acceptodds
Withdrawn from ICLR 2027

ToWo-VTON: Take-Off Wear-On Virtual Try-On via Topology-aware Decoupling

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Virtual Try-On (VTON) aims to synthesize garment transfer without physical trials. Although recent methods perform well on upper-body garments, they often fail in lower-body and full-body scenarios due to complex occlusions and inadequate modeling of body structure. To address this limitation, we introduce ToWo (Take-Off Wear-On), a plug-in two-stage paradigm that decouples body topology reconstruction from garment synthesis. Take-Off first recovers an occlusion-free body representation, and Wear-On then synthesizes the target garment conditioned on this structural prior. This decoupling improves garment placement and robustness, especially in challenging lower-body and full-body settings with large silhouette variations and cross-category transformations. Built on this paradigm, we propose ToWo-VTON, a unified topology-aware framework with structure-aware supervision. We construct a new Body-Fitting Dress-Code (BFDC) dataset via body-aware diffusion and VLM-based filtering to provide topology-consistent supervision across upper-, lower-, and full-body cases, particularly benefiting scenarios with severe occlusions. To demonstrate generality, we instantiate ToWo-VTON on both diffusion-based and MLLM-guided backbones. Diffusion models adopt a plug-in conditioning strategy with BFDC fine-tuning, while MLLM-guided models incorporate Multimodal Alignment Blocks (MABs) for semantic–topology alignment and Garment Boosting Attention (GBA) for detail enhancement. Extensive experiments demonstrate that ToWo-VTON achieves state-of-the-art performance, reducing KID by 26.1% and 51.5% across diffusion-based and MLLM-guided backbones, respectively. It also produces more coherent and perceptually realistic results across diverse garment categories.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Virtual Try-on, Topology-Aware Decoupling, Conditional Image Generation, Diffusion Models, Multimodal Large Language Models

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_HiRpOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes ToWo-VTON, a two-stage framework for virtual try-on that separates body structure reconstruction from garment synthesis. The first stage, Take-Off, employs a Body Structure Revealing (BSR) model to generate a tight-fitting body proxy from the input person image. The second stage, Wear-On, uses this proxy as structural conditioning for target garment generation. To support training with such representations, the authors construct the Body-Fitting DressCode (BFDC) dataset using diffusion-based proxy generation and vision-language model filtering. The framework is implemented on two existing diffusion-based VTON backbones, TED-VITON and BooW-VTON, as well as an MLLM-guided DiT architecture incorporating Multimodal Alignment Blocks (MABs) and Garment Boosting Attention (GBA). The paper evaluates the proposed approach on DressCode and VITON-HD through quantitative comparisons, qualitative results, and ablation studies, reporting improvements across several metrics and garment categories.

Critical strengths
  1. The paper addresses body occlusion and source-garment interference in virtual try-on, particularly for lower-body and full-body garments with substantial silhouette variations. The proposed two-stage formulation provides a straightforward approach to separating body structure estimation from garment synthesis.

  2. The proposed conditioning strategy is evaluated on both mask-based (TED-VITON) and mask-free (BooW-VTON) diffusion architectures. The paper also presents an MLLM-guided DiT implementation, illustrating how the proposed framework can be incorporated into different generative architectures.

  3. The experiments cover several garment categories on DressCode and VITON-HD, with quantitative results suggesting potential benefits in some of the evaluated settings, particularly for lower-body and dress try-on. The paper also includes qualitative comparisons, component ablations, a user study, and computational cost analysis, providing additional perspectives on the behavior and practical characteristics of the proposed approach.

Critical weaknesses
  1. BooW-VTON is used for both BSR training data generation and benchmark augmentation, raising concerns about potential distribution-specific advantages. Evaluation on independent transformations would strengthen the generalization claims.

  2. The contributions of Take-Off inference and DBFDC training are not isolated. A dedicated ablation of MAB is also missing.

  3. The reported metrics primarily measure image quality rather than body topology consistency. Quantitative analysis under different occlusion levels and garment transformations would better support the central claims.

  4. The quality of generated body proxies and VLM-based filtering is not fully characterized. The inconsistent valid/invalid proxy totals in Table 8 also require clarification.

Questions for authors
  1. Since BooW-VTON is used for both BSR training data generation and benchmark augmentation, have the authors evaluated the framework on independently generated or real-world garment transformations?

  2. Could the authors provide separate ablations for Take-Off inference and DBFDC training to clarify their individual contributions?

  3. MAB is introduced as a key component of the MLLM-guided architecture. Could the authors provide a dedicated ablation to evaluate its effectiveness?

  4. Could the authors include topology-specific metrics, such as limb consistency or pose alignment, to more directly support the claimed structural improvements?

  5. How does the performance vary with different levels of garment occlusion and silhouette changes? Are the improvements more pronounced under larger transformations?

  6. How does the proposed body proxy differ from existing body or limb reconstruction approaches? Could the authors clarify the advantages of using a complete tight-fitting proxy?

  7. What are the sample retention rates of DBFDC after VLM filtering? Has the quality of the generated proxies been independently assessed?

  8. How sensitive is the framework to the selection of tight-fitting garment templates during Take-Off inference?

  9. Could the authors clarify how the 40 image sets and participants were selected for the user study, and whether the comparisons were randomized and conducted blindly?

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

This paper addresses source-garment interference in virtual try-on, particularly for lower-body and full-body clothing with substantial silhouette changes. Its two-stage Take-Off/Wear-On pipeline first generates a tight-fitting body proxy using a Body Structure Revealing model, then synthesizes the target garment conditioned on that proxy. The authors construct the Bod˚y-Fitting DressCode (BFDC) dataset and adapt the framework to TED-VITON, BooW-VTON, and an OmniGen2-based model with multimodal alignment and garment-detail attention. Experiments show improved visual quality and garment–body alignment across DressCode and VITON-HD.

Critical strengths
  1. Improvements on both mask-based and mask-free diffusion models suggest that the body-proxy strategy is useful beyond the proposed MLLM-based implementation.

  2. On DressCode lower-body evaluation, the full model achieves a KID of 0.172 versus 0.509 for BooW-VTON, with qualitative examples supporting better limb structure and garment boundaries.

  3. The appendix evaluates proxy-failure detection, fallback inference, and oracle routing rather than discussing error propagation only qualitatively.

Critical weaknesses
  1. Most importantly, the motivation for the Take-Off stage is insufficiently justified. The paper attributes failures under substantial garment silhouette changes to inadequate body-topology modeling, but the presented examples do not clearly establish this diagnosis. In the left example of Figure 1, the person’s overall leg configuration is already discernible, yet the generated shorts have an incorrect silhouette. A plausible alternative explanation is interference from the source garment’s shape or, for mask-based methods, shape information leaked through the agnostic mask. Related mask-induced garment-shape leakage has been discussed in AnyFit (https://arxiv.org/abs/2405.18172), which addresses it through parsing-free mask augmentation. A more direct baseline would construct training triplets comprising a synthetically re-dressed person, the ground-truth person image, and the target garment, deliberately introducing large silhouette differences between the input and target clothing. Training the same single-stage backbone on these triplets would test whether stronger supervision alone resolves the observed failures. Without this comparison, the paper does not convincingly establish the need for an additional body-reconstruction stage or distinguish its benefits from improved training coverage and reduced source-garment bias.

  2. The sources of improvement are not sufficiently isolated. The w/o ToWo configuration removes both BFDC training and Take-Off inference. Consequently, the experiments cannot distinguish the benefits of inference-time body proxies from those of altered training data and additional fine-tuning. A budget-matched, factorial ablation of BFDC training and Take-Off inference is needed, together with an independent evaluation of the Multimodal Alignment Blocks.

  3. Body-proxy reliability remains insufficiently established. Image-quality metrics and attention visualizations do not directly measure pose, limb connectivity, or body-shape preservation. The default detector identifies only 54.0% of invalid proxies. The fallback experiments help, but do not establish robustness under severe occlusion.

  4. The retained BFDC size after filtering, the initial BSR training-set size, and the relationship between template counts in the main text and appendix should be clarified.

Questions for authors

Please refer to the weakness 1 2 and 4.

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

This paper addresses source-garment interference in virtual try-on, particularly when the source and target garments have substantially different silhouettes. It proposes ToWo-VTON, a two-stage framework in which Take-Off generates a body-proxy image depicting the person in tight-fitting clothing, and Wear-On synthesizes the target garment conditioned on this proxy. The authors construct the BFDC dataset and apply the approach to both diffusion-based and MLLM-guided generative models. Experiments on VITON-HD and DressCode report improvements in garment fit, limb geometry, and visual quality, particularly for lower-body garments and dresses involving substantial silhouette changes.

Critical strengths
  1. Practical relevance. The paper addresses interference from source garments under large changes in garment silhouettes. Using a body proxy depicting the person in tight-fitting clothing is well motivated and intuitive.

  2. Applicability across backbones. Improvements are reported for TED-VITON, BooW-VTON, and an MLLM-guided generative backbone, suggesting that the approach is applicable across different architectures.

  3. Broad evaluation coverage. The evaluation includes multiple garment categories, user studies, computational costs, and failure detection with fallback. Qualitative examples show more plausible garment lengths and more coherent limb structures.

Critical weaknesses
  1. Missing comparisons with recent relevant baselines. The experiments do not include RefTon [1], OmniVTON [2], or Any2AnyTryon [3]. Consequently, the claim of state-of-the-art performance is not sufficiently supported.

  2. Unclear result provenance and evaluation protocol. Only TED-VITON and BooW-VTON are explicitly marked as reproduced. Several other results differ from those in the original publications, but the corresponding evaluation configurations are not specified. It is also unclear which metrics in Table 2 are evaluated under paired or unpaired settings.

  3. Potential inconsistencies in Table 2. Qwen-Image and StableVITON have identical KID, FID, and SSIM scores on DressCode-upper, while their LPIPS scores differ by only 0.001. These near-identical results require verification. The caption stating that “FID is scaled by 100” also needs clarification.

  4. Ablations do not isolate individual contributions. The “w/o ToWo” variant removes both BFDC training and Take-Off inference, making it difficult to distinguish the benefits of the training data, additional optimization, and body-proxy conditioning. A separate ablation of the MAB module is also missing.

[1] Liuzhuozheng Li et al. RefTon: Reference person shot assist virtual Try-on. CVPR, 2026.

[2] Zhaotong Yang et al. OmniVTON: Training-Free Universal Virtual Try-On. ICCV, 2025, pp. 16702–16711.

[3] Hailong Guo et al. Any2AnyTryon: Leveraging Adaptive Position Embeddings for Versatile Virtual Clothing Tasks. ICCV, 2025, pp. 19085–19096.

Questions for authors
  1. Could you include comparisons with the methods listed above under matched input conditions and a common evaluation protocol?

  2. Are the results in each row of Table 2 taken from the original publications or obtained through your own evaluations? Please specify the checkpoints, input construction, evaluation resolution, metric implementations, and paired/unpaired settings.

  3. Could you verify the near-identical results for Qwen-Image and StableVITON and clarify whether the caption stating that FID is scaled by 100 is correct?

  4. Under controlled training data and matched training budgets, what are the separate contributions of BFDC, Take-Off, and MAB? Could you also evaluate Direct Wear-On on the full test set as a controlled comparison?

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

Related papers

Discussion (0)

Sign in to comment.