acceptodds
Under review as a conference paper at ICLR 2027

FitVTON: Fit-aware Virtual Try-On via Body-Garment Size Control

Abstract

While diffusion-based virtual try-on has achieved impressive visual realism, most methods treat the task as 2D inpainting, prioritizing texture preservation over physical plausibility. Consequently, they often produce plausible-looking images that fail to reflect authentic garment fit across diverse body shapes. We present FitVTON, a Fit-aware virtual try-on model on different bodies in the wild. FitVTON encodes Garment-Body Size, a pair of measured body and garment-size levels, through structured text prompts, and learns from simulated try-on triplets that drape five garment sizes on diverse parametric bodies. To improve the fitting effects over garment silhouettes, we introduce two auxiliary heads to predict the masks for both the garment and the exposed body. We further introduce a texture rectification stage to improve realistic appearance from simulated data. To evaluate the fitting fidelity, we curate a real-world dataset, FittingEffect3K, combining VLM-based scoring protocol. Both subjective and quantitative experiments show that FitVTON demonstrates authentic fitting fidelity, with significant sizing accuracy and shape preservation over state-of-the-art methods while maintaining competitive image quality. Code and data will be released upon acceptance.

Primary area:
generative models
Keywords:
Virtual Try-On, Image Editing, Image Processing, Flow-matching

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_KXnUOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies fit-aware virtual try-on, motivated by the observation that existing VTON systems can generate visually plausible clothing transfers while failing to reproduce how a specific garment should actually fit different body shapes. The proposed system, FitVTON, represents fit through a measurement-derived Garment-Body Size prompt and learns garment-body deformation from physics-based synthetic data rather than relying only on conventional appearance-oriented VTON supervision.

Critical strengths
  1. The paper addresses an important limitation of current virtual try-on systems: garment appearance transfer and image realism do not necessarily imply correct garment-body interaction. The proposed simulation pipeline provides controlled supervision over five garment sizes and 16 body prototypes, with multiple poses and wearing styles, producing 1,190,400 synthetic training triplets. Unlike standard paired VTON data, this construction allows the same garment design to be observed under systematically varied garment/body size relationships.
  2. FittingEffect3K and the four-dimensional fit evaluation protocol target garment-body alignment, tightness/looseness, silhouette, and local fitting artifacts rather than relying exclusively on FID/KID. FitVTON obtains the highest reported overall fit score in Table 1, is selected most often in the independent human preference study, and remains competitive on standard image-quality metrics.
Critical weaknesses
  1. FitVTON's inference procedure uses Garment-Body Size prompts generated from measurements of the person and garment. However, in the main baseline comparison, Section 4.2 states that competing approaches use their respective default prompts. This creates an important confound: part of FitVTON's advantage may arise from receiving explicit physical metadata that the competing methods do not receive in the main quantitative experiment.
  2. The claimed generality of garment-body fit control exceeds the diversity represented by the body and size model. FitVTON represents human morphology using only 16 body prototypes per gender, constructed from four height levels and four body-size levels. When garment measurements are unavailable, as in the in-the-wild evaluation, garments are simply treated as M-size; body-size levels are also visually estimated when tape measurements are unavailable.
Questions for authors
  1. Can the authors report a full FittingEffect3K comparison in which competing models receive the same body- and garment-related information available to FitVTON whenever their interfaces permit it—for example body category, height, garment size, or numerical measurements? The qualitative EXP0–EXP3 experiment in Figure 11 is useful, but a quantitative result across the benchmark would be substantially more convincing.
  2. What is the FittingEffect3K fit score before versus after Stage-II texture rectification? The current ablation shows a clear FID/KID improvement from texture rectification, but does not quantitatively demonstrate that fit control is preserved.
4 comments
AuthorsOpenReview ↗

Response to Reviewer KXnU (Part I)

We sincerely thank the reviewer for the thorough review and insightful feedback. We are encouraged that the reviewer recognizes the value of our simulation pipeline in providing systematic supervision across diverse garment-body size relationships and the effectiveness of FittingEffect3K and our fit-oriented evaluation protocol in assessing garment fitting quality. We have carefully considered all comments and hope the following clarifications address the reviewer's concerns.

Q1: Fair comparison when baselines receive body and garment metadata.

A (Q1): Thank you for raising this important point. Specialized mask-free VTON baselines use fixed task prompts (or lack text encoders, such as CatVTON) and were not trained on measurement metadata. To test whether explicit physical metadata explains FitVTON's advantage when interfaces support it, we added two quantitative evaluations:

1) Full FittingEffect3K benchmark (3,350 pairs) with body and garment metadata (Table 1, Appendix B). We evaluated the recent open-weight multi-image editor Qwen-Image-Edit-2511 on all 3,350 pairs, giving it the exact same height and body-size levels as FitVTON alongside garment size M. Even with this explicit metadata, Qwen-Image-Edit-2511 reaches 2.95 vs. FitVTON's 3.08 (Upper: 3.12 vs. 3.22, Lower: 2.80 vs. 2.99, Dress: 2.73 vs. 2.90).

2) 400-triplet controlled prompt sweep across four baselines (Appendix D.2, Table 9). Expanding Fig. 12 quantitatively, we evaluated Any2AnyTryOn, OmniTry, JCo-MVTON, and Nano Banana on 400 real try-on triplets across EXP0 (default), EXP1 (body shape, height category, garment size), EXP2 (numerical height, weight, garment length), and EXP3 (combined):

Table R7. Prompt-level sweep on 400 real try-on triplets (Appendix D.2, Table 9).

ModelEXP0 (Default)EXP1 (Body/Size)EXP2 (Numerical Metrics)EXP3 (Combined)
Any2AnyTryOn2.672.512.422.58
OmniTry2.412.282.312.25
JCo-MVTON2.952.762.752.74
Nano Banana2.892.992.942.86

Providing body categories or numerical measurements slightly lowers fit scores for the three open-source VTON models (which ignore or are perturbed by out-of-distribution prompt tokens) and changes Nano Banana by at most +0.10 (2.89 → 2.99), still well below FitVTON (3.08). Hence, FitVTON's gains come from learning physical garment-body deformation in Stage I rather than from prompt metadata alone.

AuthorsOpenReview ↗

Response to Reviewer KXnU (Part II)

W2: Diversity of the body/size model and in-the-wild assumptions.

A (W2): We sincerely thank the reviewer for this detailed comment. We agree that 16 prototypes per gender are a coarse discretization and view FitVTON as a practical first step toward controllable fit:

  • Coarse body prototypes: Garments include ease, so visible fit shifts occur mainly across ease boundaries rather than fine local variations; finer 3D shape is also underdetermined from a single photo and basic tape measurements. Importantly, the 16 prototypes serve as simulation anchors rather than a restriction on test subjects: real bodies are mapped by physical measurements (Appendix A.6, Table 5) while individual contours enter directly via the person image IaI_a. None of the FittingEffect3K participants is a synthetic prototype.
  • Defaulting unmeasured garments to M-size: Size M is the neutral anchor fitted to the mean SMPL-X body (Appendix A.4). Even under a fixed M label, body conditioning drives meaningful fit adaptation: in Fig. 17 (Appendix D.6), the same M-size Dress A falls below the knee on a 160 cm subject (Person 3) and at mid-thigh on 176–179 cm subjects (Persons 6–7). Whenever e-commerce size charts are available, the exact garment size level follows directly from Table 6.
  • Visually estimated body size: Visual estimation is used only for unconstrained in-the-wild demos (Appendix D.6, where reported height still determines the height level). All quantitative evaluations on FittingEffect3K strictly use tape-measured height, chest, waist, and hips.

We have added these points, along with the single-fabric simulation assumption, to a Limitations paragraph in Sec. 6.


Q2: FittingEffect3K fit score before vs. after Stage II texture rectification.

A (Q2): We thank the reviewer for this valuable comment. We have added this comparison across all 3,350 FittingEffect3K pairs in Appendix D.3 (Table 10 & Fig. 13):

Table R8. Fit scores before and after Stage II (Appendix D.3, Table 10).

Stage / LoRA Mixing (Stage I / Stage II)GBT/LSCLFUpperLowerWhole Fit
Before Stage II (1.0 / 0.0, Stage I only)3.212.832.803.403.262.833.04
After Stage II (0.8 / 0.2, FitVTON default)3.182.762.803.423.222.993.08

Stage II texture rectification at our default 0.8/0.2 LoRA mixing ratio preserves—and slightly improves—overall fit score from 3.04 to 3.08 (paired difference +0.04, 95% participant-bootstrap CI [+0.01, +0.08]), while delivering the substantial FID/KID gains on VITON-HD (13.53 → 10.44) and DressCode (11.66 → 5.21) shown in Table 3. Appendix D.3 also sweeps higher Stage II weights (0.6/0.4 → 2.96, 0.4/0.6 → 2.38), showing why freezing text LoRA during Stage II and blending image LoRAs at 0.8/0.2 achieves both realism and preserved fit control.

Thanks for the additional results and clarifications. I will keep my weak accept rating.

Thank you for taking the time to review our rebuttal and for your continued support. We sincerely appreciate your time and effort throughout the review process.

Reviewer_USwBOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FitVTON, a fit-aware virtual try-on model controlled by structured 'Garment-Body Size' text prompts. It builds GarmentCodeVTON, a simulation dataset of about 1.19M try-on triplets with 5 garment sizes and 16 body prototypes (4 height levels x 4 body-size levels) per gender. FLUX.1 Kontext is fine-tuned in two stages: i) LoRA training on the simulation data, with dual-branch mask supervision, to learn prompt-driven fit priors, and ii) texture rectification on VITON-HD and DressCode that updates only the image LoRA layers. The paper also collects FittingEffect3K (3,350 real try-on triplets from 10 participants and 16 garments) and a GPT-5.2-based fit score. FitVTON obtains the best fit score against four VTON baselines and Nano Banana (3.08 vs 2.82) and is preferred most often in a 20-participant study; FID/KID and ablations of the mask and texture stages are also reported.

Critical strengths
  • The paper constructs the try-on datasets with different body shape and garment size.
  • The paper constructs the real-world benchmark to evaluate fitting accuracy.
Critical weaknesses
  • Insufficient evidence for prompt-driven size control. The paper claims that structured Garment-Body Size prompts control garment fit, but the presented experiments do not isolate the contribution of these prompts from the visual conditioning. Figs 1 and 15 vary the person image together with the body description while keeping the garment-size label at M. They therefore demonstrate adaptation across bodies without establishing the independent effect of size conditioning. Fig. 11 keeps the visual inputs fixed while varying prompt information, but evaluates only the baselines. Although Table 1 reports improved fit scores, including a substantial tightness/looseness gain for lower-body garments, these results do not establish whether the gains depend on the size prompts. Evaluating FitVTON with fixed person and garment reference images while varying only the garment-size label (e.g., XS to XL) can help isolate the contribution of size conditioning and establish whether the model responds consistently to the size labels.
  • The native-backbone comparison does not isolate the fit-specific contribution. Appendix D.1 compares FitVTON with unmodified FLUX.1 Kontext using a multi-image conditioning pipeline, while acknowledging that the native model was trained for single-reference editing. Its failure may therefore reflect limitations in multi-reference conditioning or basic garment transfer, making it difficult to attribute the improvement specifically to the paper’s claimed contribution to fit control. Although the paper includes specialized VTON baselines and Nano Banana, comparisons with more latest open-weight models that support multi-reference editing, such as FLUX.2 [dev] or Qwen-Image-Edit-2511, would strengthen the evaluation. Also, evaluating these models with prompts that explicitly describe the desired tightness or looseness and hem position can help establish how much of the claimed fit control is already achievable through direct textual instructions.
Questions for authors

please check the two critical weaknesses

1 comment
AuthorsOpenReview ↗

Response to Reviewer USwB

We sincerely thank the reviewer for the thorough review and constructive feedback. We have carefully considered all comments and hope the following clarifications address the reviewer's concerns.

Q1: Isolating prompt-driven size control with fixed person and garment images.

A1: Thank you for pointing out this gap. To isolate the effect of the Garment-Body Size prompt from visual conditioning, we added a controlled experiment in Sec. 5 (Fig. 7 & Table 4) using our full two-stage FitVTON model on FittingEffect3K. Keeping the person image, reference garment image, body level, wearing style, and random seed strictly fixed, we vary only the garment-size label across XS, S, M, L, and XL. Using a human parser and each participant's measured height, we quantify physical changes from XS to XL:

Table R4. Garment geometry change from XS to XL with fixed person and garment images (Sec. 5, Table 4).

Garment CategoryΔhem (XS → XL, cm)Δwidth (XS → XL, cm)Area Ratio (XL / XS)Monotonically Ordered
Female upper+11.0+1.71.27×99.2%
Female lower+14.9+0.51.28×95.9%
Female dress+9.2+0.41.16×99.1%
Male upper+3.3+7.01.11×90.1%
Male lower+9.5+2.81.48×94.4%

With identical visual inputs, increasing the size label from XS to XL lowers the hem monotonically in 90.1%–99.2% of sequences (by 9.2–14.9 cm on female garments and male lowers, and +7.0 cm in width on male tops) and expands garment silhouette area by 11%–48%. Fig. 7 visually confirms smooth, physically plausible progressions in hem length, sleeve length, and torso ease driven solely by the size prompt.


Q2: Native-backbone comparison, stronger open-weight editors, and outcome prompts.

A2: We thank the reviewer for this valuable comment. We agree that Appendix D.1 alone only verifies that pretrained FLUX.1 Kontext lacks multi-reference try-on ability prior to our training; fit-specific control itself is isolated by the fixed-input XS–XL sweep in A1 above.

1) Comparison with open-weight multi-reference editors (Qwen-Image-Edit-2511). Following your suggestion, we evaluated Qwen-Image-Edit-2511 on all 3,350 FittingEffect3K pairs using a multi-image try-on prompt that explicitly includes the same height and body-size levels together with garment size M (Appendix B), and added it to Table 1, Fig. 4, and Fig. 15 (FLUX.2 [dev] was omitted due to rebuttal time constraints):

Table R5. Comparison with Qwen-Image-Edit-2511 under body-level and garment-size conditioning (Table 1).

MethodUpper AvgLower AvgDress AvgWhole Avg ↑
Nano Banana3.192.452.832.82
Qwen-Image-Edit-2511 (w/ body & size prompt)3.122.802.732.95
FitVTON (Ours)3.222.992.903.08

Qwen-Image-Edit-2511 is a strong baseline (2.95, surpassing Nano Banana's 2.82), yet FitVTON outperforms it across all categories (3.08 overall), with the largest margins on fit-sensitive lower-body (2.99 vs. 2.80, T/L: 2.68 vs. 2.49) and dress (2.90 vs. 2.73) categories.

2) Effect of richer attribute prompts (EXP0–EXP3) and why we avoid outcome-leaking prompts. In Appendix D.2 (Fig. 12 & Table 9), we scored four baselines across 400 real try-on triplets under four prompt settings (EXP0: base; EXP1: body/size categories; EXP2: explicit height, weight, and garment length; EXP3: combined):

Table R6. Fit scores under four prompt levels on 400 real try-on triplets (Appendix D.2, Table 9).

ModelEXP0 (Base)EXP1 (Body/Size)EXP2 (Metric)EXP3 (Combined)
Any2AnyTryOn2.672.512.422.58
OmniTry2.412.282.312.25
JCo-MVTON2.952.762.752.74
Nano Banana2.892.992.942.86

Adding body categories or metric measurements decreases fit scores for Any2AnyTryOn, OmniTry, and JCo-MVTON, and shifts Nano Banana within [−0.03, +0.10]. All variations stay within ±0.25—comparable to evaluator/rendering variance (e.g., JCo-MVTON changes <1% of pixels across EXP1–EXP3 yet varies by 0.12). Thus, text conditioning alone does not yield reliable fit control without simulation-grounded supervision.

Finally, we intentionally exclude prompts that prescribe where the hem should land or how tight the garment should look (e.g., "hem at mid-thigh, loose waist"). Those are the outputs of fit prediction, not inputs known prior to wearing: supplying them leaks the ground-truth answer into the prompt and turns try-on into oracle-guided image editing. FitVTON takes only pre-try-on body and garment measurements, leaving resulting hem position and tightness for the model to infer (clarified in Appendix D.2).

Reviewer_52vAOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses fit-aware virtual try-on, focusing on how garment fit changes with body shape and clothing size. The authors propose FitVTON, a FLUX.1 Kontext-based framework trained on synthetic try-on triplets generated through a simulation pipeline built upon GarmentCode, SMPL-X, and NVIDIA Warp. The method incorporates garment-body size information through structured text prompts, uses dual-branch mask supervision to guide garment and body geometry, and introduces a second training stage to improve texture realism using real-world data. The paper also presents FittingEffect3K, a real-world benchmark for evaluating garment fit, together with a VLM-based evaluation protocol and human preference study. The reported results show improvements in fitting accuracy over the compared baselines while maintaining competitive image quality on standard virtual try-on benchmarks.

Critical strengths
  1. The paper addresses an important yet underexplored problem in virtual try-on: modeling garment fit across different body shapes and clothing sizes, rather than focusing solely on visual realism.

  2. The simulation-based training pipeline is well motivated, and the combination of dual-branch mask supervision and texture rectification provides a reasonable approach to improving fitting consistency and visual quality.

  3. The proposed FittingEffect3K benchmark and fit-oriented evaluation protocol are valuable. The experimental results show consistent improvements over existing baselines, supported by both VLM-based scores and human preferences.

Critical weaknesses
  1. The simulation pipeline builds on existing components (GarmentCode, SMPL-X, and NVIDIA Warp). The specific technical contributions beyond their integration are not sufficiently distinguished.

  2. The discrete Garment-Body Size representation may not capture fine-grained body proportions or differences in fabric properties, limiting its ability to model more complex fitting behavior.

  3. Dual-branch mask supervision constrains 2D garment and body geometry but does not explicitly model fabric mechanics such as tension, friction, or elasticity. Its contribution to physical fitting accuracy, beyond silhouette consistency, remains unclear.

  4. Although freezing the text LoRA is intended to preserve fitting knowledge, the experiments do not directly verify whether size-control capability is retained after the second training stage.

  5. The main experiments measure similarity to real try-on images but do not establish whether the model responds consistently to changes in Garment-Body Size prompts.

  6. Despite containing 3,350 triplets, FittingEffect3K includes only 10 participants and 16 garments, raising concerns about generalization to unseen body shapes and garment designs.

Questions for authors
  1. Could the authors clarify which components are newly developed beyond the existing capabilities of GarmentCode, SMPL-X, and NVIDIA Warp? In particular, would it be possible to quantify the benefits of the proposed dynamic simulation process compared with a more conventional static garment draping pipeline?

  2. How does the method handle individuals with similar overall body-size categories but substantially different local body proportions, such as waist-to-hip ratios? Have the authors evaluated whether the discrete size representation can capture these differences? How does the method account for garments with similar dimensions but different material properties?

  3. Could the authors clarify the extent to which the proposed supervision improves physical fitting behavior rather than only 2D silhouette consistency? Additionally, would separate ablations of the garment mask branch and the exposed-body mask branch help establish their individual contributions?

  4. Could the authors report the fitting accuracy before and after the second training stage using the GB, T/L, SC, and LF metrics? It would also be helpful to examine how different LoRA mixing ratios affect the trade-off between fitting accuracy and image quality.

  5. Could the authors provide controlled experiments that directly evaluate the effect of Garment-Body Size prompts? For example, how do the generated results change when the size information is removed, altered, or randomly shuffled while keeping the person and garment images fixed? Furthermore, when using different physical sizes of the same garment design with their corresponding prompts, do garment boundaries, hem positions, and tightness exhibit consistent and physically plausible changes?

  6. Could the authors provide a more detailed breakdown of the participants' body shapes and garment categories? Given the relatively small number of independent participants and garments, it would be valuable to report participant-level and garment-level results, ideally with confidence intervals, to determine whether the improvements are consistent across different body types and clothing designs.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 52vA (Part I)

We sincerely thank the reviewer for the thorough review and insightful feedback. We are encouraged that the reviewer recognizes the importance of modeling garment fit across diverse body shapes and clothing sizes and the effectiveness of our simulation-based training pipeline and fit-oriented evaluation protocol. We have carefully considered all comments and hope the following clarifications address the reviewer's concerns.

Q1: Contributions beyond GarmentCode, SMPL-X, and NVIDIA Warp.

A1: We thank the reviewer for this valuable comment. We agree that GarmentCode, SMPL-X, and NVIDIA Warp provide important foundations for our simulation pipeline. However, directly combining these existing tools is insufficient to generate diverse garment-fitting configurations across different body shapes, clothing sizes, and poses. To address these challenges, we develop several extensions and clarify their contributions below.

1) Our Extensions to Existing Simulation Tools. GarmentCode provides sewing patterns, SMPL-X represents human body shapes and poses, and NVIDIA Warp supports XPBD-based cloth simulation. However, the original GarmentCode pipeline primarily supports garment draping on a canonical mean body in an A-pose. Building upon these components, we introduce four extensions, which are now summarized at the beginning of Appendix A:

  • Step-wise Dynamic Simulation (Appendix A.1, Algs. 1–2). The garment first settles on the A-pose body; the body then morphs to the target shape and only afterwards articulates to the target pose with adaptive midpoint subdivision, keeping the cloth coupled to the moving body throughout.
  • Tucked-in and Untucked Control (Appendix A.2, Fig. 8). Outfits are decomposed into separate upper and lower meshes simulated jointly, constraining the upper hem inside or outside the lower waistband.
  • Wrist and Ankle Stop Constraints (Appendix A.2, Fig. 9). Directional boundary projection and velocity damping prevent cuffs and hems from sliding past the hands and feet under gravity and articulation, without rigidly pinning them to the skin.
  • Five-size grading (Appendix A.4): Every design is graded into XS, S, M, L, and XL and simulated across all 16 body prototypes per gender (80 combinations per pose).

2) Why Dynamic Simulation Rather Than Static Draping? We do not run a quantitative baseline with static draping because static draping cannot generate GarmentCodeVTON:

  • Pose Diversity. Placing a 2D pattern directly onto an articulated target pose initializes cloth panels inside limbs/torso, causing immediate solver failure; static draping would restrict supervision to the A-pose.
  • Body-Shape Transitions. Placing a pattern fitted to one body directly onto a larger prototype causes severe cloth-body interpenetration. Gradually morphing from the canonical A-pose makes extreme size-body combinations (e.g., XS on a plus-size body) physically simulable.
  • Motion-Induced Drape. Dynamic coupling preserves realistic tension and folds produced as a garment moves with the body.

Together, these extensions enable the generation of diverse and physically plausible garment-fitting configurations that cannot be directly obtained using the original static draping pipeline.

AuthorsOpenReview ↗

Response to Reviewer 52vA (Part II)

Q2: Discrete Garment-Body Size, Local Body Proportions, and Fabric Properties.

A2: We thank the reviewer for raising this important concern. We agree that discrete body and garment size levels cannot fully capture fine-grained anatomical variations or differences in fabric properties. However, FitVTON is designed as a tractable foundation for controllable fit for three reasons:

  • Garment ease absorbs small proportion differences: Garments are cut with ease, so visible fit regimes shift primarily when body dimensions cross ease boundaries rather than with every centimeter of variation. Taking the median of chest, waist, and hip levels (Appendix A.6, Table 5) prevents a single atypical girth from shifting the prototype, while capturing the scale where fit visibly changes.
  • Ill-posedness from 2D inputs: A single 2D photo and basic tape measurements do not uniquely determine local 3D curvature or soft-tissue distribution, and no real-world dataset pairs fine-grained proportion shifts with ground-truth try-on for supervision or evaluation.
  • Visible local contours still enter through image conditioning: The person image IaI_a provides individual visual contours (such as waist-to-hip silhouette), which the exposed-body branch preserves, while the prompt supplies the relative physical scale between the garment and body that a 2D product photo lacks.

Fabric Properties and Limitations. We also appreciate the reviewer's concern regarding fabric properties and acknowledge that this is an important limitation of our current framework. All garments in GarmentCodeVTON currently share one set of XPBD cloth parameters, so FitVTON does not distinguish identical patterns made of different fabrics (e.g., stiff denim vs. stretch jersey). While the simulation pipeline supports varying cloth parameters, material-conditioned try-on also requires material labels at inference. With 10 participants in FittingEffect3K, we also cannot isolate waist-to-hip ratio statistically. We will add these points to a Limitations paragraph in Sec. 6.


Q3: Physical Fitting vs. 2D Silhouette, and Role of the Two Mask Branches.

A3: We thank the reviewer for this insightful comment. We agree that 2D silhouette supervision alone cannot fully characterize the physical behavior of garments. However, our framework incorporates physical fitting information through both simulation-based supervision and the complementary design of the two mask branches. We would like to clarify these aspects as follows.

1) Physical Fitting Beyond 2D Boundaries. While the mask loss operates on 2D silhouettes, the ground-truth masks (Mg,Mb)(\mathbf{M}^g, \mathbf{M}^b) are direct projections of 3D XPBD cloth-body equilibrium states, encoding physically solved hem drop, ease expansion, and body-cloth contact. Furthermore, interior physical behavior—such as tension pulls, fabric tautness, and local folds—is supervised pixel-wise by LCFM\mathcal{L}_{\text{CFM}} on the simulated renderings, and evaluated by our T/L and LF metrics. Removing dual-branch mask supervision drops not only GB and SC, but also T/L (Upper: 2.94 → 2.71, Lower: 2.68 → 2.52) and LF (Upper: 3.69 → 3.39, Lower: 3.37 → 3.23), reducing the overall fit score from 3.08 to 2.87 (Table 1).

2) Complementary Roles of the Two Mask Branches. We respectfully note that we did not train separate single-branch models during Stage I due to the cost of full-scale training on 1.19M triplets, because the two branches are designed as a coupled partition around fit-critical boundaries:

  • Garment-mask branch only: Without constraining exposed skin/limbs, supervising only Mg\mathbf{M}^g allows a trivial shortcut where the generator alters visible body contours (e.g., thinning arms, waist, or legs, or hallucinating extra clothing coverage) to match the target garment mask rather than deforming the garment around the person's true body.
  • Exposed-body branch only: Supervising only Mb\mathbf{M}^b leaves unexposed garment boundaries unconstrained—such as an untucked top hem draping over trousers, outer silhouette ease/looseness, or long hems—causing the garment region to revert toward an averaged neutral fit.

Jointly supervising both branches anchors the invariant body silhouette while forcing the garment boundary to absorb size-dependent deformation.

AuthorsOpenReview ↗

Response to Reviewer 52vA (Part III)

Q4: Fitting accuracy before/after Stage II and LoRA mixing-ratio trade-off.

A4: We thank the reviewer for this valuable comment. We have added this evaluation on all 3,350 FittingEffect3K pairs in Appendix D.3 (Table 10 & Fig. 13), varying the inference mixing ratio between the Stage I fitting LoRA and Stage II texture LoRA from 1.0/0.0 (Stage I only) to 0.4/0.6:

Table R1. LoRA mixing ratios on FittingEffect3K (Appendix D.3, Table 10). †: 95% participant-bootstrap CI of the paired difference vs. 0.8/0.2 excludes zero; Whole includes 50 dress pairs; FID/KID from Table 3.

Fitting / Texture LoRAGBT/LSCLFUpperLowerWhole FitVITON-HD FID / KIDDressCode FID / KID
1.0 / 0.0 (Stage I only)3.212.832.803.403.262.833.04†13.53 / 4.7911.66 / 7.72
0.8 / 0.2 (FitVTON default)3.182.762.803.423.222.993.0810.44 / 2.035.21 / 1.21
0.6 / 0.43.012.882.733.353.042.902.96†——
0.4 / 0.62.202.292.082.942.502.252.38†——
  • Fit is preserved after Stage II: At the default 0.8/0.2 ratio, overall fit does not degrade—it slightly improves from 3.04 to 3.08 (paired Δ = +0.04, 95% CI [+0.01, +0.08]), while substantially improving realism on VITON-HD (FID 13.53 → 10.44) and DressCode (11.66 → 5.21).
  • Trade-off across ratios: Increasing the Stage II weight to 0.6/0.4 (2.96) and 0.4/0.6 (2.38) degrades fitting. As visualized in Fig. 13, because Stage II trains on real pseudo-triplets with a fixed tall, slim, M-size prompt, an overly high texture weight overrides Stage I garment transfer and leaves the original clothing or blends both garments. Blending at 0.8/0.2 retains Stage I fit control while capturing Stage II texture realism.

Q5: Controlled experiments on Garment-Body Size prompts.

A5: We thank the reviewer for this valuable suggestion. To further validate the effectiveness of Garment-Body Size prompts, we conduct additional controlled experiments and provide the following clarifications.

1) Varying garment size (XS to XL) with fixed visual inputs. We added a dedicated controlled study in Sec. 5 (Fig. 7 & Table 4) using the full two-stage model across FittingEffect3K. Fixing the person image, reference garment, body level, wearing style, and random seed, we vary only the garment-size label from XS to XL and measure the segmented garment geometry using each participant's physical height:

Table R2. Garment geometry change from XS to XL with fixed visual inputs (Sec. 5, Table 4).

Garment CategoryΔhem (XS → XL, cm)Δwidth (XS → XL, cm)Area Ratio (XL / XS)Monotonically Ordered
Female upper+11.0+1.71.27×99.2%
Female lower+14.9+0.51.28×95.9%
Female dress+9.2+0.41.16×99.1%
Male upper+3.3+7.01.11×90.1%
Male lower+9.5+2.81.48×94.4%

Increasing the size label consistently lowers hems (+3.3 to +14.9 cm, monotonic in 90.1%–99.2% of sequences) and expands garment width and silhouette area (+11% to +48%), confirming that size-control capability is retained after Stage II.

2) Removing or mismatching prompt information. Because FitVTON is trained exclusively with structured Garment-Body Size prompts in both stages, removing the size prompt creates a train–test conditioning mismatch outside the learned text-LoRA distribution, causing noticeable degradation in visual quality and unstable garment transfer/boundaries. Conversely, altering/mismatching the size label shifts the generated hem and looseness systematically away from the ground-truth reference, as quantified by the up-to-14.9 cm hem shift and 48% area change in Table 4.

3) Physical garment sizes in FittingEffect3K. Each of the 16 designs in FittingEffect3K was captured in a single physical specimen with recorded tape measurements (mapped to its corresponding Garment-Body Size level via Table 6) worn across diverse bodies, while multi-size controllability of a fixed garment design is directly verified by the XS–XL sweep in Sec. 5.

AuthorsOpenReview ↗

Response to Reviewer 52vA (Part IV)

Q6: FittingEffect3K participant/garment breakdown and confidence intervals.

A6: We sincerely thank the reviewer for this detailed and insightful comment. We have added the full dataset composition and per-participant statistical breakdown across all 3,350 pairs in Appendix D.4 (Tables 11–12).

Participant & garment composition (Table 11). FittingEffect3K comprises 5 female (F1–F5; height 155–175 cm, chest 79–109 cm, waist 64–92 cm, hips 88–114 cm; 950 triplets) and 5 male (M1–M5; height 170–184 cm, chest 90–104 cm, waist 78–102 cm, hips 97–106 cm; 2,400 triplets) participants in 5 poses each. Garments span 16 items: female tank top, T-shirt, long-sleeve shirt, shorts, trousers, pleated skirt (150 triplets each), and 2 dresses (25 each); male zip-up jacket, hooded jacket, polo shirt, T-shirt, 2 trousers, and 2 patterned shorts (300 each). Measurement ranges rather than individual rows are reported to protect participant privacy.

Per-participant fit scores (Table 12, 3,350 pairs, GPT-5.2).

Table R3. Per-participant fit scores on FittingEffect3K (Appendix D.4, Table 12). ∗: 95% CI of the paired difference from FitVTON excludes zero; overall CIs resample participants: +0.78, +0.51, +0.53, and +0.34.

ParticipantFitVTONCatVTONAny2AnyTryOnOmniTryJCo-MVTON
F13.012.12∗2.50∗2.59∗2.50∗
F23.011.69∗2.41∗2.40∗2.32∗
F32.941.87∗2.54∗2.39∗2.59
F43.032.11∗2.43∗2.20∗2.70∗
F53.392.50∗2.33∗2.18∗2.74∗
M12.872.582.932.952.97
M23.062.613.012.752.87
M33.082.842.45∗2.842.87
M43.332.45∗2.55∗2.51∗2.99
M53.082.23∗2.55∗2.692.85
Overall3.082.30∗2.57∗2.55∗2.74∗

FitVTON ranks first on 9 of 10 participants and beats every baseline overall with a 95% participant-bootstrap CI excluding zero. Category-level performance across Upper (3.22), Lower (2.99), and Dress (2.90) is reported in Table 1; we omit a separate 16-row per-garment table in the current revision for space, and will include full per-garment statistics upon code/benchmark release.

Comments on the paper

3 comments

Thank you very much for your thoughtful follow-up and for raising your score. As fellow researchers in virtual try-on, we fully agree that reliable real-world size prediction remains a very difficult goal. How a garment fits depends not only on garment and body measurements, but also on its cut, its material properties, and the wearer's detailed body shape, and a VLM-based score cannot fully capture actual fit. We see FitVTON as an initial step that makes garment and body size explicitly controllable, intended to open up this direction and invite further work, rather than a complete solution. We are already pursuing follow-up research along these lines, including numerical garment and body-shape representations, and fit evaluation grounded in physical measurements. Hope you will see our further work soon. Thank you again for sharing your expertise; your comments have helped us improve the paper.

Dear Area Chairs

We sincerely thank you for your time and efforts in managing the review of our submission. We have provided point-to-point responses to all reviewers' comments. However, by far we didn't receive the feedback from reviewer 52vA and reviewer USwB yet. As the "Author-Reviewer" discussion phase is ending soon, we would greatly appreciate it if you could kindly remind these reviewers to provide further feedback so that we can address any remaining concerns before the end of the discussion phase.

All the best

Authors of Submission #1383

Dear Area Chairs

By far, we still haven't received a response from Reviewer 52vA yet. We believe that we have already provided thorough responses to Reviewer 52vA. Since the rebuttal period will end in 30 minutes, we would be grateful if you could kindly remind the reviewer again.

All the best

Authors of Submission #1383

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.