acceptodds
Withdrawn from ICLR 2027

WearAnything: Dress 3D Humans in Seconds via Decoupled Gaussian Garment Transfer

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We present WearAnything, a feed-forward framework for direct 3D garment transfer in Gaussian human representations, enabling efficient and controllable virtual try-on without multi-stage reconstruction or subject-specific optimization. Existing 3D virtual try-on methods are computationally expensive and often struggle with cross-category transfer involving substantial changes in garment coverage. WearAnything performs garment transfer directly in the canonical token representation of a pretrained 3D human model. The framework learns separate identity and garment latent spaces from dense canonical features, while a lightweight routing module adapts the editable region to the target garment support. Together, these components enable the recomposition of source identities and target garments across clothing categories, including transfers with substantial changes in garment coverage. To maintain 3D structure during garment editing, we introduce a 3D distillation strategy that transfers canonical feature and Gaussian geometry supervision from the frozen reconstruction model to the editable representation. The learned garment space further supports continuous outfit interpolation. Experiments demonstrate state-of-the-art 3D garment transfer quality and fast end-to-end inference within a few seconds.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Garment transfer, 3D vision, Human modeling

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_1uCzOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a feed-forward framework for garment transfer in 3D Gaussian human representations. It reconstructs avatars with LHM++ and decouples identity from garment. Specifically, it operates on the canonical token representation of the reconstructor and introduces a lightweight identity and garment decoupling design. Because a garment may change its spatial coverage during transfer, the paper learns a garment latent space that captures garment appearance, geometry and coverage. It further introduces a decoupling training strategy to strengthen identity and garment separation while preserving 3D consistency.

Contributions

A lightweight identity and garment decoupling module with routing-guided composition.

A decoupling training strategy with 3D feature and geometry distillation.

Compared to existing methods, the method achieves more accurate garment transfer.

The method is significantly more efficient, as it requires neither per-avatar optimization nor a complex multi-stage process.

Critical strengths

Compared to existing methods, the method achieves more accurate garment transfer.

The method is significantly more efficient, as it requires neither per-avatar optimization nor a complex multi-stage process.

Critical weaknesses

Blurry results. The results are blurry, which partially stems from LHM++.

The virtual try-on claim is overstated. Although the method allows transfer across garment categories, all garments remain tightly rigged to the avatar. The term "garment transfer" is therefore reasonable, but "virtual try-on" is an overclaim.

Result quality. Slight Gaussian floating artifacts remain. In addition, I am surprised that the existing methods perform so poorly in the comparisons, failing even to guarantee view consistency.

The comparison is not fully fair. M3D-VTON and GS-VTON take a reference garment image, rather than an image of a person wearing the garment, as the garment input. The artifacts in their results may stem from this mismatch.

The ablation settings are unclear. For example, how is the identity and garment decoupling removed?

Exposition. Some terms are introduced without explanation, such as the "projection-validity indicator" and the "garment prior".

Questions for authors

How exactly is each ablation variant implemented?

How were the garment inputs provided to M3D-VTON and GS-VTON, which expect a reference garment image rather than an image of a person wearing the garment?

Reviewer_9JY7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

WearAnything is a feed-forward framework for 3D garment transfer that separates identity and garment representations within the canonical token space of a frozen human reconstructor. It combines source identity and reference garment features through spatial routing, trained with token and geometry distillation alongside rendering supervision. The reported results improve on the evaluated baselines, with approximately 5–6-second inference on an RTX 3090.

Critical strengths

Flexible Garment Transfer. The framework provides an efficient approach to 3D garment transfer, accommodating changes in garment coverage and supporting continuous outfit interpolation.

Broad Baseline Comparison. The evaluation includes both 3D methods and 2D try-on pipelines followed by reconstruction, with end-to-end runtime reported.

Targeted Ablation Studies. The ablation studies evaluate the contributions of identity/garment decoupling, garment routing, and key training components, providing evidence for the effectiveness of these design choices.

Critical weaknesses

Limited novelty of the method. Existing works such as LayGA have already explored body-garment separation and compositional 3D garment transfer. While the feed-forward formulation improves efficiency, the proposed identity/garment encoders, mask-based routing, and feature composition appear to be relatively straightforward adaptations within the pretrained canonical representation.

Cross-identity preservation is not fully established. The architecture and identity-consistency loss encourage separation, but same-identity training alone cannot establish whether transfer preserves the source person’s attributes when source and reference identities differ. The appendix provides useful cross-identity examples and ArcFace results, but body-shape and non-facial appearance preservation remain less thoroughly evaluated. This limits the strength of the disentanglement claim.

3D Geometry Evaluation. The paper claims that 3D Distillation preserves garment geometry and multi-view consistency, but the evaluation mainly relies on image-based metrics. Moreover, supervision comes from a frozen LHM++ model rather than ground-truth 3D geometry. Direct geometric evaluations would better establish whether the proposed distillation improves actual reconstruction accuracy.

Questions for authors

Methodological Novelty. Given the shared canonical representation provided by the pretrained LHM++, the proposed identity/garment decoupling, mask-based routing, and feature composition appear relatively straightforward. Could the authors clarify the key methodological novelty beyond adapting existing garment separation and feature composition techniques to the canonical token space?

Cross-Identity Evaluation. The main quantitative experiments use same-identity, different-clothing pairs, while the cross-identity evaluation primarily measures identity preservation. Could the authors provide additional evaluations of garment fidelity and source body-shape preservation when transferring clothing between different identities, without requiring paired ground-truth try-on images?

3D Geometric Validation. Although the paper introduces 3D distillation to preserve garment geometry, the quantitative evaluation primarily relies on image-based metrics. Could the authors provide more direct evaluations of 3D geometric accuracy, preferably using measurements independent of the frozen LHM++ teacher, to validate the benefits of 3D distillation?

Reviewer_wFAjOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper builds a 3D model of a person wearing an outfit taken from a second photo in about five seconds. It does this by editing the internal features of a pretrained photo-to-3D human network.

  • Input: A full-body source image, a reference image of a person wearing the target outfit, and garment masks from a human parser.
  • Output: An animatable 3D Gaussian human in the LHM++ format, with renders in any pose and view.
  • Method: The frozen LHM++ maps each image to point-aligned canonical tokens. A routing head predicts which points are garment, conditioned on the projected garment mask. Two encoders compress the tokens into a small identity code and a larger garment code. Decoders produce a person feature field from the source identity and a garment feature field from the reference garment code, conditioned on the source identity. These fields are blended with the reference garment mask and decoded by the frozen LHM++ into Gaussians and images. Training uses pairs of the same person in two outfits. Token, Gaussian, image and mask losses compare the output with LHM++ applied to the second image.
  • Delta: LHM++ provides the representation and decoders. The new parts are the mask-conditioned routing in canonical space, the asymmetric identity and garment bottlenecks and the region-weighted distillation from the frozen reconstructor.
  • Experimental validation: The paper reports PSNR, SSIM and LPIPS on 50 same-identity pairs each from StreetTryOn and MVHumanNet. The method reaches 23.65 dB against 20.65 dB for OneModelForAll with LHM++, in 5.4 s against 89 s. Cross-identity transfer is evaluated with ArcFace similarity only. An ablation removes decoupling, routing, the 3D losses, the 2D losses and gradient isolation.
Critical strengths
  • Editing the canonical tokens of a feed-forward reconstructor gives simple, fast 3D try-on. Because LHM++ maps every person to the same ordered point set, garment transfer reduces to recombining per-point features. The module has only 4.12 M parameters and runs in 5.4 s, against 29 s to 1,833 s for the baselines (Table 1).

  • The method changes garment coverage across categories. Fig. 3, Fig. 5 and the supplementary turntables show transfers between trousers, shorts, skirts and dresses with plausible coverage from all sides. The per-subject and inpainting-based baselines tend to keep the source garment extent.

Critical weaknesses

There are many signs of this being a late last minute submission. This has occupied a lot of my time. This paper should not have been submitted and is clearly not ready for submission.

  • A citation is broken. Appendix A.3 shows "InsightFace ?".

  • Key baselines are missing, and the video omits the strongest ones. For static renders, the natural two-stage baseline is a 2D try-on followed by a strong static reconstructor such as SyncHuman [1]. A dedicated try-on model such as IDM-VTON [2] with LHM++ is also missing. The video compares only with M3D-VTON, GS-VTON and VTON360, and shows no baseline animation. VTON360 collapses to noise in most turntable views, far below its published results, which suggests a failed baseline run.

  • The renders are blurry despite competitive PSNR. In terms of PSNR the results look acceptable, but the visual quality is well below recent avatar methods. Even LHM++, the reconstructor that the method builds on, produces sharper results. Non-garment regions are regenerated from a 64-token identity code (Eq. 6 and 7), which may contribute to the loss of detail. PSNR and SSIM favour smooth outputs, so they do not capture this problem.

  • The quantitative protocol measures reconstruction of the reference person, not garment transfer. Each Table 1 pair shows one person in two outfits, and the output is compared with the reference image (Section 4.1). For these pairs, the ideal output equals the LHM++ reconstruction of the reference image, which is the training target. A method that ignores the source would therefore score near the top, and the paper does not report it. The two-stage baselines instead must re-synthesise the person and are penalised for any difference.

  • The evaluation does not test the claimed capabilities. Animation is the main reason to build on LHM++, yet no result evaluates animated or novel-pose renders. Dress and skirt transfers are never shown in motion, where skinning-based deformation is most likely to fail. Geometric consistency is claimed but never measured, since all numbers are image metrics on renders. Cross-identity transfer is measured only by ArcFace similarity (Table 3), with no baselines and no garment-fidelity metric, and the method scores below plain LHM++.

Additional Comments

  • The supplementary video lacks polish. It appears assembled at the last minute, with no audio and little explanation of the comparisons. For an ICLR submission I expect a more careful presentation, especially since high-quality animations are now cheap to produce.
  • Several citations are wrong or broken. OneModelForAll (Cheng et al., 2025) is an image-fusion paper, while the try-on model is OMFA [3]. Many citations also lack parentheses, for example "LayGA Lin et al. (2024)".
  • Related work omits earlier 3D garment transfer. StructLDM [4] performs 3D try-on by editing a structured latent space, and SCARF [5] and GALA [6] transfer garment layers between avatars. Forwardrobe (Jin et al., 2026, cited in the paper) also reports feed-forward garment transfer and needs a precise comparison.
  • The evaluation sets are small and selected. Each benchmark uses 50 pairs, kept after manual inspection (Appendix A.3), and no table reports variance.
  • Table 2 does not state its evaluation pairs. Its ArcFace value (0.7112) equals the cross-identity value in Table 3.
  • The interpolation needs a routing mask. Please state which mask Eq. 9 uses for intermediate values of α\alpha.

References

[1] Wenyue Chen, Peng Li, Wangguandong Zheng, Chengfeng Zhao, Mengfei Li, Yaolong Zhu, Zhiyang Dou, Ronggang Wang, Yuan Liu. "SyncHuman: Synchronizing 2D and 3D Generative Models for Single-view Human Reconstruction." NeurIPS 2025. arXiv:2510.07723.

[2] Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, Jinwoo Shin. "Improving Diffusion Models for Authentic Virtual Try-on in the Wild." ECCV 2024. arXiv:2403.05139.

[3] Jinxi Liu, Zijian He, Guangrun Wang, Guanbin Li, Liang Lin. "One Model for All: Unified Try-On and Try-Off in Any Pose via LLM-Inspired Bidirectional Tweedie Diffusion." arXiv 2025. arXiv:2508.04559.

[4] Tao Hu, Fangzhou Hong, Ziwei Liu. "StructLDM: Structured Latent Diffusion for 3D Human Generation." ECCV 2024. arXiv:2404.01241.

[5] Yao Feng, Jinlong Yang, Marc Pollefeys, Michael J. Black, Timo Bolkart. "Capturing and Animation of Body and Clothing from Monocular Video." SIGGRAPH Asia 2022. arXiv:2210.01868.

[6] Taeksoo Kim, Byungjun Kim, Shunsuke Saito, Hanbyul Joo. "GALA: Generating Animatable Layered Assets from a Single Scan." CVPR 2024. arXiv:2401.12979.

Questions for authors

I did not like the presentation of the paper and it is unlikely that I will change my rating. The paper took a lot of my time and is clearly not ready for peer-review.

  • Report LHM++ reconstruction of the reference image alone in Table 1.
  • Report a 2D try-on such as IDM-VTON [2] followed by SyncHuman [1] and by LHM++.
  • Evaluate animated and novel-pose renders, including dress and skirt transfers.
  • For cross-identity pairs, report garment fidelity and ArcFace similarity for all methods.
  • Add OneModelForAll and OOTDiffusion with LHM++ to the video, including animations.
  • Check the VTON360 runs, whose turntables collapse to noise.
  • Report a variant that keeps the source tokens outside both garment masks.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.