acceptodds
Under review as a conference paper at ICLR 2027

DirectTryOn: One-Step Virtual Try-On via Straightened Conditional Transport

Abstract

Recent diffusion- and flow-based VTON methods achieve strong results with pretrained generative models, but their reliance on multi-step sampling incurs high inference cost, while existing acceleration methods largely overlook the intrinsic structure of the try-on task. In this paper, we highlight a key observation: VTON outputs are highly constrained by the conditional inputs, suggesting that the conditional sampling trajectory can be much straighter than that in general image generation, making one-step generation a natural solution. However, limited task-specific data makes training from scratch impractical, forcing existing methods to fine-tune pretrained models whose objectives do not encourage such straight conditional trajectories. Thus, the deviation from an ideal straight path mainly comes from the mismatch between pretrained base models and the conditional nature of try-on generation, rather than from the task itself. Motivated by this insight, we encourage straighter VTON sampling trajectories through three targeted modifications: pure conditional transport, a garment preservation loss, and a self consistency loss. We further introduce a one-step distillation stage. Extensive experiments show that our method achieves state-of-the-art performance with one-step sampling, establishing a new standard for efficient and high-quality VTON.

Primary area:
generative models
Keywords:
Virtual Try-On (VTON), One-Step Generation, Conditional Transport

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_LXBkOpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DirectTryOn, a two-stage framework for efficient one-step virtual try-on. The authors argue that, unlike general image generation, virtual try-on is strongly constrained by person and garment inputs, resulting in lower conditional uncertainty and potentially straighter transport trajectories. Based on this observation, the first stage fine-tunes a pretrained flow-matching model through pure conditional transport, garment preservation loss, and self-consistency loss to encourage straighter conditional trajectories. The second stage employs Latent Adversarial Diffusion Distillation to obtain a one-step student model. Experiments on VITON-HD and DressCode demonstrate strong image generation quality, with the best FID and KID scores among the evaluated baselines. DirectTryOn achieves approximately 19× faster end-to-end inference than the fastest evaluated multi-step baseline. The authors also provide ablation studies to validate the effectiveness of the proposed training components.

Critical strengths
  • Clear motivation: The paper identifies an important limitation of existing diffusion-based virtual try-on methods: their reliance on computationally expensive multi-step sampling. The observation that strongly conditioned VTON tasks may admit straighter transport trajectories provides an interesting perspective on one-step generation.

  • Reasonable framework design: DirectTryOn combines conditional transport straightening with one-step distillation. The garment preservation and self-consistency objectives are intuitively motivated and designed to address garment fidelity and trajectory consistency.

  • Promising empirical results: The method achieves strong quantitative and qualitative performance on VITON-HD and DressCode while substantially reducing inference time. The component-wise ablations, loss sensitivity analyses, and comparisons with generic acceleration methods provide useful evidence of its effectiveness.

Critical weaknesses
  • W1. Insufficient validation of the theoretical motivation: The paper argues that VTON has lower conditional uncertainty and therefore straighter transport trajectories, making one-step generation more feasible. However, the experiments mainly provide indirect evidence through FID improvements. A more direct analysis of trajectory curvature and its relationship with generation quality would better support this central claim.

  • W2. Limited analysis of sampling efficiency: Although the paper focuses on one-step generation and provides some comparisons with 4-step and 30-step sampling, the trade-off between generation quality and sampling efficiency remains unclear. A more systematic evaluation across different sampling steps would help justify the choice of one-step generation.

  • W3. High training costs and limited generalization: The method requires substantial training resources, including lengthy fine-tuning and an additional distillation stage. Moreover, separate models are trained for VITON-HD and DressCode, leaving cross-dataset generalization unclear. Evaluations on more challenging real-world benchmarks, such as StreetTryOn, would further strengthen the paper.

Questions for authors
  • Q1. Could the authors directly measure the transport trajectory curvature before and after the proposed straightening stage? It would be helpful to analyze whether straighter trajectories consistently lead to better one-step generation quality.

  • Q2. Could the authors provide a more systematic comparison across different sampling steps (e.g., 1, 2, 4, 5, and 10), including both generation quality and inference time? This would help clarify whether one-step generation offers the best quality–efficiency trade-off.

  • Q3. Could the authors evaluate cross-dataset generalization, such as training on VITON-HD and directly testing on DressCode without additional fine-tuning? Additional evaluations on more challenging benchmarks, such as StreetTryOn, would also be valuable.

  • Q4. Given the substantial training costs, could the authors further discuss whether similar one-step performance can be achieved with fewer fine-tuning epochs or reduced training resources?

3 comments

Thank you for your review and constructive suggestions. We address your questions and concerns point by point below.

1. Direct trajectory analysis and one-step quality

Sec. 4.4 and Table 2 now measure actual model-generated ODE trajectories before distillation. SD3 Base uses 160 randomly selected text prompts, while Standard FT and PCT+GL+SC perform try-on on VITON-HD. Compared with SD3 Base, even Standard FT reduces velocity variance from 15,578.17 to 6,699.10 (57.0%) and one-/30-step latent deviation from 239.96 to 67.93 (71.7%), supporting our motivation that try-on admits straighter conditional transport.

At the same 100-epoch budget, PCT+GL+SC further reduces orthogonal deviation from 12.6219 to 12.0997 and one-/30-step latent deviation from 67.93 to 57.10 (15.9%), with lower velocity variance. The path length ratio is unchanged between the two fine-tuned models. Together with the generation-quality ablations (Sec. 4.3; Appendix C, Table 4), these results support the connection between improved trajectory consistency and one-step generation quality.

2. Quality–efficiency trade-off across sampling steps

We evaluated the straightened teacher on VITON-HD with 1, 2, 4, and 10 steps. We also include the final distilled student (Table 1):

ModelStepsFID ↓KID ↓Time (s) ↓
Teacher19.531.680.48
Teacher29.110.970.66
Teacher48.520.541.01
Teacher108.340.472.06
Distilled student18.590.560.48

More teacher steps improve quality at higher latency. Distillation brings one-step quality close to the four-step teacher while reducing latency by 2.1×. This supports one-step generation as an efficient operating point, while additional teacher steps remain useful when quality takes priority.

3. Cross-dataset and in-the-wild generalization

Appendix K, Fig. 8 provides results from the VITON-HD-trained one-step model applied directly to DressCode and StreetTryOn person images, without further fine-tuning.

4. Training cost and fewer fine-tuning epochs

As discussed in Secs. 1 and 3.3, we rely on a pretrained base model for priors on how garments appear on people. However, general-purpose pretraining learns much broader content and does not explicitly encourage VTON's near-straight denoising trajectories. We view this mismatch as a major reason for the extended fine-tuning needed to adapt the model.

A promising direction is to pretrain a smaller base model from scratch specifically on clothed people, then fine-tune it for try-on. We expect this more focused prior to enable faster convergence during adaptation, while the smaller model could reduce fine-tuning and inference costs. This remains a direction for future work.

Thank you for the additional experiments and clarifications. Regarding Q3, could you clarify whether the model trained on VITON-HD was evaluated on the full DressCode test set without additional fine-tuning? The response mainly presents qualitative results; quantitative results would provide stronger evidence of generalization. Overall, I appreciate the framework design and its improvements in generation quality and inference efficiency.

Thank you for your positive assessment of our work and for recognizing the merits of our framework, particularly its improvements in generation quality and inference efficiency.

To clarify, the model trained on VITON-HD was evaluated on DressCode without any additional fine-tuning. Following your suggestion, we have added quantitative evaluation results and a comprehensive analysis of cross-dataset generalization in Appendix K.

We sincerely appreciate your valuable suggestion, which has helped us further strengthen our experimental evaluation.

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

DirectTryOn frames VTON as low-entropy conditional transport: person and garment largely determine the output, so the path is near-straight. Stage 1 (§3.3) fine-tunes an SD3-Medium MMDiT with pure conditional transport (no unconditional samples, no CFG), a latent-MSE garment-preservation loss, a self-consistency loss over arbitrary timestep pairs, and 400 epochs. Stage 2 (§3.4) distills it into a one-step student with off-the-shelf LADD. On VITON-HD and DressCode it reports unpaired FID/KID and clock time, claiming the best 1-NFE quality at ~19× speedup (Table 1).

Critical strengths
  1. Acceleration baselines are controlled, not strawmen: Table 5 distills the same base architecture with DMD (20.04/12.08) and runs 4-step Euler/DDIM (10.77/10.16), all far behind 8.59/0.56, and Appendix C applies the identical LADD recipe to CatVTON (20.19/13.95). That rules out the "LADD-on-anything" explanation.
  2. Better, not just faster: 8.59/0.56 at 1 NFE (0.48 s) vs 9.02/0.63 at 30 NFE (9.24 s) for the best baseline (Table 1), and the teacher alone hits 8.52/0.54 in four steps (Appendix C). The unified clean-fid protocol and Table 6's reproduction breakdown are unusually honest.
Critical weaknesses
  1. Garment fidelity — the property VTON exists for — is never measured. Table 1 uses only FID/KID, blind to whether this garment's texture, logo or printed text survived, while Table 7's SSIM/LPIPS mask the clothing region to gray by construction. The §4.2 claim that garment appearance is "well preserved" thus rests on Figures 5 and 8, while Appendix I concedes complex textures are "imperfectly reconstructed"; even α is tuned by FID (Table 4), blind to that detail. A one-step model can hallucinate plausible-but-wrong garments and still win on FID, so the promise that 1 NFE costs no fidelity is unverified.
  2. The straightening is asserted, not demonstrated, and the novelty is thin. Theorem 1 (Appendix A) is a tautology whose premise (c uniquely determines y) is exactly what §1 says fails for VTON; no straightness statistic appears anywhere, so Fig. 3's "one-step keeps improving to 400 epochs" equally fits training longer (6 days on 16 PPUs, Appendix B). Stripped of labels, the method is drop-CFG/unconditional training (standard few-step practice) + latent MSE + a consistency regularizer close to the cited consistency flow matching (Yang et al., 2024a) + verbatim LADD (§3.4). And +0.43 FID over OOTDiffusion (Table 1) is smaller than the ~3-FID cross-script spread the paper documents itself (Table 6).
Questions for authors
  1. Garment metrics. Which metric distinguishes your one-step output from a 30-step teacher? Please add garment-region SSIM/LPIPS or CLIP-I against the input garment, with OCR/logo checks or a human preference study.
  2. Speed fairness. §4.1 pins all baselines to 30 steps and times CatVTON/Any2anyTryon — natively 384×512 — at 768×1024. Please give FID/KID and time vs. NFE per baseline (and at native resolution) so 19× is a Pareto statement; also state the Table 1 split/sample count.
  3. Straightness. Report velocity variance or trajectory curvature for the SD3 base, standard fine-tuning, and +SC, and show it predicts one-step FID.
3 comments

Thank you for your review and constructive suggestions. We address your questions and concerns point by point below.

1. Garment fidelity

Table 1 now reports garment CLIP-I and a user study, with protocols in Sec. 4.1. SegFormer extracts garment regions for comparison with the input garment. Because extraction can fail, we retain only samples successfully extracted for every compared method: 1,964/2,032 on VITON-HD and 4,885/5,400 on DressCode.

The 20 participants evaluated 50 randomly selected examples per dataset, selecting the result that most accurately reproduces the target garment. DirectTryOn achieves the highest garment CLIP-I (0.9027/0.9051) and garment-fidelity preference (44.0%/60.5%) on VITON-HD/DressCode.

2. Native-resolution speed and evaluation protocol

At their native 384 × 512 resolution, the 30-step baseline results on VITON-HD are:

MethodFID ↓KID ↓Time (s) ↓
CatVTON9.361.194.06
Any2anyTryon9.021.0616.32

DirectTryOn takes 0.48 s at 768 × 1024, approximately 8.5× faster than CatVTON at its native resolution. The original 19× figure refers specifically to the uniform-resolution timing protocol in Sec. 4.1, with the compared baselines run for 30 steps. These are comparisons at the stated settings, rather than a complete quality–latency frontier across all baseline step counts.

Table 1 uses the official unpaired test splits: 2,032 VITON-HD and 5,400 DressCode images for FID/KID. Its quality scores follow the common output-resolution protocol in Sec. 4.1, which differs from the native-resolution scores above. All methods are scored using the same clean-fid pipeline.

3. Trajectory straightness, one-step FID, and motivation

Sec. 4.4 and Table 2 measure actual generated ODE trajectories before distillation. SD3 Base uses 160 randomly selected text prompts, while the fine-tuned models perform try-on. Compared with SD3 Base, even Standard FT reduces velocity variance from 15,578.17 to 6,699.10 (57.0%) and one-/30-step latent deviation from 239.96 to 67.93 (71.7%), supporting our motivation that try-on admits straighter conditional transport.

At the same 100-epoch budget, PCT+GL+SC further reduces orthogonal deviation from 12.6219 to 12.0997, velocity variance from 6699.10 to 6547.49, and one-/30-step deviation from 67.93 to 57.10. The path length ratio is unchanged between the two VTON models. Complementing these trajectory measurements, Appendix C, Table 4 shows that adding SC to PCT+GL improves one-step FID from 15.79 to 11.27 and KID from 6.87 to 3.17 at 100 epochs. These results connect our straightening design with improved one-step quality.

As discussed in Sec. 3.3, the motivation is to adapt the pretrained model to the near-straight transport favored by the strongly constrained try-on task. This motivates consistency across arbitrary timestep pairs. Theorem 1 and Appendix A formalize the ideal conditional case; the trajectory measurements provide empirical support for the practical setting.

4. Ethics and misuse mitigation

We have expanded the Ethics Statement to address non-consensual image manipulation and misuse of garment designs or logos. We recommend consent and image authorization, clear labeling of generated content, and abuse reporting and removal mechanisms.

Dear Reviewer ZRNJ,

We would like to kindly follow up on our response. We would greatly appreciate it if you could take a moment to review our response and share your thoughts when convenient. Thank you for your time and consideration.

Dear Reviewer,

The discussion deadline is approaching. Please review the authors’ responses and clearly indicate whether they adequately address your concerns. Kindly update your score if appropriate.

Best regards,

AC

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

This paper proposes DirectTryOn, a two-stage method for one-step virtual try-on. It argues that the try-on result is largely determined by the person and garment images, so the conditional flow from noise to the output should be close to straight. Stage 1 fine-tunes an SD3-Medium MMDiT model without unconditional training or CFG, adds a garment preservation loss on the garment stream and a self-consistency loss that matches velocity predictions across random timestep pairs, and trains for a long time. Stage 2 distills this teacher into a one-step student with LADD. On VITON-HD and DressCode (unpaired FID/KID), the one-step model scores better than OOTDiffusion, CatVTON, Leffa and Any2anyTryon run at 30 steps, and is about 19× faster than the fastest of them.

Critical strengths
  1. The result is practical. The final model runs in one step (0.48s) and gets the best FID/KID among the evaluated baselines on both VITON-HD and DressCode.
  2. The CFG study is useful in practice. Removing unconditional training and CFG clearly helps one-step generation (Fig. 3 and Fig. 6). The appendix is also open about why the reproduced baseline numbers differ from the original papers.
Critical weaknesses
  1. The core claim mostly restates its assumption. With xt=ty+(1−t)ϵx_t = t y + (1-t)\epsilon, the optimal velocity at t=0t=0 is E[y∣c]−ϵ\mathbb{E}[y|c] - \epsilon, so one Euler step returns the conditional mean E[y∣c]\mathbb{E}[y|c] no matter what the noise is. A "straight path" is therefore the same as assuming that p(y∣c)p(y|c) is narrow. The paper assumes this but never measures it: Fig. 3 only shows that one-step FID improves with longer training, and Fig. 7 follows directly from the identity above. When the distribution is not narrow (wrinkles, occluded parts, tucking), the mean image is blurry and the sharpness comes from the LADD adversarial loss, which makes the method close to a regression-plus-GAN model. One-step GAN-based try-on methods such as HR-VITON and GP-VTON are not compared. Both datasets are studio photos, so the narrow distribution may be a property of the data, and there is no in-the-wild test.
  2. The setup is dated, and the source of the gain is unclear. The model is built on SD3-Medium (2024), while OOTDiffusion and CatVTON use SD1.5, so part of the gain may come from the stronger backbone. Newer backbones such as FLUX.2 or Z-Image already generate in a few steps, which makes going from 30 steps to 1 less important. The key control is missing: the same architecture, data and 400 epochs with standard training, followed by LADD. The DMD baseline (FID 20.04, worse than 4-step Euler) looks under-tuned. Evaluation is limited to unpaired FID/KID, with no paired metrics, no measure of garment detail, no user study, and no comparison with IDM-VTON or FitDiT.
Questions for authors
  1. What do you get with standard training (same architecture, data and 400 epochs) followed by the same LADD distillation?
  2. How does a deterministic one-step model compare, i.e., the same architecture with no noise input, trained with regression plus an adversarial loss?
  3. Can you measure straightness directly, e.g., the distance between one-step and 30-step outputs from the same noise during training?
  4. How does the method work on a newer backbone (FLUX.2 or Z-Image) and on in-the-wild images?
  5. Can you add paired SSIM/LPIPS, a garment-detail metric, and a user study against IDM-VTON and FitDiT?

Clarifications: 6. How does the self-consistency loss differ from Consistency Flow Matching (Yang et al., 2024a), which you cite? 7. How was DMD tuned, given that it is worse than 4-step Euler on the same base model?

4 comments

Thank you for your review and constructive suggestions. We address your questions and concerns point by point below.

1. Standard training + LADD

We ran the requested control with the same architecture, data, 400-epoch teacher training budget, and LADD settings and duration (Sec. 3.4 and Appendix B). Standard training + LADD achieves 10.41/2.08 FID/KID on VITON-HD, compared with 8.59/0.56 for DirectTryOn (Table 1). This isolates the benefit of our teacher-training procedure under matched conditions.

2. Conditional mean and deterministic generation

We agree with the conditional-mean calculation. Theorem 1 (Sec. 3.3; Appendix A) describes the ideal conditionally deterministic case.

As discussed in Sec. 1, our approach relies on pretrained generative priors. Noise-free regression can work for modest refinement, as in TryOn-Refiner (ICCV 2025), where the input already contains a coarse try-on result. We view these limited corrections as a smaller departure from the pretrained model's behavior. Direct try-on requires a much larger change: generating the garment replacement from the original person image. In our earlier experiments, removing noise from the same architecture and training with regression alone caused training collapse.

3. Direct trajectory measurements

Sec. 4.4 and Table 2 provide two comparisons before distillation. First, SD3 Base performs text-to-image generation using 160 randomly selected prompts, while the fine-tuned models perform try-on on VITON-HD. Compared with SD3 Base, even Standard FT reduces velocity variance from 15,578.17 to 6,699.10 (57.0%) and one-/30-step latent deviation from 239.96 to 67.93 (71.7%), supporting our motivation that try-on admits straighter conditional transport.

Second, at the same 100-epoch budget, PCT+GL+SC further reduces orthogonal trajectory deviation from 12.6219 to 12.0997 and one-/30-step latent deviation from 67.93 to 57.10 (15.9%). The path length ratio remains unchanged between the two fine-tuned models. This comparison shows how our training further aligns the model with the task's conditional structure.

4. Newer backbones and in-the-wild evaluation

Cross-dataset and in-the-wild experiments, together with the generalization discussion, are provided in Appendix K, Fig. 8. The training principles in Sec. 3.3 are transferable in principle to other conditional flow backbones, with the garment loss requiring a suitable garment representation.

5. Garment fidelity, user study, and additional comparisons

Table 1 now includes garment CLIP-I and human preference, with protocols in Sec. 4.1 and discussion in Sec. 4.2. We use SegFormer garment crops and retain only examples where extraction succeeds for every compared method. The user study involves 20 participants and 50 randomly selected examples per dataset. DirectTryOn obtains CLIP-I of 0.9027/0.9051 and preference shares of 44.0%/60.5% on VITON-HD/DressCode, respectively, the highest among the compared methods.

For paired SSIM/LPIPS, our unmasked person input already contains the target garment, allowing input preservation to score well without testing replacement (paired/unpaired protocol: Sec. 4.1). Unpaired replacement has no aligned ground truth, so the added garment metric and user study evaluate this setting directly.

The added IDM-VTON evaluation gives FID/KID of 9.10/1.06 on VITON-HD and 5.51/1.42 on DressCode, compared with our 8.59/0.56 and 5.08/0.95 (Table 1).

6. Difference from Consistency Flow Matching

As explained in Sec. 3.3, our starting point is that the strongly constrained try-on task already favors near-straight conditional transport. Our consistency loss helps the pretrained model reflect this task-inherent property. Consistency Flow Matching, by contrast, imposes straightness through consistency training for general generation, without this near-deterministic conditioning. This distinction motivates our more aggressive design: we enforce velocity consistency across arbitrary timestep pairs, whereas CFM uses nearby timesteps. Appendix C, Table 3 supports our design: arbitrary pairs yield FID 11.27, versus 14.15 with a fixed interval and 13.26 without stop-gradient.

7. DMD settings and tuning

Appendix F and Table 6 provide the configuration and results. We distill the same conventionally trained base model for 40 epochs at 1024 × 768, using AdamW with learning rate 1e-6, batch size 24, CFG 2.0, and one generator update per five fake-score updates.

Dear Reviewer gkQq,

We would like to kindly follow up on our response. We would greatly appreciate it if you could take a moment to review our response and share your thoughts when convenient. Thank you for your time and consideration.

Reviewer_gkQqOpenReview ↗

Response to rebuttal: rating maintained

I thank the authors for the response and the additional experiments. The standard-training + LADD control (10.41/2.08 vs. 8.59/0.56) is the most useful addition: it shows that the teacher-training recipe matters. The garment CLIP-I scores and the user study are also welcome.

However, I will keep my score, for three reasons.

  1. The new trajectory measurements do not support the central claim. Table 2 shows that most of the straightening comes from the task itself: standard fine-tuning already reduces the path-length ratio from 1.019 to 1.005. The proposed training leaves this ratio unchanged (1.00497 for both), and its effect on orthogonal deviation and velocity variance is within one standard deviation. Meanwhile, with the same training budget and LADD, the proposed recipe lowers FID from 10.41 to 8.59. This gain is therefore hard to attribute to "straightened conditional transport". It is consistent with my original point: a straight path is equivalent to a narrow p(y|c), which is a property of the data rather than something the method creates.

  2. Several requested comparisons are still missing. The regression + adversarial baseline (Q2) was not run. A loss that "collapses with regression alone" does not answer the question. HR-VITON, GP-VTON and FitDiT are not compared. DMD was trained for 40 epochs versus 400 for the proposed teacher, so it remains under-tuned relative to the main method.

  3. Generalization and garment fidelity remain weakly supported. Cross-dataset transfer degrades FID from 11.12 to 16.49, and the in-the-wild evidence consists of two qualitative examples. The garment CLIP-I margin over Leffa (0.9027 vs. 0.9009) is small, and CLIP features are largely insensitive to logos and printed text.

Overall, DirectTryOn is a practical recipe that combines dropping CFG, long fine-tuning, a consistency regularizer and LADD. However, the paper's main insight, that try-on admits straighter transport, follows from the conditioning itself, and the new evidence does not show that the proposed training is what produces it. I maintain my rating of weak reject.

Minor: please include the standard-training + LADD control and the IDM-VTON results in the final version of the paper.

We sincerely appreciate your thoughtful response and the time and effort you have devoted to reviewing our work. While we respect your perspective, we respectfully but firmly disagree with your remaining concerns and would like to clarify our position on each of them.

1. Novelty and the evidence for straightened conditional transport.

We would like to emphasize that our main contribution is to formulate the virtual try-on task as a low-entropy conditional transport problem (Lines 99–100), which, to the best of our knowledge, has not been discussed in prior work. We do not claim to have created this property. Rather, all our proposed modifications are designed to exploit this property and help the pretrained model (SD3) better realize it (Lines 76–80).

Furthermore, as shown in Table 2, our PCT+GL+SC model consistently improves the trajectory measurements over Standard FT. Specifically, Velocity Variance decreases from 6699.10 ± 1034.68 to 6547.49 ± 975.64, One-/30-Step Deviation decreases from 67.93 ± 10.55 to 57.10 ± 9.75, One-/30-Step RMSE decreases from 0.15320 ± 0.02378 to 0.12877 ± 0.02200, and Orthogonal Deviation decreases from 12.6219 ± 1.7290 to 12.0997 ± 1.6893. These measurements collectively provide quantitative evidence that our proposed PCT+GL+SC further improves the straightness of the conditional transport trajectory. In particular, our first-stage model is trained entirely under the flow-matching framework and achieves one-step denoising. If the denoising trajectory were not effectively straightened, it would be difficult to explain the substantial FID improvement from 22.44 to 9.53.

2. Baselines and comparison fairness.

As explained in our previous rebuttal, we have already explored directly training the try-on task using a regression loss. However, this approach causes the model to deviate too far from the pretrained model and eventually collapse, regardless of whether mask-based or mask-free training is used. Therefore, we have strong reasons to believe that a regression-based method with adversarial loss would not provide a viable baseline.

Regarding FitDiT, it is trained jointly on VITON-HD and DressCode, whereas all the compared methods in our paper are trained and evaluated separately on each dataset. Therefore, including FitDiT would not constitute a fair comparison under our experimental setting. As for HR-VITON and GP-VTON, they are relatively outdated methods, and we do not expect them to outperform the more recent baselines considered in our paper.

The DMD variant is specifically implemented to address the question of whether standard training combined with a general one-step distillation method can achieve comparable performance (Lines 964–967). Moreover, our LADD model is also trained for only 40 epochs. Therefore, we consider this comparison fair.

3. Generalization and garment fidelity evaluation.

We believe that the performance gap between in-distribution and out-of-distribution data under one-step denoising is acceptable, as already discussed in Lines 1115–1122. Regarding garment fidelity, considering that our one-step method still slightly outperforms the 30-step Leffa, we believe this provides further evidence of the effectiveness of our approach.

Regarding the evaluation metrics, since garments can undergo deformation in try-on results, pixel-level metrics such as SSIM are not meaningful for evaluating garment fidelity. Therefore, we adopt CLIP-I, as suggested by Reviewer ZRNJ.

Minor Comment. We will include the standard-training + LADD control and the IDM-VTON results in the final version of the paper.

Finally, as you pointed out in your first concern, we believe that our results have already demonstrated that the try-on task itself admits straighter conditional transport. This is precisely our central claim and the primary contribution of this work. Thank you again for your response, which gives us another opportunity to further clarify our method and contributions.

Comments on the paper

1 comment
Area_Chair_fuewOpenReview ↗

Reminder: Discussion Deadline — November 18

Dear Reviewers,

Thank you for your valuable time and contributions to ICLR. This is a friendly reminder to review the authors' responses and engage in discussions with them. The discussion deadline is November 18.

Thank you for your continued efforts!

ACs

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.