a single image and consecutive SMPL-X poses, we extract image features, VLM-derived garment descriptions, and temporal motion features. 2) We construct an independent 3D garment asset through garment extraction, semantics-aware seam closing, skinning initialization, and base-appearance decomposition. The Garment Dynamic Module further predicts motion-dependent geometry, appearance, … view at source ↗
Figure 3
Figure 3. Garment continuity initialization. Top: seam closing followed by diffused skinning-weight reassignment. Bottom: improved spatial continuity under animation. 3.3 Garment Dynamic Module The static garment asset cannot fully capture pose-dependent non-rigid deformation and appearance variation, partic￾ularly for loose garments. We therefore introduce a residual-based Garment Dynamic Module that predicts motion￾conditio… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison under challenging target poses. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Garment composition. Our composition strategy preserves the static body appearance while integrating the dynamically animated garment. 4.3 Ablation Studies We evaluate the key components of Forwardrobe on four NeuMan subjects and four UBC Fashion subjects, as reported …
Figure 6
Figure 6. Figure 6: Applications. Forwardrobe supports garment editing and 3D virtual try-on. The resulting avatars preserve the edited or transferred garments and can be freely animated with arbitrary motion sequences. References [1] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, J…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 4 linked inside Pith

  1. [23]

    Disentangled clothed avatar generation with layered representation

    Weitian Zhang, Yichao Yan, Sijing Wu, Manwen Liao, and Xiaokang Yang. Disentangled clothed avatar generation with layered representation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 11327–11338, 2025

    2025

  2. [1]

    Nerf: Representing scenes as neural radiance fields for view synthesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. InEuropean conference on computer vision, pages 405–421. Springer, 2020

    2020

  3. [2]

    3d gaussian splatting for real-time radiance field rendering.ACM Trans

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, George Drettakis, et al. 3d gaussian splatting for real-time radiance field rendering.ACM Trans. Graph., 42(4):139–1, 2023

    2023

  4. [3]

    Gaussian head avatar: Ultra high-fidelity head avatar via dynamic gaussians

    Yuelang Xu, Benwang Chen, Zhe Li, Hongwen Zhang, Lizhen Wang, Zerong Zheng, and Yebin Liu. Gaussian head avatar: Ultra high-fidelity head avatar via dynamic gaussians. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1931–1941, 2024

    1931

  5. [4]

    Sfdm: Robust decomposition of geometry and reflectance for realistic face rendering from sparse-view images

    Daisheng Jin, Jiangbei Hu, Baixin Xu, Yuxin Dai, Chen Qian, and Ying He. Sfdm: Robust decomposition of geometry and reflectance for realistic face rendering from sparse-view images. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 26409–26419, 2025

    2025

  6. [5]

    Animatable neural radiance fields for modeling dynamic human bodies

    Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. InProceedings of the IEEE/CVF international conference on computer vision, pages 14314–14323, 2021

    2021

  7. [6]

    Arah: Animatable volume rendering of articulated human sdfs

    Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. InEuropean conference on computer vision, pages 1–19. Springer, 2022

    2022

  8. [7]

    Avatarrex: Real-time expressive full-body avatars.ACM Transactions on Graphics (TOG), 42(4):1–19, 2023

    Zerong Zheng, Xiaochen Zhao, Hongwen Zhang, Boning Liu, and Yebin Liu. Avatarrex: Real-time expressive full-body avatars.ACM Transactions on Graphics (TOG), 42(4):1–19, 2023

    2023

Show all 34 references
  1. [8]

    Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling

    Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19711–19722, 2024

    2024

  2. [9]

    Layga: Layered gaussian avatars for animatable clothing transfer

    Siyou Lin, Zhe Li, Zhaoqi Su, Zerong Zheng, Hongwen Zhang, and Yebin Liu. Layga: Layered gaussian avatars for animatable clothing transfer. InACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024

    2024

  3. [10]

    Gaussian wardrobe: Compositional 3d gaussian avatars for free-form virtual try-on

    Zhiyi Chen, Hsuan-I Ho, Tianjian Jiang, Jie Song, Manuel Kaufmann, and Chen Guo. Gaussian wardrobe: Compositional 3d gaussian avatars for free-form virtual try-on. InProceedings of the International Conference on 3D Vision (3DV), 2026

    2026

  4. [11]

    Vid2avatar: 3d avatar reconstruction from videos in the wild via self-supervised scene decomposition

    Chen Guo, Tianjian Jiang, Xu Chen, Jie Song, and Otmar Hilliges. Vid2avatar: 3d avatar reconstruction from videos in the wild via self-supervised scene decomposition. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12858–12868, 2023

    2023

  5. [12]

    Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization

    Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 84–93, 2020

    2020

  6. [13]

    Econ: Explicit clothed humans optimized via normal integration

    Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J Black. Econ: Explicit clothed humans optimized via normal integration. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 512–523, 2023. 10

    2023

  7. [14]

    Lhm: Large animatable human reconstruction model for single image to 3d in seconds

    Lingteng Qiu, Xiaodong Gu, Peihao Li, Qi Zuo, Weichao Shen, Junfei Zhang, Kejie Qiu, Weihao Yuan, Guanying Chen, Zilong Dong, et al. Lhm: Large animatable human reconstruction model for single image to 3d in seconds. InProceedings of the IEEE/CVF International Conference on Co...

    2025

  8. [15]

    Zero-shot reconstruction of animatable 3d avatars with cloth dynamics from a single image

    Joohyun Kwon, Geonhee Sim, and Gyeongsik Moon. Zero-shot reconstruction of animatable 3d avatars with cloth dynamics from a single image. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18182–18192, 2026

    2026

  9. [16]

    D 3-human: Dynamic disentangled digital human from monocular video

    Honghu Chen, Bo Peng, Yunfan Tao, and Juyong Zhang. D 3-human: Dynamic disentangled digital human from monocular video. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 10836–10846, 2025

    2025

  10. [17]

    Ani3dhuman: Photorealistic 3d human animation with self-guided stochastic sampling

    Qi Sun, Can Wang, Jiaxiang Shang, Yingchun Liu, and Jing Liao. Ani3dhuman: Photorealistic 3d human animation with self-guided stochastic sampling. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12651–12662, 2026

    2026

  11. [18]

    Persona: personalized whole-body 3d avatar with pose-driven deformations from a single image

    Geonhee Sim and Gyeongsik Moon. Persona: personalized whole-body 3d avatar with pose-driven deformations from a single image. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 12670–12680, 2025

    2025

  12. [19]

    Multi-garment net: Learning to dress 3d people from images

    Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. InProceedings of the IEEE/CVF international conference on computer vision, pages 5420–5430, 2019

    2019

  13. [20]

    Clothcap: Seamless 4d clothing capture and retargeting.ACM Transactions on Graphics (ToG), 36(4):1–15, 2017

    Gerard Pons-Moll, Sergi Pujades, Sonny Hu, and Michael J Black. Clothcap: Seamless 4d clothing capture and retargeting.ACM Transactions on Graphics (ToG), 36(4):1–15, 2017

    2017

  14. [21]

    Mpmavatar: Learning 3d gaussian avatars with accurate and robust physics-based dynamics.Advances in Neural Information Processing Systems, 38:140008–140034, 2025

    Changmin Lee, Jihyun Lee, and Tae-Kyun Kim. Mpmavatar: Learning 3d gaussian avatars with accurate and robust physics-based dynamics.Advances in Neural Information Processing Systems, 38:140008–140034, 2025

    2025

  15. [22]

    Physavatar: Learning the physics of dressed 3d avatars from visual observations

    Yang Zheng, Qingqing Zhao, Guandao Yang, Wang Yifan, Donglai Xiang, Florian Dubost, Dmitry Lagun, Thabo Beeler, Federico Tombari, Leonidas Guibas, et al. Physavatar: Learning the physics of dressed 3d avatars from visual observations. InEuropean Conference on Computer Vision, ...

    2024

  16. [24]

    Monocloth: Reconstruction and animation of cloth-decoupled human avatars from monocular videos

    Daisheng Jin and Ying He. Monocloth: Reconstruction and animation of cloth-decoupled human avatars from monocular videos. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 5503–5511, 2026

    2026

  17. [25]

    Disco4d: Disentangled 4d human generation and animation from a single image

    Hui En Pang, Shuai Liu, Zhongang Cai, Lei Yang, Tianwei Zhang, and Ziwei Liu. Disco4d: Disentangled 4d human generation and animation from a single image. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 26331–26344, 2025

    2025

  18. [26]

    Expressive body capture: 3d hands, face, and body from a single image

    Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition...

    2019

  19. [27]

    Sapiens2.arXiv preprint arXiv:2604.21681, 2026

    Rawal Khirodkar, He Wen, Julieta Martinez, Yuan Dong, Su Zhaoen, and Shunsuke Saito. Sapiens2.arXiv preprint arXiv:2604.21681, 2026

    2026 arXiv

  20. [28]

    Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193, 2023

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193, 2023

    2023 arXiv

  21. [29]

    Dwnet: Dense warp-based network for pose-guided human video generation.arXiv preprint arXiv:1910.09139, 2019

    Polina Zablotskaia, Aliaksandr Siarohin, Bo Zhao, and Leonid Sigal. Dwnet: Dense warp-based network for pose-guided human video generation.arXiv preprint arXiv:1910.09139, 2019

    1910 arXiv

  22. [30]

    X-avatar: Expressive human avatars

    Kaiyue Shen, Chen Guo, Manuel Kaufmann, Juan Jose Zarate, Julien Valentin, Jie Song, and Otmar Hilliges. X-avatar: Expressive human avatars. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16911–16921, 2023

    2023

  23. [31]

    Neuman: Neural human radiance field from a single video

    Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. InEuropean Conference on Computer Vision, pages 402–418. Springer, 2022

    2022

  24. [32]

    Idol: Instant photorealistic 3d human creation from a single image

    Yiyu Zhuang, Jiaxi Lv, Hao Wen, Qing Shuai, Ailing Zeng, Hao Zhu, Shifeng Chen, Yujiu Yang, Xun Cao, and Wei Liu. Idol: Instant photorealistic 3d human creation from a single image. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2630...

    2025

  25. [33]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 586–595, 2018

    2018

  26. [34]

    Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191, 2024

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191, 2024. 12

    2024 arXiv

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.