acceptodds
Under review as a conference paper at ICLR 2027

FitFlow: Decoupled Flow Matching Model for Fit-aware Virtual Try-On

Abstract

Virtual try-on (VTON) has evolved from merely generating visually appealing results to focusing on the precise fit between the garment and the target body. Existing methods typically rely on simulated data and leverage numerical control or a structured prompt to achieve garment fit. However, such control signals lack spatial precision for continuous fit adjustment. While the reliance on synthetic data fails to capture fine-grained fit textures (e.g., elasticity and subtle draping), leading to visual realism degradation. Motivated by the nature of flow-matching trajectories, where high-noise timesteps establish global layout and low-noise timesteps restore fine details, we propose FitFlow, a decoupled flow matching model that leverages large-scale synthetic data to learn global body-aware fitting and real try-on pairs at late timesteps to recover fine-grained fit textures. We introduce a simple Garment Size Encoding, that warps the garment image onto a canonical grid based on its relative size to the target body, providing continuous fit control while maintaining photorealistic garment appearance. We also develop a measurement-annotated try-on dataset and evaluation protocol for authentic try-on and fit analysis. Experiments demonstrate that our method achieves improved fit consistency, spatial controllability, and visual fidelity over state-of-the-art baselines.

Primary area:
generative models
Keywords:
Virtual Try-on, Generative model

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

This paper proposes a decoupled flow matching framework for measurement-guided virtual try-on. The method assigns fitting supervision from synthetic data to the high-noise stage and supervision from real images to the low-noise stage, connecting the two training stages via a structural bridging mechanism. Additionally, the paper introduces GSE, which maps the dimensional relationships between garments and the human body to image space to enable continuous fitting control. The authors also constructed a dataset with measurement annotations and proposed relative fitting metrics such as HFE and SFE to evaluate the accuracy of garment length. Experimental results in the paper demonstrate that this method achieves good performance in terms of fitting accuracy, controllability, and visual quality.

Critical strengths

1.The research question and motivation are fairly clear. The paper provides a relatively clear analysis of the respective strengths and limitations of real-world fitting data and synthetic data, laying a sound foundation for research into how to combine these two types of data to achieve controlled garment fit.

2.The experimental results and appendix provide relatively comprehensive information. The authors not only performed ablation studies on the main modules but also analyzed the segmentation timesteps and provided detailed explanations of data construction, handling of evaluation failures, and the setup for reproducing the baseline results.

Critical weaknesses

1.The technical innovation is limited. The three main contributions proposed in the paper are relatively straightforward: time-step-based synthetic/real supervised allocation and bridging, clothing-human scale encoding, and keypoint-based fitting metrics. Although these designs demonstrate some effectiveness, the overall technical approach is relatively simple, and the methodological innovation embodied by their combination remains limited.

2.The current experiments are insufficient to support some of the paper’s main claims. Although the experiments show that phased training outperforms directly mixing synthetic and real data, they have not directly demonstrated that synthetic data is better suited for early-stage global fitting, while real data is better suited for late-stage detail refinement. The absence of a reverse comparison (real-early versus synthetic-late) makes it difficult to determine whether the improvement stems from this specific division of responsibilities or simply from training the two data domains separately over time steps.

Questions for authors

1.Can the author provide more direct evidence to verify the division of labor proposed in the “flow trajectory”? For example, one could first use real data and then synthetic data, or analyze intermediate results from different time periods to verify whether the overall fit structure was indeed primarily formed in the early stages, while the specific details of the fit were mainly optimized in the later stages.

2.Quantitative Comparison of GSE. The appendix describes a controlled comparison between GSE, numerical conditioning, and prompt conditioning, but presents only qualitative results. Could the authors provide the corresponding HFE, SFE, and IoU results?

3.Since HFE and SFE are calculated based on keypoints detected by DeepFashion2 and MMPose, is the positioning of these keypoints on the generated images reliable enough? To what extent will positioning errors affect the final reported fitting score?

4 comments
AuthorsOpenReview ↗

Responses to Reviewer tCfE(Part1)

We sincerely thank the reviewer for the thorough review and insightful feedback. We are encouraged that the reviewer recognizes the sound foundation our work establishes for combining synthetic and real data to achieve controlled garment fit, as well as the simplicity and effectiveness of our framework. We have carefully considered all comments and hope the following clarifications address the reviewer’s concerns.

Q1: Concerns Regarding the Technical Contributions of Timestep-Based Supervision, Structural Bridging, Garment Size Encoding, and Keypoint-Based Fit Evaluation.

A1: We sincerely thank the reviewer for recognizing the importance of fit-aware VTON, as well as our comprehensive experimental analysis and promising results. To address the concern regarding methodological novelty, we further clarify our contributions from three perspectives.

1) Decoupled Flow Matching Framework. Although timestep-dependent training has been explored in prior work, FitFlow's key contribution lies in aligning the complementary characteristics of synthetic and real data with the coarse-to-fine generation process of Flow Matching, establishing a unified cross-domain training framework. To address potential structural inconsistencies arising from stage-specific supervision, we introduce the Structural Bridge to connect synthetic structural priors with real appearance supervision through a continuous training path. This mechanism goes beyond simple data allocation, requiring careful coordination of cross-domain supervision, timestep partitioning, and training path design. Its effectiveness is further validated by our ablation studies.

2) Garment Size Encoding (GSE). Incorporating physical garment measurements into VTON remains challenging. Unlike FIT's Numerical Embedding and FitVTON's Text Embedding, GSE leverages the garment-length-to-body-height ratio to proportionally scale the entire garment within a canonical grid, providing simple, efficient, and geometrically interpretable size control. As also acknowledged by Reviewer BYF2 in Critical Strengths, GSE offers "Simple and effective conditioning for body-relative garment scale." and “Its simplicity makes it a useful contribution to measurement-guided garment placement.”

3) Fit-Aware VTON Benchmark and Evaluation. Existing datasets, such as DressCode and VITON-HD, lack explicit body and garment measurements. To address this limitation, we construct a real-world try-on benchmark with measurement annotations. Furthermore, conventional metrics (FID, KID, SSIM, and LPIPS) cannot directly assess garment fit. We therefore introduce HFE and SFE to provide more direct quantitative evaluation of size-aware garment placement.

Overall, FitFlow's contributions extend beyond individual components to the unified development of cross-domain complementary supervision, geometric size encoding, and a dedicated benchmark and evaluation protocol for fit-aware VTON. We hope these clarifications help the reviewer better appreciate the novelty and broader contributions of our work and respectfully invite the reviewer to reconsider their assessment.

AuthorsOpenReview ↗

Responses to Reviewer tCfE(Part2)

Q2: Validation of Stage-Specific Supervision Through Reverse Data Allocation and Intermediate results Visualization.

A2: We thank the reviewer for this insightful suggestion. To further validate the proposed division of labor along the flow trajectory, we provide two additional experiments:

1) Reverse Allocation. We reverse the supervision order by assigning real data to high-noise timesteps and synthetic data to low-noise timesteps. As shown in Table R5, FitFlow outperforms the reversed strategy across all seven metrics, supporting the effectiveness of the proposed synthetic-early/real-late allocation.This result has been incorporated into Table 2 of the revised manuscript.

Table R5. Quantitative comparison between reversed and proposed timestep supervision strategies. KID is reported in units of 10−310^{-3}. Best results are highlighted in bold.

MethodFID ↓KID ↓SSIM ↑LPIPS ↓HFE(%) ↓SFE(%) ↓IoU ↑
Reverse Allocation15.4191.9820.84250.12257.718.400.8023
FitFlow (Ours)13.9300.5160.85260.11525.545.860.8423

2) Intermediate Results Visualization. In Fig. 2, we have provided attention visualizations suggesting that early high-noise timesteps primarily shape macro geometry and body–garment alignment, whereas late low-noise timesteps focus on localized textures and wrinkles. To further clarify this observation and address the reviewer's concern, we additionally visualize intermediate sampling results at different timesteps. These results further illustrate the progression from overall garment structure to fine-grained appearance details during generation. We have included these visualizations as Fig. 16 in the revised manuscript. Together, these results provide both quantitative and qualitative evidence supporting our stage-specific supervision strategy and the proposed coarse-to-fine division of labor along the flow trajectory.

Q3: Quantitative Comparison of GSE with Numerical and Prompt Conditioning Using HFE, SFE, and IoU.

A3: We thank the reviewer for this careful comment. We would like to clarify that the quantitative results corresponding to the qualitative comparisons in the Appendix are already reported in Table 4 of the main manuscript, including HFE, SFE, and IoU.

Q4: Reliability of DeepFashion2 and MMPose Keypoint Detection and the Impact of Localization Errors on HFE and SFE.

A4: We sincerely thank the reviewer for this insightful comment. We acknowledge that keypoint localization errors may affect HFE and SFE. We would like to clarify our evaluation protocol from two aspects.

1) Keypoint Reliability. As described in Appendix A (Real-Data Curation) and Fig. 10(a), our real-world dataset undergoes manual inspection and filtering to ensure reliable human pose estimation, garment landmarks, and segmentation masks. Samples with missing or unreliable keypoints are removed during data curation. We have provided additional details on these filtering criteria in Appendix A of the revised manuscript.

2) Handling of Evaluation Failures. As detailed in Appendix B (HFE and SFE Qualification), we average multiple hem landmarks to reduce localization noise and evaluate both sleeves independently. Samples with unreliable keypoint detection are assigned the maximum fitting error of 100% rather than excluded. This transparent and consistent evaluation protocol avoids artificially optimistic scores, while evaluation over 2,000 test samples helps reduce the influence of individual localization errors on the aggregate results.

The reverse-allocation comparison addresses my main concern about the choice to use synthetic data early and real data late. I am raising my rating from 2 to 3. In the final version, please align the fit claims with the evidence and use manual annotations to check whether keypoint errors affect the HFE and SFE comparisons.

AuthorsOpenReview ↗

Thank You for Your Constructive Feedback

Dear Reviewer tCfE,

We sincerely thank you for your positive feedback and reconsideration of our work. We are glad that the reverse-allocation experiment helps address your main concern. In the final version, we will carefully refine our fit-related claims to better align with the experimental evidence and conduct manual annotation-based validation to assess the potential impact of keypoint estimation errors on HFE and SFE.

We greatly appreciate your constructive suggestions, which have helped us improve the paper.

Best wishes,
The Authors

Reviewer_BYF2OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FitFlow is a measurement-guided virtual try-on framework that trains a single LoRA-adapted flow-matching generator using synthetic and real data assigned to different timestep intervals. Synthetic cross-size samples supervise high-noise timesteps (t>0.95t>0.95), while real samples supervise lower-noise timesteps, with the aim of learning body–garment structure and realistic appearance, respectively. During training, a structural bridge adds latent-space paths between real targets and re-noised predictions from an auxiliary model fine-tuned exclusively on synthetic data; inference uses a single sampling trajectory. Garment Size Encoding (GSE) is a parameter-free scheme that conveys garment length relative to body height by resizing the garment foreground on a canonical grid. The authors also construct a measurement-annotated real-world dataset and a synthetic dataset, and propose body-relative hem and sleeve fitting errors. Experiments on the curated benchmark report improved fit accuracy and image quality over baselines on most metrics and include component ablations. A separate experiment compares GSE with MLP- and prompt-based conditioning on the FIT benchmark under a shared backbone.

Critical strengths
  1. Effective integration of synthetic and real supervision.
    The paper addresses a practical challenge in fit-aware VTON by combining controlled size variation from synthetic data with appearance supervision from real images. The ablations support the benefits of timestep partitioning and the structural bridge, showing improvements in garment placement and image quality. The method also retains a single sampling trajectory at inference.

  2. Simple and effective conditioning for body-relative garment scale.
    GSE expresses garment length relative to body height through image-space scale, without a learned measurement encoder. Both the component ablation and the shared-backbone comparison support its effectiveness over the evaluated alternatives. Its simplicity makes it a useful contribution to measurement-guided garment placement.

Critical weaknesses
  1. The evidence for size-aware fit is mainly limited to garment-length control.
    The paper is motivated by overall body–garment fit, including draping and stretching. However, GSE explicitly encodes garment length relative to body height, and the continuous-control examples mainly show changes in hem position. In the first three rows of Figure 11, sleeve extent and apparent looseness appear to change relatively little. These examples demonstrate length responsiveness, but do not establish coherent changes in overall garment size or fit. Although circumference and sleeve measurements are collected, their role in conditioning is unclear. SFE and IoU provide useful evidence on cuff placement and silhouette, but broader size-dependent fit remains less directly supported.

  2. Generalization and baseline fairness are not sufficiently established.
    The complete framework is evaluated primarily on the authors’ curated benchmark, without reported results on VITON-HD or DressCode. Appendix C describes the FIT comparison as evaluating a GSE-only variant rather than the complete framework. The paper also does not clearly specify whether the trainable baselines in the main comparison received comparable training data, measurement inputs, and target-domain adaptation. The Real-Only ablation provides a useful in-domain reference, but does not resolve this uncertainty for external baselines. This makes the source of the reported gains harder to assess.

Questions for authors
  1. Is the intended control garment length or overall garment size? How do measurements other than garment length and body height enter the generator, and what evidence supports changes in overall fit beyond hem placement?

  2. Were the trainable baselines in the main comparison fine-tuned on the same real training split, or were public checkpoints used directly? Please clarify their access to synthetic data, measurement inputs, masks or prompts, and adaptation budgets.

  3. Which garment condition—GSE-encoded or original—is used during inference above and below tst_s? Appendix C specifies different representations during training, but the corresponding inference procedure is unclear.

  4. Does the “FitFlow (Ours)” row in Table 3 correspond to the GSE-only variant described in Appendix C?

Suggested additional evaluation: Results on established benchmarks such as VITON-HD and DressCode would help assess general try-on quality beyond the curated dataset. Such comparisons should explain how missing measurements are handled and provide comparable adaptation for trainable baselines.

4 comments
AuthorsOpenReview ↗

Responses to Reviewer BYF2(Part1)

We sincerely thank the reviewer for the thorough review and insightful feedback. We are encouraged that the reviewer recognizes the effective integration of synthetic and real supervision and the simplicity and effectiveness of GSE for body-relative garment scale control. We have carefully considered all comments and hope the following clarifications address the reviewer's concerns.

Q1: Clarification of Overall Garment Size Control, the Role of Additional Measurements, and Other Evidence for Size-Dependent Fit Beyond Hem Placement.

A1: We sincerely thank the reviewer for this detailed and insightful comment. This concern is closely related to W1 regarding the scope of garment size control. We would like to clarify the following three aspects.

1) Clarification of GSE-based Size Control. GSE controls the overall garment scale rather than merely adjusting the hem position. Specifically, it uses the garment-length-to-body-height ratio to uniformly scale the entire garment within a canonical grid, preserving the relative geometry of sleeves, waist, and garment body. Other measurements, such as sleeve length and waist circumference, are not independently provided to the generator. Meanwhile, local fit characteristics, including draping, stretching, and looseness/tightness, are learned from training data. Therefore, sleeve changes may appear less pronounced than hem displacement, as observed in Fig. 13.

2) Evidence for Overall Fit Beyond Hem Placement. Conventional VTON metrics (FID, KID, SSIM, and LPIPS) primarily assess image quality and perceptual similarity rather than body–garment fit. However, reliably measuring waist circumference, shoulder width, or 3D garment–body clearance from generated 2D images remains challenging due to pose variations, occlusions, and non-rigid deformations (Appendix B). Therefore, we adopt HFE and SFE to evaluate longitudinal garment placement and Silhouette IoU to assess overall 2D garment shape consistency, providing measurable evidence of size-aware adaptation while acknowledging the limitations of 2D evaluation.

3) Future Work. We acknowledge that a more comprehensive evaluation of physical garment fit remains an important direction. In future work, we plan to extend FitFlow to 3D by jointly reconstructing body and garment surfaces and measuring region-wise garment–body clearance, enabling more detailed evaluation of tightness, ease, and draping beyond longitudinal fit.


Q2: Generalization to Public Benchmarks (DressCode and VITON-HD).

A2: We thank the reviewer for this valuable comment. We address the concerns regarding generalization to public benchmarks and baseline adaptation separately. Unlike conventional VTON, FitFlow requires body and garment measurements for size-aware conditioning. However, DressCode and VITON-HD do not provide these measurements, which prevented direct evaluation of the complete framework in our original submission. For the FIT dataset, which is synthetic, we evaluate only the GSE-based variant rather than the complete FitFlow framework, as described in Appendix C. To assess generalization without measurement annotations, we fine-tuned a no-GSE variant of FitFlow on DressCode and evaluated it under the unpaired setting. The results are summarized below.

Table R3. Quantitative comparison on DressCode (unpaired).* KID is reported in units of 10−310^{-3}. Best and second-best results are highlighted in bold and underlined, respectively.

MethodFID ↓KID ↓
Any2AnyTryOn4.8290.758
CatVTON7.4002.619
OmniTry5.3981.028‾\underline{\text{1.028}}
RefTON5.3101.360
FitVTON5.2111.208
FitFlow (Ours, no-GSE)5.198‾\underline{\text{5.198}}1.029

FitFlow achieves competitive performance on DressCode (FID = 5.198, KID = 1.029), demonstrating its generalization to a public VTON benchmark without measurement-based conditioning. We also provide additional qualitative comparisons in Fig. 6, and the quantitative comparison previously given as Table R3 in the response has been added to the revised manuscript as Table 3. We will include more comprehensive evaluations on DressCode and VITON-HD in the camera-ready version.

AuthorsOpenReview ↗

Responses to Reviewer BYF2(Part2)

Q3: Fairness of Baseline Comparisons, Training Data, Measurement Conditioning, and Adaptation Settings.

A3: We thank the reviewer for this valuable comment. Our benchmark and DressCode both primarily consist of fashion e-commerce images, while our benchmark additionally provides body and garment measurements. Given the similarity in data sources, we initially evaluated the baselines without additional target-domain fine-tuning. Moreover, except for FitVTON, the compared methods do not explicitly support measurement-based conditioning.

To ensure a fairer comparison, we conduct additional target-domain adaptation experiments. Specifically, we fine-tune Any2AnyTryOn, CatVTON, OmniTry, and RefTON on the same 2,000 real training pairs for 5 epochs, using a global batch size of 8, a learning rate of 1×10−41 \times 10^{-4}, and a resolution of 768×1024768 \times 1024. These methods are adapted using only real data, without additional synthetic training. As FitVTON is the only baseline with an explicit synthetic fitting stage, we retain its original two-stage strategy, using 15,200 synthetic triplets for fitting training, followed by 2,000 real training pairs for texture adaptation. For measurement conditioning, we adopt a heuristic mapping strategy that converts body height and bust, waist, and hip measurements into four height categories (short, medium-short, medium-tall, and tall) and four body-size categories (slim, medium-slim, medium-plus-size, and plus-size). We further construct garment prompts based on garment category, while the other baselines retain their original conditioning interfaces without explicit measurement inputs.

Table R4: Comparison with Target-Domain Adapted Baselines. Δ denotes the change relative to the original results reported in the main paper. Best results are highlighted in bold.

MethodHFE(%) ↓SFE(%) ↓IoU ↑ΔHFEΔSFEΔIoU
Any2AnyTryOn11.787.460.7659+0.25−0.33+0.0016
CatVTON11.457.070.7715+0.43+0.27−0.0151
OmniTry13.878.320.7501+0.83+0.33−0.0012
RefTON9.667.330.7999−0.23+0.16−0.0051
FitVTON8.627.790.7801+0.34−0.26−0.0036
FitFlow (Ours)5.545.860.8423———

As shown in the Table R4, target-domain adaptation yields mixed changes across different baselines, with improvements in some metrics but degradation in others. FitFlow consistently achieves the best performance across all three metrics, with lower HFE (5.54 vs. 8.62) and SFE (5.86 vs. 7.07), as well as higher IoU (0.8423 vs. 0.7999), compared with the strongest fine-tuned baseline for each metric. We have added these results and corresponding analyses to Appendix D in the revised manuscript.


Q4: Clarification of GSE-Encoded and Original Garment Conditioning Across Inference Timesteps.

A4: We sincerely thank the reviewer for this detailed comment. We would like to clarify that during inference, the GSE-encoded garment is used at high-noise timesteps (t∈(ts,1]t \in (t_s,1]), while the original garment image is used at low-noise timesteps (t∈(0,ts]t \in (0,t_s]).As illustrated in Fig. 3 and Appendix C (Lines 891–893):

In this low-noise interval, the original garment image is used without GSE, since excessive resizing may reduce fine-grained garment details, while structural guidance has already been introduced through the synthetic-guided bridge anchor.

We have explicitly clarified this inference procedure in the revised manuscript.


Q5: Clarification of Whether the FitFlow Results in Table 3 Correspond to the GSE-Only Variant.

A5: We thank the reviewer for this careful comment. We would like to clarify that the "FitFlow (Ours)" row in Table 4 corresponds to the GSE-only variant, as described in the implementation details of Appendix C (Lines 903–906). We have explicitly clarified this inference procedure in the revised manuscript.

Thanks for your response and additional experiments. Most of my concerns have been addressed. I remain original rating 3.

AuthorsOpenReview ↗

Thank You for Your Constructive Feedback

Dear Reviewer BYF2,

We sincerely thank you for your feedback and for acknowledging our additional experiments and clarifications. We are glad that our responses have addressed most of your concerns. We will carefully incorporate the additional experimental results and clarifications into the final version to further strengthen the paper.

We greatly appreciate your constructive suggestions, which have helped us improve the quality and clarity of our work.

Best wishes,

The Authors

Reviewer_WyZ7OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FitFlow learns overall garment layout from synthetic data during the early generation stage and textures and wrinkles from real photographs during later stages. It scales the garment input image according to the ratio of garment length to body height and connects the stages through a training path that corrects synthetic-guided predictions toward real photographic targets.

Critical strengths
  1. It generates garment lengths that better match the specified measurements. It represents the garment-length-to-height ratio through input image scale, reducing hem-position error on the authors’ benchmark.
  2. It learns fit from synthetic data while retaining realistic textures from photographs. It assigns the two data sources to different generation stages and uses a single model at inference without a separate refinement model.
Critical weaknesses
  1. The comparisons supporting the method’s originality appear insufficient. The idea of using synthetic and low-quality data according to noise level has precedent in Ambient Diffusion Omni, while GSE resizes images according to measurement ratios. The technical novelty therefore rests on the specific construction of the training paths, but direct comparisons with relevant prior methods appear to be lacking.
  2. Validation in practical settings is limited. I wonder whether the method maintains its performance on photos taken by ordinary users and on complex poses. Users would likely want to preview garment fit using photos taken with their own phone cameras.
  3. Garment identity preservation during length changes is not evaluated using a dedicated metric. The authors appear to use SSIM and LPIPS to assess this indirectly, but these metrics do not sufficiently establish whether the same garment’s design is preserved across different sleeve-length settings. Although qualitative comparisons show the edited results, a quantitative evaluation is needed to assess how well patterns, colors, logos, buttons, and garment type are preserved.
Questions for authors

Overall, I am pleased with the results achieved in this work. Please refer to the weaknesses section. I would like to see a simple quantitative comparison of garment identity preservation.

4 comments
AuthorsOpenReview ↗

Responses to Reviewer WyZ7 (Part1)

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

Q1: Timestep-Dependent Training and Comparisons with Prior Work.

A1: We thank the reviewer for this valuable comment. We acknowledge that timestep-dependent data utilization has been explored in prior work, including Ambient Diffusion Omni[1]. However, FitFlow differs from this method in the following aspects.

1) Differences from Ambient Diffusion Omni.

  • Motivation. Ambient Diffusion Omni is motivated by the observation that low-quality and high-quality data exhibit similar distributions at high noise levels. In contrast, FitFlow is motivated by the coarse-to-fine generation process of flow matching, where global garment structure is established before local appearance details are refined. This motivates us to exploit the complementary properties of synthetic and real try-on data.
  • Data Characteristics. Ambient Diffusion Omni primarily considers data with quality differences caused by degradations such as blur and compression, while largely preserving the underlying semantic structure. In contrast, FitFlow combines physically simulated synthetic data and real-world try-on data. Both provide valuable supervision, but exhibit distinct domain characteristics: synthetic data provides accurate fitting structures, whereas real data captures realistic appearance details.
  • Training Strategy. Ambient Diffusion Omni employs a classifier to determine the noise levels at which lower-quality data can be incorporated. In contrast, FitFlow assigns synthetic structural supervision to high-noise timesteps and real appearance supervision to low-noise timesteps. Moreover, we explicitly construct a continuous training path connecting the Synthetic Prior and Real Target through the Structural Bridge, addressing the structural inconsistencies that may arise from stage-specific supervision.

2) Additional Comparisons with Prior Methods.

Latent Forcing[2] jointly models latent and pixel representations of the same images using separate noise schedules, while eDiff-I[3] trains specialized denoising experts for different noise intervals and switches between them during inference.

Table 2 in the original manuscript already compares different decoupled training strategies and validates the effectiveness of the Structural Bridge. To further address the reviewer's concern, we conduct two additional experiments:

  • Interleaved Timestep Training (inspired by Latent Forcing): Assigns synthetic and real data to two interleaved, non-overlapping timestep grids spanning the entire noise schedule, without stage-specific supervision.
  • Timestep-Specific Experts (inspired by eDiff-I): Trains separate experts for high- and low-noise intervals and switches between them during inference.

These comparisons investigate whether data mixing or timestep-specific modeling alone can achieve comparable performance. The results further support the effectiveness of FitFlow's cross-domain complementary supervision and Structural Bridge.

Table R1. Quantitative comparison with timestep-dependent training strategies.* KID is reported in units of 10−3. Best results are highlighted in bold.

MethodFID ↓KID ↓SSIM ↑LPIPS ↓HFE(%) ↓SFE(%) ↓IoU ↑
eDiff-I-inspired16.11621.6730.84390.13398.2010.250.7587
Latent Forcing-inspired14.39781.4500.84670.13398.258.640.7873
FitFlow (Ours)13.93000.5160.85260.11525.545.860.8423

These experiments further evaluate the benefits of generation-stage-aware cross-domain supervision and the Structural Bridge, highlighting FitFlow's distinction from existing timestep-dependent training approaches. We have added these results and corresponding analyses to Appendix E in the revised manuscript.


Q2: Generalization to User-Captured Photos and Complex Poses.

A2: We thank the reviewer for this valuable suggestion. We agree that evaluating FitFlow on user-captured photographs and complex poses is important for assessing its practical applicability. To address this concern, we have added qualitative results using smartphone photographs of a volunteer featuring diverse poses and real-world backgrounds, as shown in Fig. 15 of Appendix F in the revised manuscript

AuthorsOpenReview ↗

Responses to Reviewer WyZ7 (Part2)

Q3: Quantitative Evaluation of Garment Identity Preservation, Including Garment Type, Patterns, Colors, Logos, and Buttons.

A3: We thank the reviewer for this valuable suggestion. We agree that conventional VTON metrics, such as SSIM and LPIPS, cannot fully capture fine-grained garment identity preservation. To address this concern, we provide additional quantitative evaluation and clarify two key designs of FitFlow.

1) Quantitative Evaluation of Garment Identity Preservation. Following the reviewer's suggestion, we select 200 garment–person pairs and generate 1,000 try-on images under five length adjustments (-20, -10, 0, +10, and +20 cm). Inspired by Beyond Global Realism [4], we employ GPT-5.6 to evaluate five garment attributes: garment type, pattern, color, logo, and button preservation, using a five-point scale. For each generated image, GPT-5.6 independently compares the try-on result with the original garment reference and assigns a preservation score ranging from 1 (poor preservation) to 5 (excellent preservation). Attributes that are absent from the reference garment or cannot be reliably observed are excluded from the corresponding evaluation. The results are presented in Table R2.

Table R2. Fine-grained garment identity preservation under different length adjustments (↑).

Metric−20 cm−10 cm0 cm+10 cm+20 cm
Garment Type3.723.884.054.013.78
Pattern3.653.873.983.923.69
Color4.014.124.074.144.03
Logo3.883.734.013.773.89
Button3.563.814.034.024.17

As shown in Table R2, FitFlow achieves generally consistent garment identity preservation across different length adjustments, with most scores remaining close to 4.0 even under substantial length changes length changes (±20\pm 20 cm). In particular, color preservation remains stable across all adjustment levels, while garment type and pattern preservation show moderate degradation under extreme length changes. These results demonstrate that FitFlow can achieve continuous garment-length control while largely maintaining fine-grained garment attributes.

2) Garment Identity Preservation in FitFlow. FitFlow preserves garment identity through two complementary designs. First, GSE-based proportional scaling uniformly scales the entire garment within a canonical grid, maintaining the relative proportions and spatial arrangements of garment components, including patterns, logos, and buttons. Second, timestep-dependent supervision (Fig. 3 and Appendix C, Lines 891–893) uses GSE- resized images at high-noise timesteps (t∈(ts,1]t \in (t_s,1]) for structural learning and original images at low-noise timesteps (t∈(0,ts]t \in (0,t_s]) for fine-grained appearance preservation. Together, these designs enable continuous length control while maintaining garment identity.

References

[1] Giannis Daras, Adrian Rodriguez-Munoz, Adam Klivans, Antonio Torralba, and Constantinos Daskalakis. Ambient diffusion omni: Training good models with bad data. arXiv preprint arXiv:2506.10038, 2025.

[2] Alan Baade, Eric Ryan Chan, Kyle Sargent, Changan Chen, Justin Johnson, Ehsan Adeli, and Li Fei-Fei. Latent forcing: Reordering the diffusion trajectory for pixel-space image generation. arXiv preprint arXiv:2602.11401, 2026.

[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. eDiff-I: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022.

[4] Kaidong Zhang, Yukang Ding, Xiaoyu Liu, and Ying Chen. Beyond global realism: Virtual try-on evaluation and optimization with dimension-wise garment fidelity assessment. arXiv preprint arXiv:2608.29804, 2026.

Thank you for the response. My concerns about novelty, practical settings, and garment identity preservation have been addressed. Please include the garment identity evaluation in the final version, together with the scoring prompt/rubric and baseline scores, and add the Ambient Diffusion Omni discussion to the related work. I will keep my score of 3.

AuthorsOpenReview ↗

Thank You for Your Constructive Feedback

Dear Reviewer WyZ7,

We sincerely thank you for your positive feedback and for acknowledging that our additional experiments and clarifications have addressed your concerns. In the final version, we will incorporate the garment identity preservation evaluation, including the scoring prompt, evaluation rubric, and baseline scores. We will also expand the related work section to discuss Ambient Diffusion Omni and further clarify the distinctions from our approach.

We greatly appreciate your constructive suggestions, which have helped us improve the quality and clarity of our paper.

Best wishes,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.