Pith. sign in

REVIEW 5 major objections 5 minor 19 references

FW-VTON: Flattening-and-Warping for Person-to-Person Virtual Try-on

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A three-stage pipeline—flatten, warp, integrate—achieves state-of-the-art person-to-person virtual try-on.

desk verdict Useful dataset and strong flattening results, but the VITON-HD SOTA claim is undercut by proxy warping supervision and uneven baselines. read the letter →

arxiv 2507.16010 v1 pith:HCSUO7XP submitted 2025-07-21 cs.CV

Zheng Wang , Xianbing Sun , Shengyi Wu , Jiahui Zhan , Jianlou Si , Chi Zhang , Liqing Zhang , Jianfu Zhang This is my paper
classification cs.CV
keywords person-to-personvirtualtry-ongarmentflatteningwarpingdiffusionmodeldualU-NetreferencenetP2P-VTONdatasettry-off
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper confronts the harder virtual try-on setting where the desired garment is only visible on another person, not as a flat product photo. It argues that the right way to handle occluded, distorted garments is to decompose the task into three learned stages: flattening the garment into a clean front view, warping it to the target person's pose, and integrating it with the target image. The paper introduces a new dataset of multi-pose images of the same person wearing the same garment to support training and evaluation. Across VITON-HD and the new P2P-VTON dataset, the proposed FW-VTON reports the best FID, KID, and user-preference scores among compared methods, and its flattening module also outperforms prior try-off models.

What carries the argument

The central machinery is the three-stage decomposition built on a shared dual U-Net backbone: a reference net initialized from Stable Diffusion's U-Net whose features are concatenated spatially with the denoising U-Net's features through self-attention, letting the denoising stream consult garment details while synthesizing. Flattening conditions the reference net on the cut garment and feeds the whole source person image to CLIP for pose-and-fit context; warping concatenates the target garment-region mask with the cut garment in the reference net's input; integration overlays the warped garment on the agnostic target image as an 8-channel latent input and drops cross-attention. Each module is trained separately for 45,000 steps and sampled with DDIM at 20 steps.

What would settle it

Train the warping module only on PFAFN-synthesized VITON-HD pairs and evaluate it on the P2P-VTON test set, where the same garment appears in multiple real poses and ground-truth warped garments can be constructed from those pairs. If the warping module's accuracy drops sharply relative to a version trained on real multi-pose pairs, the VITON-HD gains are imitation of PFAFN rather than generalizable warping.

Watch

Extended reading notes

Core claim

On the paper's own terms, FW-VTON establishes that person-to-person virtual try-on can be handled without any flat garment image by explicitly reconstructing a flat garment from the clothed source, warping it to the target body, and then fusing the warped structure with the flat garment's details. The flattening stage uses a dual U-Net with the full source image (not just the cut garment) as CLIP conditioning, which the authors show is far better than conditioning on the isolated garment. The warping stage concatenates the target garment mask with the cut garment image in the reference net, and the integration stage removes cross-attention entirely, relying on the warped garment as explicit structure guidance and the flattened garment as detail reference. The claimed result is state-of-the-art performance on both VITON-HD (FID 8.53) and the new P2P-VTON dataset (FID 18.78), plus a try-off module that outperforms TryOffDiff and TryOffAnyone on SSIM, LPIPS, FID, KID, and DISTS.

Load-bearing premise

The VITON-HD evaluation assumes that warping targets synthesized by the PFAFN warping module from flat garment images are valid supervision for person-to-person warping, so the learned warping module may inherit PFAFN's biases rather than learning genuine garment warping from clothed people.

Editorial extensions

If this is right

  • Person-to-person try-on no longer requires a flat garment image, so a user can try on a garment seen only in a photo of someone else wearing it.
  • Decomposing the task into flattening, warping, and integration outperforms single-stage and two-stage baselines; ablations show removing either the flattening or the warping module degrades FID and KID.
  • The try-off subtask, extracting a flat garment from a dressed person, can be solved effectively with a dual U-Net and full-image CLIP conditioning, giving state-of-the-art reconstruction scores.
  • The new P2P-VTON dataset, with multiple poses of the same person in the same garment, supports training and benchmarking in this setting and reveals that performance still drops for more diverse poses, pointing to headroom for future work.
  • In image-only diffusion conditioning, cross-attention for text can be omitted without hurting quality, simplifying the integration module and reducing computational cost.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The P2P-VTON dataset's multi-pose groups could enable fully supervised training of the warping module with genuine ground-truth warps, removing the reliance on PFAFN-synthesized supervision and potentially strengthening the case that the warping module generalizes.
  • The finding that full source-image CLIP conditioning beats cut-garment conditioning in flattening suggests that pose and body context carry essential reconstruction cues; the same principle may transfer to other garment or object reconstruction tasks where occlusion is heavy.
  • Because the three stages are modular, each can be upgraded or benchmarked independently; for instance, a different warping backbone or a stronger flattening prior could be dropped into the pipeline without retraining the other stages.
  • If the VITON-HD margin largely reflects imitation of PFAFN's warping patterns, the P2P-VTON results, where baselines were trained on real multi-pose data, are the more trustworthy evidence of genuine person-to-person warping ability.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper introduces FW-VTON, a three-stage pipeline for person-to-person virtual try-on: a flattening module extracts a flat garment representation from a clothed source person, a warping module aligns the garment to the target pose, and an integration module synthesizes the final try-on image. The authors also introduce a new P2P-VTON dataset of paired clothed-person images with corresponding flat garment annotations. They report state-of-the-art results on VITON-HD and on the new dataset, including a large user-study margin, and report strong try-off (garment extraction) results on VITON-HD.

Significance. If the claims are fully supported, the contribution is meaningful: the three-stage decomposition is a sensible approach to the under-explored person-to-person setting, and the new P2P-VTON dataset fills a real gap in available supervision for this task. The try-off results (Table 3) are a clear strength, as the flattening module outperforms two recent dedicated baselines by a substantial margin. However, the central SOTA claim is weakened by evaluation asymmetries and by the use of PFAFN-synthesized warps as training targets on VITON-HD. The paper does not ship code or data, and the favorable VITON-HD numbers may partly reflect imitation of the teacher warp distribution rather than a genuinely better warping model. With fairer baselines, error bars, and validation on real person-to-person pairs, the contribution would be solid.

major comments (5)
  1. [Supplement A (FW-VTON training)] The warping module on VITON-HD is trained on PFAFN-synthesized warped garment images: the supplement states that because VITON-HD has no garment images under diverse poses, the authors adopt the PFAFN warping module to synthesize warped garment images from flat garment inputs and use those synthesized pairs as training data. This makes the warping module a student of PFAFN, so the claim that the warping module eliminates dependency on flat garment images is not demonstrated on VITON-HD. The small FID margin in Table 1 (8.53 vs. 8.86 for Stable-VTON) may partly reflect successful imitation of PFAFN's warp distribution rather than superior warping ability. The authors should evaluate the warping module on the P2P-VTON dataset, which has real person-to-person pairs, and report quantitative warp-quality metrics there.
  2. [§4.1 Compared Methods, Tables 1 and 2] The baseline comparison is inconsistent. DCI-VTON and LaDI-VTON on VITON-HD use pre-trained warping modules and only the final diffusion model is trained, while Stable-VTON and IDM-VTON are fine-tuned or retrained; CatVTON is not retrained at all, using its released checkpoint. CatVTON is then omitted from the P2P-VTON comparison (Table 2), even though it is a strong recent baseline designed to handle concatenated image inputs, which is relevant to the person-to-person setting. The SOTA claim on both tables would be much more convincing if all baselines were retrained under the same protocol, including CatVTON (e.g., by feeding it cut garments from the source person), or if the authors explicitly justified the omission and reported the effect of using a released checkpoint.
  3. [§4.3 Table 4 ablation (w/o Warping)] The 'w/o Warping' ablation replaces the learned warping module with the PFAFN warping module. Because the learned module was trained on PFAFN-generated warps on VITON-HD (Supplement A), this ablation compares the student to its teacher. The observed improvement (9.57 vs. 8.53 FID) may only reflect the student's ability to fit the teacher's outputs, not intrinsic superiority of the learned warp mechanism. A cleaner ablation would train the warping module on real person-to-person pairs from P2P-VTON and perform the ablation on that dataset, or would report the distance between the learned warps and PFAFN warps on held-out VITON-HD pairs.
  4. [§4.1 Evaluation Metrics and user study] The user study protocol is under-specified. The paper reports selection proportions that sum to 100 in Tables 1 and 2, but it does not state whether judges performed a forced-choice ranking among all methods per image, how many judges rated each pair, whether the judges saw the source person image and target person image, or whether any quality guidance was provided. A single preference proportion of 66.24% can be difficult to interpret without knowing the number of alternatives and the inter-rater variability. Please provide the full protocol, including the exact question asked, the number of ratings per image, and any measures of agreement.
  5. [Tables 1 and 2, FID/KID] No variance or confidence intervals are reported for any FID or KID value. Given the small margins in Table 1 (e.g., 8.53 vs. 8.86), it is not possible to tell whether the difference is significant. FID/KID estimates depend on the number of samples and the feature-extractor preprocessing; the authors should report error bars over the test-set samples or over multiple evaluation runs, and they should state the exact sample count used for the FID/KID computation in the person-to-person setting where ground truth is not available.
minor comments (5)
  1. [Supplement A] The supplement uses the name 'FW-VITON' once (in the Stable-VITON description); this is inconsistent with the main text's 'FW-VTON'.
  2. [Supplement B, Table 7] Table 7's caption reads 'Ablation study of CLIP input in the flattening module', but the table actually shows combinations of try-off and try-on methods and reports FID/KID. The caption appears to be a copy-paste error and should be corrected to describe the actual experimental setup.
  3. [§3.2 Warping Module] The sentence 'extending its input channels from 4 to 5 through zero-parameter padding' is unclear; please specify whether the weights for the new channel are initialized to zero and how this interacts with the VAE-encoded inputs.
  4. [§3.4 P2P-VTON Dataset] The dataset description reports group sizes (2,787 groups with two images, 916 with three, etc.) but does not give the total number of images or the total number of unique persons/garments. Adding these totals would help readers understand the dataset's scale and diversity.
  5. [§4.2 Qualitative Results] The text says 'Similar artifacts are observed in IDM-VTON' but Figure 4 shows only a single qualitative example per method; please provide additional qualitative examples to support the general claim about artifacts.

Circularity Check

1 steps flagged · score 4.0 of 10

VITON-HD warping 'prediction' is trained on PFAFN-synthesized targets, so the flat-garment-independence claim is only externally grounded by the new P2P-VTON dataset; the rest of the pipeline is self-contained.

  1. fitted input called prediction [Supplementary Section A (FW-VTON paragraph) and Section 4.3 / Table 4]
    "Since the VITON-HD dataset does not provide garment images under diverse poses, we adopt the PFAFN warping module to synthesize warped garment images from flat garment inputs, conditioned on the target person’s pose. These synthesized pairs are then used as training data for our warping module."

    On VITON-HD, the warping module's supervised target is PFAFN's warped output, so its predicted G_w is, by construction, a fit to PFAFN. The paper then presents VITON-HD FID (Table 1: 8.53) and the Table 4 ablation (Ours 8.53 vs PFAFN 'w/o Warping' 9.57) as evidence for the module that 'eliminat[es] the dependency on flat garment images.' That comparison is student-vs-teacher inside a PFAFN-generated target distribution, and the 'w/o Warping' condition further feeds PFAFN warps to an integration module trained on the learned module's outputs. The VITON-HD numbers therefore do not demonstrate flat-garment-free warping; only the P2P-VTON experiments supply real multi-pose warp supervision. This is a partial, evidence-level circularity rather than an equation-level equivalence.

full rationale

The derivation chain is mostly self-contained: the flattening module is trained and evaluated against ground-truth flattened garment images (Table 3, SSIM/LPIPS/FID/DISTS), and the integration module is measured by FID/KID and a user study against real image distributions; neither stage reduces to its own output. The self-citations to DCI-VTON are baseline comparisons rather than load-bearing premises. The one significant circularity is the VITON-HD warp supervision: the warping module is trained to reproduce PFAFN's warps synthesized from flat garments, and the 'w/o Warping' ablation replaces the module with PFAFN, so the VITON-HD SOTA margin (8.53 vs 8.86) and the ablation margin (8.53 vs 9.57) are measured inside a target distribution defined by PFAFN. This does not make the whole method circular, because the P2P-VTON dataset provides genuine person-to-person pose pairs and the model wins there as well (FID 18.78 vs 19.01; user study 61.70), and flattening and integration retain independent evaluation. The paper's own supplement discloses the proxy supervision, which mitigates but does not remove the circularity of using the VITON-HD numbers as evidence for flat-garment-free warping. Overall score 4.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on hand-set training hyperparameters and on the assumption that existing parsing, pre-trained diffusion, and PFAFN warping tools are reliable enough to generate supervision. No new physical entities are introduced. The P2P-VTON dataset is a new resource, but it is not released, so it cannot currently serve as independent evidence.

free parameters (4)
  • DDIM inference steps = 20
    Used for all modules at inference; hand-chosen, no sensitivity analysis reported.
  • Training steps per module = 45,000
    Each module (warping, flattening, integration) trained for 45,000 steps; no convergence criterion or early stopping described.
  • Learning rate = 1e-5
    Adam optimizer setting; conventional, not tuned.
  • Baseline fine-tuning schedules = 40 or 130 epochs depending on method
    Unequal training budgets across baselines affect the fairness of the SOTA comparison.
assumptions (4)
  • domain assumption Pre-trained Stable Diffusion weights provide a usable generative prior for all three modules.
    All modules are initialized from the SD-Image-Variations checkpoints (Supplementary A).
  • domain assumption Human parsing maps from Li et al. 2020 and Papandreou et al. 2017 are accurate enough for garment extraction and agnostic image construction.
    Parsing maps S_p and S_g are used at every preprocessing step; errors in parsing are not analyzed.
  • ad hoc to paper PFAFN-synthesized warped garments are valid training targets for the warping module on VITON-HD.
    Supplementary A states VITON-HD lacks diverse-pose garment images, so PFAFN generates the warping ground truth; if PFAFN is biased, the learned warping inherits the bias.
  • domain assumption FID, KID, and the user study are valid proxies for try-on quality when ground truth is unavailable.
    Used for P2P-VTON test pairs; no error bars or statistical tests are reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FW-VTON: Flattening-and-Warping for Person-to-Person Virtual Try-on." pith.science (2026). https://pith.science/paper/HCSUO7XP

@misc{pith2026250716010,
  author       = {Pith},
  title        = {Pith review of: FW-VTON: Flattening-and-Warping for Person-to-Person Virtual Try-on},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HCSUO7XP}},
  note         = {Machine review of arXiv:2507.16010}
}
read the original abstract

Traditional virtual try-on methods primarily focus on the garment-to-person try-on task, which requires flat garment representations. In contrast, this paper introduces a novel approach to the person-to-person try-on task. Unlike the garment-to-person try-on task, the person-to-person task only involves two input images: one depicting the target person and the other showing the garment worn by a different individual. The goal is to generate a realistic combination of the target person with the desired garment. To this end, we propose Flattening-and-Warping Virtual Try-On (\textbf{FW-VTON}), a method that operates in three stages: (1) extracting the flattened garment image from the source image; (2) warping the garment to align with the target pose; and (3) integrating the warped garment seamlessly onto the target person. To overcome the challenges posed by the lack of high-quality datasets for this task, we introduce a new dataset specifically designed for person-to-person try-on scenarios. Experimental evaluations demonstrate that FW-VTON achieves state-of-the-art performance, with superior results in both qualitative and quantitative assessments, and also excels in garment extraction subtasks.

Figures

Figures reproduced from arXiv: 2507.16010 by the authors.

Figure 1
Figure 1. Examples of results from FW-VTON for person-to-person virtual try-on task. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed FW-VTON framework. The overall pipeline includes three stages as illustrated in (a). (b) shows the detailed architecture of [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Some examples of our P2P-VTON dataset. We concatenate this latent with Gaussian noise 𝜖 ∈ R 𝐻 ×𝑊 ×4 , result￾ing in an 8-channel input to the denoising U-Net. To accommodate this, we modify the first convolutional layer of the diffusion model by introducing zero-padded parameters. We also remove the cross-attention layers from both the reference net and the denoising U-Net. In Stable Diffusion [Rombach et al. 2022],… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Qualitative results for person-to-person VTON task on the VITON-HD dataset. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Qualitative results for the try-off subtask on the VITON-HD dataset. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Ablation examples of our two modules. Note that in the person-to [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Qualitative results for the person-to-person VTON task on our P2P-VTON dataset. Please note that the garment image is not available in our task and [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: More qualitative results for the person-to-person VTON task on the VITON-HD dataset. [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Ablation examples of G as CLIP input. from the publicly released SD-Image-Variations-Diffusers check￾point. Each module (warping, flattening, and integration) is trained separately for 45,000 steps using 8×NVIDIA A800 GPUs, with each training process taking approximate…
Figure 10
Figure 10. Figure 10: More qualitative results for the try-off subtask on the VITON-HD dataset. Due to the data augmentation applied during training, our results feature a [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

19 extracted references · 11 canonical work pages

  1. [4]

    Generative adversarial networks. Commun. ACM 63, 11 (2020), 139–144. Junhong Gou, Siyu Sun, Jianfu Zhang, Jianlou Si, Chen Qian, and Liqing Zhang

    work page 2020

  2. [5]

    Method SSIM ↑ LPIPS↓ FID↓ KID↓ DISTS↓ G𝑐 as CLIP input 0.658 0.399 19.71 6.6 0.2251 G as CLIP input 0.739 0.363 10.04 1.7 0.2016 Table

    Ablation study of concatenatingP𝑚 withG𝑐 in Warping Module.G𝑐 denotes the cut garment image from source person image,P𝑚 denotes the corresponding garment mask region in target person image. Method SSIM ↑ LPIPS↓ FID↓ KID↓ DISTS↓ G𝑐 as CLIP input 0.658 0.399 19.71 6.6 0.2251 G as CLIP input 0.739 0.363 10.04 1.7 0.2016 Table

    work page

  3. [6]

    IEEE Transactions on Pattern Analysis and Machine Intelligence 44, 6 (2020), 3260–3271

    Self-correction for human parsing. IEEE Transactions on Pattern Analysis and Machine Intelligence 44, 6 (2020), 3260–3271. Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaoou Tang

    work page 2020

  4. [8]

    arXiv preprint arXiv:2010.02502 (2020)

    Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020). Riza Velioglu, Petra Bevandic, Robin Chan, and Barbara Hammer

    arXiv 2020

  5. [9]

    arXiv preprint arXiv:2411.18350 (2024)

    TryOffDiff: Virtual-Try-Off via High-Fidelity Garment Reconstruction using Diffusion Models. arXiv preprint arXiv:2411.18350 (2024). Chenhui Wang, Tao Chen, Zhihao Chen, Zhizhong Huang, Taoran Jiang, Qi Wang, and Hongming Shan

    arXiv 2024

  6. [10]

    In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24, Kate Larson (Ed.)

    FLDM-VTON: Faithful Latent Diffusion Model for Virtual Try-on. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24, Kate Larson (Ed.). International Joint Conferences on Artificial Intelligence Organization, 1362–1370. https://doi.org/10.24963/ijcai.2024/151 Main Track. Zhou Wang, Alan C Bovik, Hamid R Sh...

    work page doi:10.24963/ijcai.2024/151

  7. [12]

    arXiv preprint arXiv:2412.08573 (2024)

    TryOffAnyone: Tiled Cloth Generation from a Dressed Person. arXiv preprint arXiv:2412.08573 (2024). Zhenyu Xie, Zaiyu Huang, Xin Dong, Fuwei Zhao, Haoye Dong, Xijin Zhang, Feida Zhu, and Xiaodan Liang

    arXiv 2024

  8. [13]

    arXiv preprint arXiv:2403.01779 (2024)

    Ootdiffusion: Outfit- ting fusion based latent diffusion for controllable virtual try-on. arXiv preprint arXiv:2403.01779 (2024). Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang

    arXiv 2024

Show all 19 references
  1. [14]

    Upper Clothes

    More qualitative results for the person-to-person VTON task on the VITON-HD dataset. FW-VTON: Flattening-and-Warping for Person-to-Person Virtual Try-on • 11 A SUPPLEMENTARY MATERIALS OF FW-VTON: FLATTENING-AND-WARPING FOR PERSON-TO-PERSON VIRTUAL TRY-ON This section provides ...

    2023

  2. [15]

    from the publicly released SD-Image-Variations-Diffusers check- point

    Ablation examples of G as CLIP input. from the publicly released SD-Image-Variations-Diffusers check- point. Each module (warping, flattening, and integration) is trained separately for 45,000 steps using 8×NVIDIA A800 GPUs, with each training process taking approximately 6 ho...

    2020

  3. [17]

    G denotes the original garment image, andG𝑐 represents the cut garment region from G

    Ablation study of CLIP input in the flattening module. G denotes the original garment image, andG𝑐 represents the cut garment region from G. integrating the cut garment image,G𝑐, with the corresponding gar- ment mask region from the target person image,P𝑚. We explored three di...

    2013

  4. [18]

    This method effectively incorporates garment mask information to enhance the warping process, outperforming the other two settings

    As shown, the third setting—concatenatingP𝑚 withG𝑐 in the reference net’s input channel—achieves the best performance. This method effectively incorporates garment mask information to enhance the warping process, outperforming the other two settings. B.3 Effect of Skipping the...

    2022

  5. [19]

    Ablation study of whether skip the cross-attention layer in dual U-Net. • w/o Cross-Attention: Our default setting, which removes all cross-attention layers in both the reference and denoising U- Nets (as described in the Method section of the main paper). • w/ Cross-Attention...

    2021

  6. [2004]

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

    Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13, 4 (2004), 600–612. Ioannis Xarchakos and Theodoros Koukopoulos

    2004

  7. [2013]

    arXiv preprint arXiv:1312.6114 (2013)

    Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013). Sangyun Lee, Gyojung Gu, Sunghyun Park, Seunghwan Choi, and Jaegul Choo

    2013 arXiv

  8. [2017]

    arXiv preprint arXiv:1711.05101 5 (2017)

    Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101 5 (2017). Davide Morelli, Alberto Baldrati, Giuseppe Cartella, Marcella Cornia, Marco Bertini, and Rita Cucchiara

    2017 arXiv

  9. [2018]

    arXiv preprint arXiv:1801.01401 (2018)

    Demystifying mmd gans. arXiv preprint arXiv:1801.01401 (2018). Seunghwan Choi, Sunghyun Park, Minsoo Lee, and Jaegul Choo

    2018 arXiv

  10. [2020]

    Chongjian Ge, Yibing Song, Yuying Ge, Han Yang, Wei Liu, and Ping Luo

    Image quality assess- ment: Unifying structure and texture similarity.IEEE transactions on pattern analysis and machine intelligence 44, 5 (2020), 2567–2581. Chongjian Ge, Yibing Song, Yuying Ge, Han Yang, Wei Liu, and Ping Luo. 2021a. Disentangled cycle consistency for highly...

    2020

  11. [2024]

    arXiv preprint arXiv:2407.15886 (2024)

    Catvton: Concatenation is all you need for virtual try-on with diffusion models. arXiv preprint arXiv:2407.15886 (2024). Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli

    2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.