acceptodds
Under review as a conference paper at ICLR 2027

Seeing More Details: Multi-Reference Virtual Try-On with Fine-Grained Garment Representation

Abstract

Virtual try-on (VTON) has achieved remarkable progress with the advancement of diffusion models. However, most existing methods focus on single-image guidance under in-shop settings, often producing unsatisfactory results when handling large pose variations, complex occlusions, or fine-grained texture transfer. To address this limitation, we propose a mask-free VTON method, termed MR-VTON, which leverages Multi-Reference garment images captured from diverse viewpoints and wearing poses to preserve the garment details and realistic try-on appearance, particularly in in-the-wild scenarios. The proposed model employs a dual-branch architecture to dynamically adapt reference image selection and fine-grained feature aggregation. Specifically, the first branch introduces a router-based garment selection mechanism that adaptively selects a single image as the primary reference to anchor garment identity features, with minimal occlusion and maximum clothing information. The second branch introduces a target pose-guided feature aggregation module that integrates garment details and viewpoint information from multiple reference images, thereby enhancing spatial consistency and fine-grained detail preservation. The features from both branches are concatenated and injected into a Diffusion Transformer architecture as reference features, facilitating VTON across diverse viewpoints and poses. Furthermore, we construct an evaluation benchmark for in-the-wild multi-reference VTON, named MRG-4K, to fill the gap in this emerging task. Extensive experiments on the proposed benchmark and zero-shot academic benchmarks demonstrate that our method consistently outperforms existing approaches and achieves state-of-the-art performance in both single-reference and multi-reference try-on tasks.

Primary area:
generative models
Keywords:
VTON, Diffusion

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_vCmeOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MR-VTON, a mask-free multi-reference virtual try-on method for in-the-wild scenarios built on Qwen-Image-Edit, which combines adaptive main-reference garment selection (token-level attention pooling with inter-image gating, soft routing during training and hard top-1 selection at inference) and target-pose-guided feature aggregation (multi-layer cross-attention with target pose queries, reference pose keys, and reference content values, augmented by segment embeddings and 2D RoPE), with both branch outputs concatenated and injected into a diffusion transformer. The authors also introduce MRG-4K, an in-the-wild multi-reference VTON benchmark.

Critical strengths

It addresses a practically relevant but underexplored setting. The focus on in-the-wild multi-reference garment images is more realistic than the fixed-view, in-shop setting used by many earlier VTON methods.

Critical weaknesses

1) Insufficient validation of MR-VTON's effectiveness and its generalization to real-world multi-reference datasets. Since the benchmark uses synthetically generated try-on edited images as inputs and video frames as ground truth, MR-VTON's high performance may partly reflect its adaptation to artifacts or stylistic biases introduced by the editing model (such as boundary blur, residual textures, or color biases), rather than genuine virtual try-on capability. This may limit its generalization to real-world inputs with multiple reference garment images.

2) The validation and analysis of whether MRG-4K can truly be used for the multi-reference try-on task are insufficient. Following from the above issue, could the artifacts or stylistic biases introduced by the editing model (such as boundary blur, residual textures, or color biases) be mistakenly treated by try-on methods as garment edge masks? Can such edited MRG-4K truly be used for multi-reference try-on methods? The authors should conduct further validation and analysis on this matter.

Questions for authors
  1. Insufficient validation of MR-VTON's effectiveness and its generalization to real-world multi-reference datasets. The proposed benchmark relies on synthetically generated try-on edits as inputs and video frames as ground truth. Consequently, its high performance may not fully reflect the model's genuine virtual try-on capability, but could partly result from adapting to artifacts or stylistic biases introduced by the image-editing model used to construct the benchmark. For example, if all test inputs are generated by the same editing model, they may contain consistent boundary blur, residual textures, or color biases. MR-VTON may learn to correct these specific artifacts and perform well on the proposed benchmark, while generalizing less effectively to real images containing multiple reference garment images. The authors should evaluate MR-VTON on additional real-image benchmarks to determine whether the reported improvements persist beyond this synthetically constructed benchmark. Datasets such as DeepFashion, MPV, and Eevee may provide useful complementary evaluation settings. In particular, Eevee appears especially relevant because it provides a setting closer to realistic multi-reference try-on. The authors should clarify how each dataset is adapted to the multi-reference setting and report results on suitable external benchmarks.

2) In Table 2 (Quantitative Comparison), the only compared method with multi-reference input is MV-VTON. To my knowledge, several recent works take multiple reference garments as input for try-on, such as Robust-MVTON (CVPR 2025) and VC-VTON (TIP 2026).

3) Figure 2 should include more technical details to provide a clearer understanding of MR-VTON.

3 comments

W1 & W2 & Q1:

Thank you for your insightful comment. We acknowledge the concern that synthetic editing artifacts may introduce unintended visual cues, potentially affecting the validity of MRG-4K as a multi-reference VTON benchmark. We would like to clarify the following points:

(1) Benchmark Construction and Synthetic Artifact Mitigation. We acknowledge that synthetically edited images may introduce artifacts or stylistic biases. To mitigate this issue, we apply a rigorous data cleaning process during training data construction to filter out low-quality edited images and samples with noticeable artifacts. Moreover, the ground-truth images are original real-world video frames rather than synthetic outputs, ensuring that the model learns to reconstruct realistic garment appearances and textures. More importantly, our evaluations on VITON-HD, DressCode, and StreetTryOn are all conducted in a zero-shot manner using real person images as inputs, without any dataset-specific fine-tuning or synthetically edited test inputs. The competitive results on these established benchmarks provide evidence that our model generalizes beyond synthetic editing artifacts and can effectively handle real-world virtual try-on scenarios.

(2) Real-World Multi-Reference Evaluation on Eevee. To further validate the generalization ability of MR-VTON, we conduct additional zero-shot experiments on the Eevee dataset without any dataset-specific fine-tuning. Specifically, within each garment category (e.g., dresses, lower-body, and upper-body garments), we select multiple reference images from the same video frame and a different video frame as the target person image. This produces a multi-reference evaluation setting consisting of multiple reference garment images and a real target person image, without relying on synthetically edited inputs. Following this protocol, we construct 1,000 evaluation samples, including 250 dresses, 250 lower-body garments, and 500 upper-body garments.

As shown in Table R1 below, our method achieves the lowest FID of 21.12, outperforming all eight competing methods, including specialized VTON approaches and recent general-purpose image editing models. In particular, MR-VTON outperforms Nano Banana Pro (22.15), GPT-Image-2 (23.47), and FireRed-Image-Edit (24.37). These results demonstrate that the effectiveness of MR-VTON extends beyond the synthetically constructed MRG-4K benchmark, providing further evidence of its robustness and generalization capability in real-world multi-reference scenarios.

Table R1: Experiments on the Eevee Dataset.

MethodFID ↓
OOTDiffusion34.94
CatVTON31.23
MV-VTON32.75
FitDiT29.42
Qwen-Image-Edit-251125.88
FireRed-Image-Edit-1.124.37
Nano Banana Pro22.15
GPT-Image-223.47
Ours (MR-VTON)21.12

Q2:

Thank you for your valuable suggestion. We agree that comparisons with recent multi-reference VTON methods are important for a comprehensive evaluation. We would like to clarify the following points:

(1) Availability of Existing Methods. We acknowledge the relevance of Robust-MVTON (CVPR 2025) [1] and VC-VTON (TIP 2026) [2]. However, to the best of our knowledge, their official implementations and pretrained checkpoints are not publicly available, making it difficult to reproduce their results under our MRG-4K evaluation protocol.

(2) Additional Comparisons with Recent Methods. To address this concern, we have conducted additional experiments with three recent VTON methods, RefTon [3], OmniVTON [4], and Any2AnyTryOn [5], on our MRG-4K benchmark. The quantitative results are presented in Table R2 below. Our method achieves the best performance on four out of five metrics. In particular, compared with RefTon, MR-VTON reduces FID from 25.28 to 21.79 (13.81%) and KID from 22.79 to 16.47 (27.73%), while improving DiffSim from 0.7713 to 0.8014. These results further demonstrate the effectiveness of our proposed approach compared with recent VTON methods.

Table R2: Comparison of Our Method with Other State-of-the-Art Methods.

MethodFID ↓KID ↓LPIPS ↓PSNR ↑DiffSim ↑
Any2AnyTryOn27.5437.840.213317.800.7709
OmniVTON27.9024.810.241317.090.7558
RefTon25.2822.790.223817.860.7713
Ours (MR-VTON)21.7916.470.221318.140.8014

Q3:

Thank you for your valuable suggestion. In our original submission, Figure 2 was primarily intended to illustrate the core architecture of MR-VTON. To provide a clearer understanding of the technical details, we have uploaded a revised architecture diagram in the revision, with the following three improvements:

(1) Routing Strategy. We explicitly illustrate the different routing strategies adopted during training and inference. Specifically, soft routing is employed during training to enable continuous gradient propagation, while hard Top-1 routing is adopted during inference to select the most informative reference.

(2) Segment Embedding. We clarify that the garment content features and pose features extracted from multiple reference images are augmented with segment embeddings after encoding. These embeddings enable the model to distinguish features originating from different reference images, facilitating effective multi-reference feature aggregation.

(3) Feature Integration. We further clarify how the outputs of the two complementary branches are integrated. Specifically, the features produced by the Adaptive Main Reference Selector and the Target Pose-Guided Feature Aggregation module are concatenated and subsequently fed into the downstream modules for virtual try-on generation.

These revisions provide a more detailed illustration of the internal mechanisms and feature interactions, making the overall MR-VTON framework easier to understand.

References

[1] Zhang N, Li Y, Du D, et al. Robust-MVTON: Learning Cross-Pose Feature Alignment and Fusion for Robust Multi-View Virtual Try-On. CVPR, 2025.

[2] Ma, Zhiyuan, et al. VC-VTON: Towards Across-View and Multi-Posture Driven Virtual Try-On via Spatiotemporal-Aware View Consistency Training. TIP, 2026.

[3] Liuzhuozheng Li et al. RefTon: Reference Person Shot Assist Virtual Try-on. CVPR, 2026.

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

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

We sincerely thank you for your time and effort in reviewing our paper. Your constructive comments have been very helpful in improving the quality of our work.

As the author-reviewer discussion period will end in approximately one day, we would greatly appreciate it if you could take a moment to review our rebuttal and additional experimental results, and let us know whether our responses have adequately addressed your concerns.

Thank you again for your time and consideration.

Reviewer_y2cmOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MR-VTON, a multi-reference virtual try-on method that integrates multiple images of the same garment through primary reference selection and target pose-guided feature aggregation. The method builds on Qwen-Image-Edit to generate try-on results. The authors construct the MRG dataset and evaluate garment-detail preservation and generalization on their proposed benchmark and established public benchmarks.

Critical strengths
  1. Practical relevance. Using multiple references to address occlusion, viewpoint variation, and missing garment details is well motivated. The dataset also offers potential value to the community.

  2. Experimental support for the overall design. Comparisons with independently trained single-reference and direct-concatenation baselines, together with dual-branch ablations, provide evidence for the overall design. Cross-dataset evaluation and a user study provide additional support.

Critical weaknesses
  1. Insufficient comparison with directly related work. The motivation for reference-guided virtual try-on and the use of synthetic training pairs are related to ideas explored in RefTon[1], yet no direct experimental comparison is provided. This makes it difficult to assess the incremental contributions of the method and dataset. Recent relevant baselines, including OmniVTON[2] and Any2AnyTryon[3], are also missing from the comparisons.

  2. Unequal input information in the main comparisons. Some baselines receive only one reference, whereas MR-VTON uses multiple references. The reported gains may therefore reflect both additional reference information and algorithmic improvements. Although Table 4 includes a multi-reference direct-concatenation baseline, the main results should also include comparisons under matched reference-input conditions.

  3. Unexplained discrepancy in the CatVTON results. Table 3 reports a VITON-HD unpaired FID of 21.96 for CatVTON, substantially higher than the 9.015 reported in its original paper. The checkpoint and evaluation configuration are not fully specified, limiting confidence in the claimed relative advantage.

  4. Unclear handling of missing reference poses. The paper presents results using flat-lay garment references, which do not contain a person, but does not explain how the pose-guided aggregation branch operates without a reference human pose. Targeted ablations isolating the contribution of pose correspondence are also missing.

Questions for authors
  1. Please report quantitative comparisons separately under single-reference and multi-reference settings, specifying the number of references, reference sets, and evaluation protocol used for each method.

  2. Please compare directly with RefTon[1] under input configurations supported by both methods and clarify the incremental contributions of the reference aggregation mechanism and dataset relative to prior work.

  3. Please explain why Table 3 reports a VITON-HD unpaired FID of 21.96 for CatVTON, compared with 9.015 in the original paper. Please specify the checkpoint, garment–person pairing, mask preprocessing, inference and evaluation resolutions, and metric implementations.

  4. The paper presents results using flat-lay garment references, for which no reference human pose is available. How does the pose-guided aggregation branch operate in this setting?

[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.

8 comments

W1 & Q2:

Thank you for your valuable comments. We acknowledge that RefTon [1] also explores reference-guided virtual try-on and synthetic training data construction. However, our work differs in two key aspects:

(1) Explicit Multi-Reference Aggregation. RefTon incorporates additional visual references through conditional image token concatenation, whereas our method explicitly models the complementary and potentially conflicting information across multiple references. Specifically, the Adaptive Main Reference Selector identifies the most informative reference to preserve global garment identity, while the Target Pose-Guided Feature Aggregation module retrieves complementary local garment details from different references. This design enables more effective exploitation of multi-view garment information.

(2) Multi-Reference Benchmark Construction. RefTon constructs its VFR dataset by extending existing VTON benchmarks with additional reference person images synthesized using Flux-Kontext, where the garment appearance is preserved while the person's identity and pose are modified. In contrast, our MRG-4K benchmark is constructed from real-world e-commerce garment images and video frames, rather than relying on synthetically generated reference images. Specifically, MRG-4K contains 1,296 test samples, each comprising 2–5 garment reference images with diverse viewpoints, poses, and occlusions. This design provides a dedicated benchmark for evaluating how effectively multi-reference VTON methods exploit complementary garment information to preserve garment identity and recover fine-grained details under realistic conditions.

We have added direct comparisons with RefTon, OmniVTON [2], and Any2AnyTryOn [3] on our MRG-4K benchmark, as shown in Table R1 below.

Table R1: Comparison of Our Method with Other State-of-the-Art Methods.

MethodFID ↓KID ↓LPIPS ↓PSNR ↑DiffSim ↑
Any2AnyTryOn27.5437.840.213317.800.7709
OmniVTON27.9024.810.241317.090.7558
RefTon25.2822.790.223817.860.7713
Ours21.7916.470.221318.140.8014

Our method achieves the best performance on four out of five metrics. Compared with RefTon, our method reduces FID by 13.81 and KID by 27.73, while improving DiffSim from 0.7713 to 0.8014.

References

[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.

W2 & Q1:

Thank you for your valuable comment. To address the concern regarding unequal reference inputs, we have reorganized our quantitative comparisons into single-reference and multi-reference settings, as shown in Table R2 below. We would like to clarify the following points:

(1) Single-Reference Comparisons on MRG-4K. To ensure a fair comparison under matched input conditions, we additionally train and evaluate a single-reference variant of our method, denoted as Ours (single-reference), using only one garment reference image during both training and inference. As shown in Table R2, this variant achieves an FID of 25.45, KID of 22.80, and PSNR of 17.78, substantially outperforming existing mask-guided VTON methods while achieving competitive performance comparable to powerful closed-source image editing models, such as GPT-Image-2 and Nano Banana Pro. These results demonstrate the effectiveness of our method even under the single-reference setting, without benefiting from additional reference information.

(2) Multi-Reference Comparisons on MRG-4K. We further evaluate methods under the multi-reference setting. The methods marked with an asterisk, including Qwen-Image-Edit-2511*, FireRed-Image-Edit-1.1*, GPT-Image-2*, and Nano Banana Pro*, are provided with multiple garment reference images during inference. Notably, all four methods exhibit performance degradation when moving from single-reference to multi-reference inputs, suggesting that simply providing additional references does not necessarily improve try-on performance and may introduce feature interference. In contrast, MR-VTON achieves the best performance across all five metrics, with an FID of 21.79, KID of 16.47, LPIPS of 0.2213, DiffSim of 0.8014, and PSNR of 18.14. These results highlight the effectiveness of our proposed dual-branch architecture in selectively integrating complementary information from multiple references rather than simply concatenating them.

Table R2: Single-Reference and Multi-Reference Experiments on MRG-4K. * denotes multi-reference garment images.

CategoryMethodFID ↓KID ↓LPIPS ↓DiffSim ↑PSNR ↑
Single-ReferenceOOTDiffusion54.87220.140.26440.724215.35
CatVTON37.9057.900.25780.730017.15
MV-VTON35.3679.280.24710.735217.54
FitDiT41.33104.210.26580.737516.07
Qwen-Image-Edit-251130.8940.880.29130.730115.26
FireRed-Image-Edit-1.124.8426.470.23670.785517.37
GPT-Image-224.8621.660.28400.771417.15
Nano Banana Pro23.2721.990.24720.785417.65
Ours (single-reference)25.4522.800.24010.774017.78
────────────────────────────────────
Multi-ReferenceQwen-Image-Edit-2511*38.6077.860.43720.645213.55
FireRed-Image-Edit-1.1*56.58230.560.41620.638615.72
GPT-Image-2*30.5538.100.37360.710915.47
Nano Banana Pro*23.3722.670.25630.783017.52
Ours (MR-VTON)21.7916.470.22130.801418.14

W3 & Q3:

Thank you for pointing out this discrepancy. We would like to clarify that the original CatVTON paper reports an unpaired FID of 9.015 on VITON-HD, where the model is specifically trained and evaluated on this dataset.

In contrast, Table 3 of our paper is intended to evaluate the generalization ability of different VTON methods. Specifically, all compared models are directly transferred from the experiments in Table 2 to VITON-HD, DressCode, and StreetTryOn, without any additional dataset-specific fine-tuning.

Notably, even compared with the FID of 9.015 reported in the original CatVTON paper, our method achieves a comparable FID of 9.28 on VITON-HD, despite not being specifically trained on this dataset. This further demonstrates the strong zero-shot generalization capability of MR-VTON across different VTON benchmarks.

W4 & Q4:

Thank you for raising this important question. We would like to clarify that our dual-branch architecture is primarily designed to exploit complementary information from multiple garment references with available human pose information. However, the absence of reference poses does not necessarily lead to the failure of the entire framework, as explained below.

(1) Role of Pose-Guided Aggregation. As illustrated in Fig. 6, the Target Pose-Guided Feature Aggregation branch primarily contributes to fine-grained garment details, such as the neckline, while having a relatively limited impact on the overall garment appearance. Furthermore, the ablation results in Table 6 demonstrate that the model still achieves competitive performance even without the pose-guided aggregation branch. These observations indicate that pose guidance mainly enhances local garment detail preservation rather than being essential for overall garment generation.

(2) Handling Missing Reference Poses and Generalization. For flat-lay garment images, where human pose information is unavailable, we use empty pose inputs as placeholders. Although the pose-guided aggregation branch cannot establish meaningful pose correspondences in this setting, the model can still generate reasonable try-on results by leveraging the garment appearance information learned from diverse training data.

Furthermore, as shown in Table 3, our method demonstrates strong zero-shot generalization across VITON-HD, DressCode, and StreetTryOn without dataset-specific fine-tuning, achieving an FID of 4.69 on DressCode, 31.31 in Model-to-Street, and a Dress FID of 28.00 in Street-to-Street. These results further support the robustness and generalization capability of our framework across diverse VTON scenarios and reference configurations.

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

This paper addresses a new problem in multi-reference in-the-wild virtual try-on (VTON), where garment details are provided by multi-view model images. It proposes a dual-branch framework, in which one branch performs optimal view selection and the other performs pose-guided feature aggregation. The representation sequences derived from the two branches are then injected into a DiT generator for image synthesis. In addition, a new benchmark is collected to support the experiments.

Critical strengths
  1. A new benchmark for multi-reference in-the-wild VTON is collected.

  2. The dual-branch framework effectively decouples optimal view selection from pose-guided feature aggregation.

  3. Experiments on the new test set MRG-4K and the existing VITON-HD, DressCode, and StreetTryOn datasets demonstrate the effectiveness of the proposed method.

  4. The paper includes extensive ablation study.

Critical weaknesses
  1. The training strategy is insufficiently described. The model appears to use only one edited image as reference and to predict ground-truth video frames (Lines 154–155). How many reference garment images will be sampled during training?

  2. For the unpaired evaluation setting, a VLM-as-judge evaluation using closed-source APIs should be adopted.

  3. The test set MRG-4K contains only about 1K+ samples. FID may be unreliable for such a small-scale test set, especially when the number of samples is below 2,048.

  4. State-of-the-art baselines from year 2026 are missing in Tables 1 and 2.

  5. An ablation study on the number of reference images used in inference used is missing.

  6. I am concerned about whether the curated benchmark will be released to the community, since it is collected from e-commerce sources and may involve copyright restrictions. If it cannot be released, it will be difficult to follow this work.

Questions for authors
  1. Why is soft routing used for training instead of Gumbel-softmax routing?

  2. What happens if pose estimation for the target person fails? Will the pose-guided aggregation branch also fail?

  3. According to Table 8, the total number of test images is 4,598, which differs from 4,505 reported in Line 755. Could the authors clarify this discrepancy?

4 comments

W1:

Thank you for your comment. We would like to clarify that our model is trained with a variable number of garment reference images, ranging from 1 to 5, rather than a single reference image.

Specifically, our training dataset contains 86,627 samples comprising 431,650 images, collected from e-commerce product pages and real try-on video frames. To construct the training triplets, we apply random garment edits to real try-on images using state-of-the-art image editing models. Each training sample consists of an edited person image, 1–5 reference garment images, and a corresponding ground-truth image. During training, the synthetically edited person image and the sampled garment references are used as inputs, while the original real-world video frame serves as the ground truth.

As detailed in Table 7 and Fig. 7(a) of the Appendix, the training set contains 12,103 samples (13.97%) with one reference image, 20,627 samples (23.81%) with two, 22,276 samples (25.71%) with three, 19,894 samples (22.97%) with four, and 11,727 samples (13.54%) with five. This diverse reference configuration enables the model to learn from both single-reference and multi-reference scenarios, facilitating effective garment information aggregation across different numbers of references.

W2:

Thank you for your valuable suggestion. While VLM-as-a-judge evaluations based on closed-source APIs may be sensitive to model selection, prompting strategies, and model versions, we agree that they can provide valuable complementary assessments of perceptual quality. Therefore, to address your concern, we have conducted additional VLM-as-a-judge evaluations using three advanced VLMs: GPT-6-Astra, Gemini-4-Argon, and Qwen4-Max. We adopt a five-point rating scale (1–5), where higher scores indicate better performance, and evaluate the generated results from the following four perspectives:

(1) Clothes Consistency: Overall garment fidelity and preservation of fine-grained details.

(2) Background Consistency: Degree of unintended background changes relative to the input image.

(3) Human & Pose Consistency: Preservation of person identity and consistency of human pose.

(4) Aesthetic Quality: Overall visual quality and absence of severe artifacts, such as limb distortion and garment collapse.

The quantitative results are presented in Table R1 below.

Table R1: VLM-as-Judge Evaluation Conducted Using Closed-Source APIs under the Unpaired Evaluation Setting.

MethodClothesBackgroundHuman & PoseAestheticAverage
OOTDiffusion1.583.752.721.542.39
CatVTON2.493.803.442.753.12
MV-VTON1.633.863.832.102.86
FitDiT2.153.222.461.662.37
Qwen-Image-Edit-25112.183.573.713.323.19
FireRed-Image-Edit-1.13.152.662.712.422.74
GPT-Image-23.813.363.444.033.66
Nano Banana Pro3.693.943.703.893.81
MR-VTON (Ours)3.494.034.053.793.84

As shown in Table R1, MR-VTON achieves the highest overall mean score of 3.84, outperforming all competing methods, including Nano Banana Pro (3.81) and GPT-Image-2 (3.66). Notably, our method achieves the best performance in Background Consistency (4.03) and Human & Pose Consistency (4.05), demonstrating its effectiveness in preserving the original person and background while performing garment replacement. Furthermore, our method achieves competitive performance in Clothes Consistency and Aesthetic Quality.

These additional VLM-based evaluations complement the conventional FID evaluation and provide further evidence of the effectiveness of MR-VTON across multiple perceptual dimensions. By incorporating assessments from three different VLMs, we also aim to reduce reliance on any single model's subjective judgments and provide a more comprehensive evaluation of virtual try-on quality.

W3:

Thank you for your comment. We would like to respectfully clarify the rationale behind our evaluation protocol and address your concerns regarding the reliability of FID on relatively small test sets.

(1) FID is a widely adopted evaluation metric in the virtual try-on (VTON) field, even on benchmarks with relatively small test sets. For example, the StreetTryOn benchmark contains only 909 samples for each of the Shop-to-Street, Model-to-Street, and Street-to-Street (Top) settings. In comparison, our MRG-4K benchmark includes 1,296 test samples, exceeding the sample sizes of these established evaluation settings. Therefore, we believe that reporting FID is both necessary and consistent with common evaluation practices in VTON research.

(2) We do not rely solely on FID to evaluate model performance. We additionally report multiple complementary metrics, including KID, PSNR, LPIPS, and DiffSim, which assess different aspects of generation quality and are generally less sensitive to evaluation sample size than FID. Furthermore, we conduct a user study to assess perceptual quality from a human perspective. Together, these quantitative metrics and subjective evaluations provide a comprehensive and reliable assessment of the compared methods, mitigating the potential limitations of FID under relatively small sample sizes.

W4:

Thank you for your valuable comment. We have carefully surveyed recent state-of-the-art VTON methods published in 2026. However, some relevant methods, such as VC-VTON [1], have not released their official implementations or pretrained checkpoints, making direct comparisons difficult. Therefore, we have added comparisons with RefTon [2], as well as two other recent methods, Any2AnyTryOn [3] and OmniVTON [4], as presented in Table R2 below. Our method achieves the best performance on four out of five metrics.* Compared with RefTon, our method reduces FID by 13.81% and KID by 27.73%, while improving DiffSim from 0.7713 to 0.8014.

Table R2: Comparison of Our Method with Other State-of-the-Art Methods.

MethodPublicationFID ↓KID ↓LPIPS ↓PSNR ↑DiffSim ↑
Any2AnyTryOnICCV 202527.5437.840.213317.800.7709
OmniVTONICCV 202527.9024.810.241317.090.7558
RefTonCVPR 202625.2822.790.223817.860.7713
Ours (MR-VTON)—21.7916.470.221318.140.8014

We would also like to clarify that our original evaluation already included several recent general-purpose image editing models, including Qwen-Image-Edit-2511, FireRed-Image-Edit-1.1, Nano Banana Pro, and GPT-Image-2, which represent competitive image editing approaches available in 2026. These additional experiments provide a more comprehensive evaluation of our method against both recent specialized VTON approaches and general-purpose image editing models.

References

[1] Ma, Zhiyuan, et al. VC-VTON: Towards Across-View and Multi-Posture Driven Virtual Try-On via Spatiotemporal-Aware View Consistency Training. TIP, 2026.

[2] Liuzhuozheng Li et al. RefTon: Reference Person Shot Assist Virtual Try-on. CVPR, 2026.

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

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

W5:

Thank you for your valuable suggestion. During inference, our model supports a variable number of reference images, ranging from 2 to 5, and directly utilizes all available references for each sample. To further investigate the effect of reference quantity, we conduct an additional ablation study by selecting 400 samples from MRG-4K, each containing at least three reference images. For the single-, two-, and three-reference settings, reference images are randomly sampled from the available references of each sample.

As shown in Table R3, our method achieves competitive performance even with a single reference image, demonstrating the effectiveness of our training strategy and the strong capability of the model under limited reference information. Furthermore, increasing the number of references consistently improves all evaluation metrics, indicating that our method effectively exploits complementary information across multiple references to enhance garment representation and virtual try-on quality. These results further validate both the effectiveness of our model and its ability to adapt to different numbers of reference images.

Table R3: Ablation Study on the Number of Reference Images Using 400 Samples from MRG-4K.

MethodFID ↓KID ↓DiffSim ↑PSNR ↑
w/ Single Reference Image32.089.480.803618.23
w/ Two Reference Images31.878.700.804118.24
w/ Three Reference Images31.658.520.804618.30

W6:

Thank you for your interest in our work and for raising this important concern. We fully recognize the importance of reproducibility, data privacy, and accessibility to the research community.

As detailed in Appendix A.2.1 (Data Construction), our dataset is annotated with semantic attributes, which are first automatically inferred and then manually verified. To protect privacy, we blur the eye regions of all livestream subjects and anonymize visible livestream information, including usernames, platform identifiers, and on-screen interface elements.

After the data cleaning, annotation, and privacy protection procedures, there are no remaining restrictions preventing the public release of our curated dataset. Therefore, we will promptly release our code and the curated MRG-4K benchmark to the public upon acceptance, enabling future research, ensuring reproducibility, and facilitating fair comparisons with our method.

Q1:

Thank you for your insightful question. We adopt soft routing rather than Gumbel-Softmax during training for the following three reasons:

(1) Stable Gradient Propagation. Soft routing computes a weighted combination of reference features:

Xsoft=∑i=1KwiXi,wi=exp⁡(gi)∑j=1Kexp⁡(gj)X_{soft} = \sum_{i=1}^{K} w_i X_i,\quad w_i = \frac{\exp(g_i)}{\sum_{j=1}^{K}\exp(g_j)}

This enables continuous gradient propagation across all reference branches, facilitating stable optimization.

(2) Training Stability. Gumbel-Softmax introduces stochastic noise and requires temperature tuning, which may increase optimization instability. In our experiments, we observed that soft routing provides more stable training than Gumbel-Softmax routing.

(3) Training and Inference Strategies. Soft routing allows the model to learn reference importance scores during training, while hard Top-1 routing selects the most informative reference during inference to avoid feature ambiguity. Our routing ablation confirms that hard routing during inference consistently outperforms soft routing across all evaluation metrics.

Q2:

Thank you for your insightful question. We acknowledge that inaccurate or failed target pose estimation may affect the effectiveness of the pose-guided aggregation branch. However, this does not necessarily lead to the failure of the entire framework. As demonstrated in our ablation study in Fig. 6, removing the Target Pose-Guided Feature Aggregation module primarily affects fine-grained garment details, such as the neckline, while having only a limited impact on the overall try-on appearance. This observation can be explained by the following two aspects:

(1) Pose Estimation as Auxiliary Guidance. The target pose is primarily used to guide the cross-attention mechanism in retrieving complementary garment features from multiple references. Therefore, inaccurate pose estimation may reduce the accuracy of local feature alignment and affect fine-grained garment details, rather than directly preventing garment generation. This is consistent with the observations in Fig. 6.

(2) Complementary Role of the Main Reference Selection Branch. Our framework contains two complementary branches. The Adaptive Main Reference Selector (Branch 1) selects the most informative garment reference based on visual features without relying on target pose estimation. Consequently, even when pose guidance is inaccurate or unavailable, the model can still obtain essential garment appearance information through Branch 1, helping preserve overall garment identity and appearance.

Q3:

Thank you for pointing out this discrepancy. We would like to clarify that the 4,598 images calculated from Table 8 represent the total number of reference image instances, rather than the total number of unique images in the test set:

263×2+354×3+385×4+294×5=4,598.263 \times 2 + 354 \times 3 + 385 \times 4 + 294 \times 5 = 4,598.

Specifically, each test sample consists of 2–5 reference garment images, one input person image, and one ground-truth image. However, as described in Appendix A.2.1 (Data Construction), our dataset is constructed by extracting frames from videos. Consequently, the same reference garment images may be shared across multiple samples with different input person images, resulting in repeated reference image instances.

Therefore, while Table 8 counts 4,598 reference image instances, the 4,505 images reported in Line 755 represent the total number of unique images in the entire test set, including reference garment images, input person images, and ground-truth images, after removing duplicates. We will clarify this distinction in the revised manuscript to avoid confusion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.