Pith. sign in

REVIEW 3 major objections 5 minor 60 references

Virtual try-on can now add, remove, and swap clothing layers while preserving what is underneath, by splitting training into two stages: learning general try-on skills from videos, then learning layering logic from a small real dataset.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 04:07 UTC pith:V5LDMR7D

load-bearing objection The task definition and two-stage pipeline are genuinely new and the ablations back the story, but the private LVTON benchmark's train/test overlap risk makes the headline SOTA claim uninterpretable until the split is clarified. the 3 major comments →

arxiv 2607.22924 v1 pith:V5LDMR7D submitted 2026-07-24 cs.CV

Layering Virtual Try-On

Chun Feng , Bowei Chen , Mengyi Shan , Ira Kemelmacher-Shlizerman This is my paper
classification cs.CV
keywords virtual try-onlayeringgarment editingimage editingdiffusion modelsdata generationvideo miningpose misalignment
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces Layering Virtual Try-On (LVTON): instead of replacing a single garment, the model takes a person image, a garment image, and an instruction ("add", "remove", "swap") and edits only the targeted layer, keeping inner layers visible and realistic. The authors argue that standard mask-based try-on methods cannot do this because masking erases the inner outfit and single-item datasets contain no layered examples. They propose a two-stage paradigm: first learn general deformation and identity-preservation priors from automatically synthesized, pose-misaligned pairs mined from fashion videos; then fine-tune on a small real-world layering dataset to learn occlusion and composition logic. They report state-of-the-art results on their new benchmark and on traditional single-garment benchmarks, with zero-shot transfer from stage one. A sympathetic reader would care because sequential outfit editing—adding a jacket over a shirt, then removing it—is the common real-world use case e-commerce and augmented-reality try-on need.

Core claim

The paper claims that the failure of direct fine-tuning on scarce, pose-variant layering data is not a data-quantity problem but a task-decomposition problem. By disentangling LVTON into general VTON priors and specific layering knowledge, the model can acquire deformation and identity skills from abundant videos (29,151 synthetic pairs from 362 videos) and then learn add/remove/swap logic from only 5,768 real training pairs. Temporal reversal—reversing a dressing sequence into an undressing sequence—doubles the real data and teaches removal. The resulting model outperforms prior mask-free and commercial baselines on the LVTON benchmark and, after task-specific fine-tuning, achieves the best

What carries the argument

The central mechanism is a mask-free, pose-misaligned training-pair construction from videos. For each video with a constant outfit, frames are segmented and re-clothed by an inpainting model; training pairs pair one inpainted frame (source person) with a different frame (target pose and garment). Because the garment source and target pose come from different frames, the model must learn true deformation and placement rather than pixel copying. Stage 2 adds a temporal-reversal augmentation over real layering videos to teach addition and removal, and both stages fine-tune the same pre-trained image-editing diffusion transformer with low-rank adapters.

Load-bearing premise

The results assume the test set is fully disjoint from the training set at the level of people and outfits; the paper only says both come from the same 60 videos and never states that subjects or clips were separated.

What would settle it

Compute overlap between stage-2 training and test frames using face or outfit-level identity matching across all 5,768 training pairs and 532 test pairs; if any test frame shares a subject or outfit with a training pair, the reported LVTON gains are inflated. Re-splitting the data at the video level and rerunning the main comparison would settle whether the margin survives.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Sequential edits become possible: add, remove, or swap one layer at a time while inner layers remain, enabling an outfit-customization loop.
  • The stage-one video-mining recipe provides a scalable way to obtain pose-variant, mask-free try-on training data without manual pairing, attacking the data bottleneck for any mask-free try-on task.
  • Because stage-one priors transfer zero-shot to standard benchmarks, a single general try-on model can serve both layering and classic single-garment replacement, reducing the need for task-specific datasets.
  • Fine-tuning on only a few thousand real pairs suffices once general priors are in place, suggesting data-efficient deployment for new garment interactions.
  • Reversal augmentation teaches removal from addition-only footage, effectively doubling available supervision for undressing operations.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the train/test split of the new benchmark is not separated at video or subject level, the headline LVTON gains could partly reflect memorization of the same outfits seen in fine-tuning; a held-out-subject test would be a sharper check.
  • The decomposition suggests a testable scaling law: performance should saturate once stage-one data supplies rich deformation priors, so adding more synthetic videos has diminishing returns—already hinted by the 4x saturation point in the paper's ablation.
  • Because stage-one artifacts come from upstream segmentation and inpainting, automated filtering or higher-quality synthesis tools could directly improve the priors without changing the two-stage architecture.
  • The purely 2D model lacks explicit geometric reasoning, so extending the same two-stage logic to multi-view or 3D layered representations is a natural next step, though it would require volumetric constraints not present here.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper introduces Layering Virtual Try-On (LVTON), a task formulation, a private benchmark, and a two-stage training method. Stage 1 learns general VTON priors from a large synthetic dataset generated automatically from fashion videos via segmentation and inpainting, with pose-variant training pairs. Stage 2 fine-tunes the model on a small real-world layering dataset, augmented by temporal reversal, to teach add/remove/swap composition logic. The authors report state-of-the-art results on their LVTON benchmark, state-of-the-art results on VITON-HD and DressCode when fine-tuned, and zero-shot transfer for the stage-1 model. They also propose a Masked Preservation Score (MPS) for layer-specific evaluation and provide extensive ablations and qualitative results.

Significance. If the claims hold, this is a meaningful contribution: the LVTON task is practically important, the two-stage disentanglement is a sensible and data-efficient strategy, and the video-based synthetic data pipeline is scalable and well described. The ablations in Table 4 support the central architectural intuition, and the public-benchmark results (Table 2) provide independent evidence that the learned priors transfer. The paper is generally well written and includes candid limitations in Sec. 5 and Appendices D, F, and G. However, the headline LVTON state-of-the-art claim rests on a private benchmark whose train/test split is not demonstrably clean, and a key hyperparameter is selected on that same benchmark. These issues are fixable but currently prevent full acceptance.

major comments (3)
  1. [Sec. 4.1, App. B.2/C] The LVTON benchmark split is not specified with respect to subjects, outfits, or videos. The paper states that 5,768 training pairs and 532 test pairs come from the same 60 videos, and Appendix B.2 describes temporal clustering and frame sampling within videos, but nowhere states that test videos, subjects, outfits, or temporal clusters are disjoint from training. If a test frame shares a person, outfit, or video with a training pair, Table 1 and the user study measure near-duplicate reconstruction or per-subject recall rather than layering generalization. Please specify the exact split protocol at the video/subject/outfit level, and ideally release the benchmark or identity metadata. Also report confidence intervals or bootstrapped error bars for the 532-image test set; the current table has none.
  2. [Fig. 9, Sec. 4.5] The 4.0x stage-1/stage-2 data ratio is selected using the LVTON benchmark itself, and the paper does not describe a separate validation split. Because the same test set appears to have influenced this hyperparameter, the reported LVTON SOTA numbers are optimistically biased. Please clarify whether a held-out validation set was used, or re-run the comparison with a proper train/validation/test partitioning and report the selected ratio on the validation set.
  3. [Appendix E, Sec. 4.3] The MPS metric's LPIPS normalization is not standard. Zeroing out non-mask regions and scaling the global average by H*W/|Munchanged| can bias the result because zero-padding affects convolutional receptive fields and VGG feature statistics. The reported improvement (SSIM +0.005, LPIPS -0.006) is small and no significance testing is provided. Please provide a more principled masked-LPIPS formulation or justify the normalization carefully, and report confidence intervals for the MPS comparison.
minor comments (5)
  1. [Table 3, Sec. 4.3] The user study used 30 participants. Please report recruitment details, rating scales, and inter-rater agreement or variance.
  2. [Fig. 9] The figure legend is unclear; it is difficult to tell which curve corresponds to SSIM and which to LPIPS. Please add explicit labels or a legend.
  3. [Appendix A] Training step counts and the single-GPU setup are useful. Please also report wall-clock training time and approximate total GPU hours.
  4. [Sec. 5, App. G] The limitation that the stage-2 dataset contains only female subjects in indoor scenes is important and should be stated earlier in the paper, as it tempers the in-the-wild generalization claim in Sec. 4.4.
  5. [Sec. 4.2, Table A1] The 'Previous SOTA' row in Table A1 lacks a citation; please specify which method it refers to.

Circularity Check

2 steps flagged

LVTON SOTA claim is partially self-referential: the 4.0x stage-1/stage-2 ratio is chosen on the LVTON benchmark itself, and the 532 test pairs come from the same 60 videos as the 5,768 training pairs with no stated split. External VITON-HD/DressCode results keep the core method independently grounded.

specific steps
  1. fitted input called prediction [Sec. 4.5, Fig. 9; Tab. 1]
    "Ablation study on stage 1 data scale using the LVTON benchmark. The curves illustrate SSIM and LPIPS as a function of the data ratio between stage 1 and stage 2. ... We adopt the saturation point of 4.0x (green star) as the optimal configuration."

    The 4.0x stage-1/stage-2 data ratio is selected by evaluating on the LVTON benchmark, and the final model ('Ours') is then evaluated on the same benchmark in Tab. 1. Thus the headline LVTON SOTA is obtained with a hyperparameter fitted to the test set; the result is partly forced by construction rather than an independent prediction.

  2. other [Sec. 4.1; Sec. 3.2; Appendix C]
    "we curate a real-world LVTON dataset of 5,768 training pairs and 532 test pairs from 60 videos ... We partition video frames into temporal clusters {C1, C2, ..., CK} ... we sample a source frame Fi in Cm and a target frame Fj in Cm+1 from adjacent clusters to form the training tuple."

    Training and test pairs are described as coming from the same 60 videos and are constructed by the same clustering/sampling procedure. The paper never states that test videos, subjects, outfits, or temporal clusters are disjoint from training. If a test frame shares a person/outfit/video with a training pair, the LVTON benchmark measures near-duplicate reconstruction or memorization of stage-2 data rather than layering generalization, so the claimed SOTA on LVTON partially reduces to fitting the training videos.

full rationale

The two-stage derivation itself is not circular: stage 1 builds mask-free, pose-variant synthetic pairs from videos and inpainting; stage 2 fine-tunes on real layering pairs; ablations and zero-shot evaluation on VITON-HD/DressCode provide independent evidence that the stage-1 priors transfer. No load-bearing self-citation or imported uniqueness theorem is present. However, the headline 'state-of-the-art on our LVTON benchmark' is weakened by two evaluation-level circularities: the key data-ratio hyperparameter is selected on the LVTON benchmark itself, and the benchmark's test pairs are drawn from the same 60 videos as the stage-2 training pairs with no stated split. These issues affect the self-benchmarked claim but not the external-benchmark results, so the circularity is partial rather than total.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 2 invented entities

The method rests on a chain of domain assumptions about video data, upstream foundation models, and self-evaluation; no external benchmark or released artifacts independently verify the central claim.

free parameters (5)
  • Stage 1 to Stage 2 data ratio = 4.0x (green star)
    Selected as the saturation point from an ablation on the LVTON benchmark (Fig. 9), i.e., tuned on the same evaluation set used for headline results.
  • LoRA rank and alpha = r=32, alpha=32
    Fixed by hand for all experiments (Appendix A); affects model capacity but is not fitted to a specific outcome.
  • Upper/lower body segmentation ratio = 0.7 / 0.3
    Heuristic split prioritizing upper-body layering when generating synthetic stage-1 pairs (Sec. B.1).
  • VLM sliding-window size W = 6 frames
    Hand-set window for detecting outfit-change transitions in stage-2 videos (Sec. B.2).
  • Blur-detection Laplacian threshold = unspecified ("predefined threshold")
    Filters frames by Laplacian variance (Sec. B.1); the exact threshold is not reported, so data filtering is not exactly reproducible.
axioms (6)
  • domain assumption Fashion videos contain long stretches where a single subject wears the same outfit while changing pose.
    Stage 1 constructs (F'_i, G_j, P_j, T, F_j) from such videos (Sec. 3.1, B.1); if outfit or pose variability is insufficient, the synthetic prior collapses.
  • domain assumption The VLM sliding-window detector correctly partitions real videos into stable outfit clusters.
    Stage 2 training tuples depend on detected transitions C_m -> C_{m+1} (Sec. B.2); detection errors mislabel add/remove/swap supervision.
  • domain assumption SAM2 segmentation and the Flux Kontext inpainting model are accurate enough that noisy synthetic pairs can be filtered by hand without biasing the learned prior.
    The paper acknowledges artifacts and relies on aggressive discarding (Sec. B.1, Fig. A1); the filtering is manual and its quality is not quantified.
  • domain assumption Stage 2 fine-tuning on a small real dataset removes the synthetic-to-real domain gap introduced by Stage 1.
    Argued in Sec. B.1 and supported by Table A1 (KID drops from 1.550 to 0.200), but not proven for all garment types and layering interactions.
  • standard math Standard denoising objective and LoRA fine-tuning of Qwen-Image-Edit preserve pre-trained editing ability while learning the new task.
    Training details are in Sec. 3.3 and Appendix A; standard diffusion and LoRA assumptions are used without formal guarantees.
  • domain assumption FID/KID on a 532-image test set is a meaningful comparison even though the paper says absolute FID is inflated.
    Table 1 note; the small test set makes FID noisy, so SOTA comparisons may be unstable.
invented entities (2)
  • LVTON benchmark (5,768 train / 532 test pairs from 60 videos) no independent evidence
    purpose: Defines the new layering task and serves as the primary evaluation set for the SOTA claim.
    No public release, no external benchmark to cross-check, and train/test pairs come from the same 60 videos with no stated subject-level split (Sec. 4.1).
  • Masked Preservation Score (MPS) no independent evidence
    purpose: Quantifies preservation of non-target regions in layering edits.
    Depends on segmentation masks the authors acknowledge can be inaccurate (Appendix E); no external validation is provided.

pith-pipeline@v1.3.0-alltime-deepseek · 21161 in / 13986 out tokens · 132789 ms · 2026-08-01T04:07:44.476968+00:00 · methodology

0 comments
read the original abstract

In the real world, fashion is about layering: adding a jacket over a shirt, or a sequence of adding and removing layers, rather than just a single-layer swap. This fundamental real-world task remains a challenge in existing Virtual Try-On (VTON) methods, which excel at single-layer replacement but are not designed to layer or de-layer an existing outfit. This paper proposes Layering Virtual Try-On (LVTON), a layering benchmark and method that preserves an existing outfit while enabling sequential layering. We find that current VTON paradigms are fundamentally ill-equipped for LVTON, as their reliance on cloth-agnostic representations and single-item datasets discards essential layering context. Our key insight is that the LVTON challenge must be disentangled into two distinct competencies: (1) General VTON Priors (e.g., deformation, identity preservation) and (2) Specific Layering Knowledge (e.g., layering order and occlusion reasoning). First, our model obtains general VTON priors by being trained on data produced by an automatic data generation pipeline that synthesizes samples from fashion videos via segmentation and inpainting. Second, the model is fine-tuned on a small, dedicated LVTON dataset to learn the layering logic. Our method achieves state-of-the-art results on our LVTON benchmark and demonstrates superior generalizability on traditional VTON benchmarks, setting new state-of-the-art results when fine-tuned and exhibiting zero-shot capabilities.

Figures

Figures reproduced from arXiv: 2607.22924 by Bowei Chen, Chun Feng, Ira Kemelmacher-Shlizerman, Mengyi Shan.

Figure 1
Figure 1. Figure 1: We propose a Layering Virtual Try-On (LVTON) method that allows for swapping, adding, and removing clothing layers on any given person image. Given a person image, a garment image, and a textual instruction (e.g., swap, add, or remove), it generates a try-on image reflecting the desired garment edit while preserving inner layers when necessary. We showcase three sequential editing examples that exhibit nat… view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of the Layering VTON (LVTON) challenge and the motivation for our two-stage paradigm. (a) LVTON task aims to composite a new garment (e.g., a sweater vest) over an existing outfit while preserving inner layers. (b) Traditional VTON paradigms mainly rely on a cloth-agnostic representation by masking out the original garment, which discards the inner layer. (c) The direct training dilemma: a mod… view at source ↗
Figure 3
Figure 3. Figure 3: Our two-stage pipeline. (a) The first stage builds general VTON priors by synthesizing mask-free, pose-mismatched training pairs from videos. This forces the model to learn robust spatial deformation, ensuring accurate alignment even during standard fixed-pose inference. (b) The second stage uses a small, real-world layering dataset to teach the model specific layering knowledge (e.g., “add”). instance is … view at source ↗
Figure 4
Figure 4. Figure 4: Visualizations of fine-tuning data. Top: A real-world “add” sequence collected from video. Bottom: The corresponding “remove” sequence generated by temporal reversal augmentation. 3.2 Stage 2: Fine-tuning for Layering Knowledge While stage 1 equips the model with strong general VTON priors, it lacks supervision over explicit compositional operations (e.g., “add”). We therefore fine-tune the model on a real… view at source ↗
Figure 5
Figure 5. Figure 5: An overview of model architecture. We apply Low-Rank Adaptation (LoRA) to fine-tune Qwen-Image-Edit. The model conditions on four inputs: (1) a text prompt (e.g., “add a light gray turtleneck sweater”), processed by the Qwen2.5-VL; (2) a noised person image latent, created by adding noise at timestep t; (3) a garment image latent; and (4) a pose image latent. image tokenizer, (2) Qwen2.5-VL [2] as the mult… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative results of LVTON. Our method successfully composites the target garment while preserving the inner layer. Baselines often erase inner details or introduce artifacts. Regions corresponding to the errors made by the baselines are highlighted with red boxes in their respective predictions for comparison (e.g., Nano Banana fails to preserve the lower portion of the black dress). 4.1 Layering Virtua… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison on the VITON-HD [9] and DressCode [38] benchmarks. Ours represents our complete two-stage model. Stage 1-only represents our stage 1 model evaluated zero-shot (analysis in Sec. 4.5). Red boxes highlight artifacts and failure cases in the baseline results. Ours consistently avoids these issues, demonstrating superior fidelity and generalization. 4.2 Traditional Virtual Try-On Datasets… view at source ↗
Figure 8
Figure 8. Figure 8: In-the-wild layering editing. Despite being fine-tuned on a strictly limited stage 2 dataset, our model successfully generalizes to complex, out-of-distribution scenarios sourced from the internet. It accurately executes specific compositional instructions (“add,” “swap,” and inner-layer swaps) while naturally preserving the visual integrity of the non-target layers and overall high image fidelity. While o… view at source ↗
Figure 9
Figure 9. Figure 9: Ablation study on stage 1 data scale using the LVTON benchmark. The curves illustrate SSIM and LPIPS as a function of the data ratio between stage 1 and stage 2. Increasing the ratio yields consistent improvements, validating the efficacy of stage 1 priors. We adopt the saturation point of 4.0× (green star) as the optimal configuration. Validation of general VTON priors. To confirm that stage 1 provides ro… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

60 extracted references · 24 linked inside Pith

  1. [1]

    In: Proceedings of the Asian Conference on Computer Vision

    Aggarwal, A., Wang, J., Hogue, S., Ni, S., Budagavi, M., Guo, X.: Layered-garment net: Generating multiple implicit garment layers from a single image. In: Proceedings of the Asian Conference on Computer Vision. pp. 3000–3017 (2022)

    2022

  2. [2]

    5-vl technical report

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., et al.: Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923 (2025)

    Pith/arXiv arXiv 2025

  3. [3]

    arXiv preprint arXiv:2408.07009 (2024)

    Baldridge, J., Bauer, J., Bhutani, M., Brichtova, N., Bunner, A., Castrejon, L., Chan, K., Chen, Y., Dieleman, S., Du, Y., et al.: Imagen 3. arXiv preprint arXiv:2408.07009 (2024)

    arXiv 2024

  4. [4]

    1 kontext: Flow matching for in- context image generation and editing in latent space

    Batifol, S., Blattmann, A., Boesel, F., Consul, S., Diagne, C., Dockhorn, T., English, J., English, Z., Esser, P., Kulal, S., et al.: Flux. 1 kontext: Flow matching for in- context image generation and editing in latent space. arXiv e-prints pp. arXiv–2506 (2025)

    2025

  5. [5]

    arXiv preprint arXiv:1801.01401 (2018)

    Bińkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying mmd gans. arXiv preprint arXiv:1801.01401 (2018)

    Pith/arXiv arXiv 2018

  6. [6]

    In: Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval

    Bonifacio, L., Abonizio, H., Fadaee, M., Nogueira, R.: Inpars: Unsupervised dataset generation for information retrieval. In: Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. pp. 2387–2392 (2022)

    2022

  7. [7]

    arXiv preprint arXiv:2509.23951 (2025)

    Cao, S., Chen, H., Chen, P., Cheng, Y., Cui, Y., Deng, X., Dong, Y., Gong, K., Gu, T., Gu, X., et al.: Hunyuanimage 3.0 technical report. arXiv preprint arXiv:2509.23951 (2025)

    Pith/arXiv arXiv 2025

  8. [8]

    In: European Conference on Computer Vision

    Chen, M., Chen, X., Zhai, Z., Ju, C., Hong, X., Lan, J., Xiao, S.: Wear-any-way: Manipulable virtual try-on via sparse correspondence alignment. In: European Conference on Computer Vision. pp. 124–142. Springer (2024)

    2024

  9. [9]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Choi, S., Park, S., Lee, M., Choo, J.: Viton-hd: High-resolution virtual try-on via misalignment-aware normalization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14131–14140 (2021)

    2021

  10. [10]

    In: European Conference on Computer Vision

    Choi, Y., Kwak, S., Lee, K., Choi, H., Shin, J.: Improving diffusion models for authentic virtual try-on in the wild. In: European Conference on Computer Vision. pp. 206–235. Springer (2024)

    2024

  11. [11]

    arXiv preprint arXiv:2407.15886 (2024)

    Chong, Z., Dong, X., Li, H., Zhang, S., Zhang, W., Zhang, X., Zhao, H., Jiang, D., Liang, X.: Catvton: Concatenation is all you need for virtual try-on with diffusion models. arXiv preprint arXiv:2407.15886 (2024)

    Pith/arXiv arXiv 2024

  12. [12]

    arXiv preprint arXiv:2507.06261 (2025)

    Comanici, G., Bieber, E., Schaekermann, M., Pasupat, I., Sachdeva, N., Dhillon, I., Blistein, M., Ram, O., Zhang, D., Rosen, E., et al.: Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261 (2025)

    Pith/arXiv arXiv 2025

  13. [13]

    arXiv preprint arXiv:2508.08488 (2025)

    Deria, A., Mahapatra, D., Bozorgtabar, B., Chakraborty, M., Chakraborty, S., Roy, S.: Muga-vton: Multi-garment virtual try-on via diffusion transformers with prompt customization. arXiv preprint arXiv:2508.08488 (2025)

    Pith/arXiv arXiv 2025

  14. [14]

    IEEE Access11, 47304–47320 (2023)

    Eigenschink, P., Reutterer, T., Vamosi, S., Vamosi, R., Sun, C., Kalcher, K.: Deep generative models for synthetic data: A survey. IEEE Access11, 47304–47320 (2023)

    2023

  15. [15]

    In: Forty-first international conference on machine learning (2024) 16 C

    Esser, P., Kulal, S., Blattmann, A., Entezari, R., Müller, J., Saini, H., Levi, Y., Lorenz, D., Sauer, A., Boesel, F., et al.: Scaling rectified flow transformers for high-resolution image synthesis. In: Forty-first international conference on machine learning (2024) 16 C. Feng et al

    2024

  16. [16]

    Fan, L., Chen, K., Krishnan, D., Katabi, D., Isola, P., Tian, Y.: Scaling laws of synthetic images for model training... for now. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7382–7392 (2024)

    2024

  17. [17]

    arXiv preprint arXiv:2508.13632 (2025)

    Feng, Y., Zhang, L., Cao, H., Chen, Y., Feng, X., Cao, J., Wu, Y., Wang, B.: Omnitry: Virtual try-on anything without masks. arXiv preprint arXiv:2508.13632 (2025)

    Pith/arXiv arXiv 2025

  18. [18]

    arXiv preprint arXiv:2403.12803 (2024)

    Fu, Y., Chen, C., Qiao, Y., Yu, Y.: Dreamda: Generative data augmentation with diffusion models. arXiv preprint arXiv:2403.12803 (2024)

    Pith/arXiv arXiv 2024

  19. [19]

    In: Proceedings of the 31st ACM International Conference on Multimedia

    Gou, J., Sun, S., Zhang, J., Si, J., Qian, C., Zhang, L.: Taming the power of diffusion models for high-quality virtual try-on with appearance flow. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 7599–7607 (2023)

    2023

  20. [20]

    arXiv preprint arXiv:2501.15891 (2025)

    Guo, H., Zeng, B., Song, Y., Zhang, W., Zhang, C., Liu, J.: Any2anytryon: Leverag- ing adaptive position embeddings for versatile virtual clothing tasks. arXiv preprint arXiv:2501.15891 (2025)

    Pith/arXiv arXiv 2025

  21. [21]

    Hammoud, H.A.A.K., Itani,H., Pizzati, F.,Torr, P., Bibi, A., Ghanem,B.: Synthclip: Are we ready for a fully synthetic clip training? arXiv preprint arXiv:2402.01832 (2024)

    Pith/arXiv arXiv 2024

  22. [22]

    Advances in neural information processing systems30(2017)

    Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems30(2017)

    2017

  23. [23]

    In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)

    Honovich, O., Scialom, T., Levy, O., Schick, T.: Unnatural instructions: Tuning language models with (almost) no human labor. In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 14409–14428 (2023)

    2023

  24. [24]

    ICLR1(2), 3 (2022)

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. ICLR1(2), 3 (2022)

    2022

  25. [25]

    In: ACM SIGGRAPH 2024 conference papers

    Huang, B., Yu, Z., Chen, A., Geiger, A., Gao, S.: 2d gaussian splatting for geomet- rically accurate radiance fields. In: ACM SIGGRAPH 2024 conference papers. pp. 1–11 (2024)

    2024

  26. [26]

    Huang, Y., Zhang, P., Liu, R., Liang, J.: Can generated images serve as a viable modality for text-centric multimodal learning? arXiv preprint arXiv:2506.17623 (2025)

    arXiv 2025

  27. [27]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Jafarian, Y., Park, H.S.: Learning high fidelity depths of dressed humans by watching social media dance videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12753–12762 (2021)

    2021

  28. [28]

    arXiv preprint arXiv:2411.10499 (2024)

    Jiang, B., Hu, X., Luo, D., He, Q., Xu, C., Peng, J., Zhang, J., Wang, C., Wu, Y., Fu, Y.: Fitdit: Advancing the authentic garment details for high-fidelity virtual try-on. arXiv preprint arXiv:2411.10499 (2024)

    Pith/arXiv arXiv 2024

  29. [29]

    arXiv preprint arXiv:2303.04132 (2023)

    Josifoski, M., Sakota, M., Peyrard, M., West, R.: Exploiting asymmetry for synthetic training data generation: Synthie and the case of information extraction. arXiv preprint arXiv:2303.04132 (2023)

    Pith/arXiv arXiv 2023

  30. [30]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Kim, J., Gu, G., Park, M., Park, S., Choo, J.: Stableviton: Learning semantic correspondence with latent diffusion model for virtual try-on. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8176–8185 (2024)

    2024

  31. [31]

    ACM Transactions on Graphics (TOG)42(6), 1–13 (2023)

    Lee, D., Kang, H., Lee, I.K.: Clothcombo: Modeling inter-cloth interaction for draping multi-layered clothes. ACM Transactions on Graphics (TOG)42(6), 1–13 (2023)

    2023

  32. [32]

    In: European Conference on Computer Vision

    Lee, S., Gu, G., Park, S., Choi, S., Choo, J.: High-resolution virtual try-on with mis- alignment and occlusion-handled conditions. In: European Conference on Computer Vision. pp. 204–219. Springer (2022)

    2022

  33. [33]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Li, N., Shih, K.J., Plummer, B.A.: Enhancing virtual try-on with synthetic pairs and error-aware noise scheduling. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 21238–21247 (2025)

    2025

  34. [34]

    arXiv preprint arXiv:2502.10663 (2025)

    Li, R., Jin, X., et al.: Real: Realism evaluation of text-to-image generation models for effective data augmentation. arXiv preprint arXiv:2502.10663 (2025)

    Pith/arXiv arXiv 2025

  35. [35]

    arXiv preprint arXiv:2012.02952 (2020)

    Liu, R., Xu, G., Jia, C., Ma, W., Wang, L., Vosoughi, S.: Data boost: Text data augmentation through reinforcement learning guided conditional generation. arXiv preprint arXiv:2012.02952 (2020)

    Pith/arXiv arXiv 2012

  36. [36]

    arXiv preprint arXiv:1711.05101 (2017)

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)

    Pith/arXiv arXiv 2017

  37. [37]

    In: Proceedings of the 31st ACM international conference on multimedia

    Morelli, D., Baldrati, A., Cartella, G., Cornia, M., Bertini, M., Cucchiara, R.: Ladi- vton: Latent diffusion textual-inversion enhanced virtual try-on. In: Proceedings of the 31st ACM international conference on multimedia. pp. 8580–8589 (2023)

    2023

  38. [38]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Morelli, D., Fincato, M., Cornia, M., Landi, F., Cesari, F., Cucchiara, R.: Dress code: High-resolution multi-category virtual try-on. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2231–2235 (2022)

    2022

  39. [39]

    arXiv preprint arXiv:2009.10195 (2020)

    Ng, N., Cho, K., Ghassemi, M.: Ssmba: Self-supervised manifold based data aug- mentation for improving out-of-domain robustness. arXiv preprint arXiv:2009.10195 (2020)

    Pith/arXiv arXiv 2009

  40. [40]

    arXiv preprint arXiv:2408.00714 (2024)

    Ravi, N., Gabeur, V., Hu, Y.T., Hu, R., Ryali, C., Ma, T., Khedr, H., Rädle, R., Rolland, C., Gustafson, L., et al.: Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714 (2024)

    Pith/arXiv arXiv 2024

  41. [41]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Shivashankar, C., Miller, S.: Semantic data augmentation with generative models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 863–873 (2023)

    2023

  42. [42]

    arXiv preprint arXiv:1409.1556 (2014)

    Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)

    Pith/arXiv arXiv 2014

  43. [43]

    arXiv preprint arXiv:2407.16224 (2024)

    Sun, K., Cao, J., Wang, Q., Tian, L., Zhang, X., Zhuo, L., Zhang, B., Bo, L., Zhou, W., Zhang, W., et al.: Outfitanyone: Ultra-high quality virtual try-on for any clothing and any person. arXiv preprint arXiv:2407.16224 (2024)

    Pith/arXiv arXiv 2024

  44. [44]

    Advances in Neural Information Processing Systems36, 48382–48402 (2023)

    Tian, Y., Fan, L., Isola, P., Chang, H., Krishnan, D.: Stablerep: Synthetic images from text-to-image models make strong visual representation learners. Advances in Neural Information Processing Systems36, 48382–48402 (2023)

    2023

  45. [45]

    In: Proceedings of the IEEE/CVF winter conference on applications of computer vision

    Valvano, G., Agostino, A., De Magistris, G., Graziano, A., Veneri, G.: Controllable image synthesis of industrial data using stable diffusion. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 5354–5363 (2024)

    2024

  46. [46]

    arXiv preprint arXiv:2504.13078 (2025)

    Velioglu, R., Bevandic, P., Chan, R., Hammer, B.: Enhancing person-to-person virtual try-on with multi-garment virtual try-off. arXiv preprint arXiv:2504.13078 (2025)

    Pith/arXiv arXiv 2025

  47. [47]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Wang, S., Feng, Y., Lan, T., Yu, N., Bai, Y., Xu, R., Wang, H., Xiong, C., Savarese, S.: Text2data: Low-resource data generation with textual control. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39, pp. 21252–21260 (2025)

    2025

  48. [48]

    In: Pro- ceedings of the 61st annual meeting of the association for computational linguistics (volume 1: long papers)

    Wang, Y., Kordi, Y., Mishra, S., Liu, A., Smith, N.A., Khashabi, D., Hajishirzi, H.: Self-instruct: Aligning language models with self-generated instructions. In: Pro- ceedings of the 61st annual meeting of the association for computational linguistics (volume 1: long papers). pp. 13484–13508 (2023)

    2023

  49. [49]

    IEEE transactions on image processing 13(4), 600–612 (2004) 18 C

    Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600–612 (2004) 18 C. Feng et al

    2004

  50. [50]

    arXiv preprint arXiv:2508.02324 (2025)

    Wu, C., Li, J., Zhou, J., Lin, J., Gao, K., Yan, K., Yin, S.m., Bai, S., Xu, X., Chen, Y., et al.: Qwen-image technical report. arXiv preprint arXiv:2508.02324 (2025)

    Pith/arXiv arXiv 2025

  51. [51]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Xie, Z., Huang, Z., Dong, X., Zhao, F., Dong, H., Zhang, X., Zhu, F., Liang, X.: Gp-vton: Towards general purpose virtual try-on via collaborative local-flow global-parsing learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 23550–23559 (2023)

    2023

  52. [52]

    arXiv preprint arXiv:2601.05853 (2026)

    Xu, Y., Dingliana, J.: Layergs: Decomposition and inpainting of layered 3d human avatars via 2d gaussian splatting. arXiv preprint arXiv:2601.05853 (2026)

    arXiv 2026

  53. [53]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Xu, Y., Gu, T., Chen, W., Chen, A.: Ootdiffusion: Outfitting fusion based latent diffusion for controllable virtual try-on. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39, pp. 8996–9004 (2025)

    2025

  54. [54]

    arXiv preprint arXiv:2004.11546 (2020)

    Yang, Y., Malaviya, C., Fernandez, J., Swayamdipta, S., Bras, R.L., Wang, J.P., Bha- gavatula, C., Choi, Y., Downey, D.: Generative data augmentation for commonsense reasoning. arXiv preprint arXiv:2004.11546 (2020)

    Pith/arXiv arXiv 2004

  55. [55]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Yang, Z., Zeng, A., Yuan, C., Li, Y.: Effective whole-body pose estimation with two-stages distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4210–4220 (2023)

    2023

  56. [56]

    arXiv preprint arXiv:2304.08821 (2023)

    Yin, Y., Kaddour, J., Zhang, X., Nie, Y., Liu, Z., Kong, L., Liu, Q.: Ttida: Con- trollable generative data augmentation via text-to-text and text-to-image models. arXiv preprint arXiv:2304.08821 (2023)

    Pith/arXiv arXiv 2023

  57. [57]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586–595 (2018)

    2018

  58. [58]

    arXiv preprint arXiv:2405.00448 (2024)

    Zhang, X., Lin, E., Li, X., Luo, Y., Kampffmeyer, M., Dong, X., Liang, X.: Mmtryon: Multi-modal multi-reference control for high-quality fashion generation. arXiv preprint arXiv:2405.00448 (2024)

    Pith/arXiv arXiv 2024

  59. [59]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Zhu, L., Li, Y., Liu, N., Peng, H., Yang, D., Kemelmacher-Shlizerman, I.: M&m vto: Multi-garment virtual try-on and editing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1346–1356 (2024) A. IMPLEMENTATION DETAILS 19 This supplementary material provides details on implementation specifics and deeper methodologi...

    2024

  60. [60]

    deform-and-place

    for5 , 000steps. For the traditional benchmarks [9,38], stage 1 is also trained for20 , 000steps, while stage 2 is fine-tuned for3, 000steps. We report results using the last checkpoint for all experiments and use a single NVIDIA H200 GPU. Our implementation is built using PyTorch (version2.7.0+cu128). B Dataset Curation Details This section provides a de...