Pith. sign in

REVIEW 6 major objections 4 minor 33 references

DiffFit: Disentangled Garment Warping and Texture Refinement for Virtual Try-On

T0 review · 6 major / 4 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 DiffFit claims that decoupling geometric garment warping from texture refinement in two latent diffusion stages produces higher-fidelity virtual try-on results than single-stage or GAN-based alternatives.

desk verdict DiffFit's two-stage design is sensible and clearly described, but the central SOTA claim is undercut by a KIDp inconsistency between tables and a selective baseline set. read the letter →

arxiv 2506.23295 v1 pith:6QVDTURD submitted 2025-06-29 cs.CV

Xiang Xu This is my paper
classification cs.CV
keywords virtualtry-onlatentdiffusiongarmentwarpingcross-modalfusiontwo-stagegenerationtexturerefinementhigh-fidelitysynthesis
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

DiffFit is a proposal for how to organize a diffusion-based virtual try-on system: instead of asking one generative model to simultaneously deform a garment and paint its fabric, the paper splits the job into a first stage that warps the garment to the target pose and a second stage that renders the final image from the warped garment, the original garment, and the person. The paper reports that this split improves measured fidelity on the DressCode and VITON-HD benchmarks, including the best SSIM (0.986 and 0.916) and paired FID (4.06 and 6.62) among the compared methods. A sympathetic reader would care because the result suggests the failure modes of virtual try-on—blurry textures, misaligned seams, lost logos—are separable and can be attacked one at a time rather than jointly.

What carries the argument

The load-bearing machinery is the two-stage progressive pipeline built on latent diffusion. Stage I's cross-attention warping U-Net takes image encodings of person, garment, and mask, projects them through a trainable Q-Former into token sequences, and injects them into the denoising U-Net's cross-attention layers, so the garment is deformed to the body without explicit keypoint annotations. Stage II's cross-modal fusion U-Net concatenates the VAE latents of the person, the reference garment, and the warped garment and also cross-attends to the warped garment's embedding, which preserves texture while the person image maintains identity and pose. The Q-Former projection and the retained original-garment condition are the two components the ablations single out as essential.

What would settle it

Train a single-stage latent diffusion model with the same person, mask, and garment inputs and no separate warping stage, under identical data and sampling settings, and compare paired FID and SSIM on VITON-HD and DressCode; if it matches or beats DiffFit, the two-stage disentanglement claim is refuted.

Watch

Extended reading notes

Core claim

The paper's central claim is that virtual try-on is best treated as two sequential diffusion problems rather than one. Stage I (CA W-UNet) uses cross-attention over Q-Former tokens from the person, the clothing, and a mask to produce a warped garment image aligned to the body. Stage II concatenates the VAE latents of the person, the original garment, and the warped garment along the channel dimension and injects the warped garment's embedding through cross-attention into a denoising U-Net, generating the final try-on image. On DressCode the method reports FIDp 4.06, KIDp 1.51, FIDu 6.28, SSIM 0.986, and on VITON-HD LPIPS 0.041, SSIM 0.916, FIDp 6.62, KIDp 1.08, FIDu 8.21. Ablations show the projection layer and the original-garment condition each contribute substantially to these scores.

Load-bearing premise

The load-bearing premise is that separating garment warping from texture rendering is worth the extra stage; the paper never tests a single-stage model with the same conditioning, so if such a model matches the results, the two-stage design is not the source of the gains.

Editorial extensions

If this is right

  • If DiffFit is right, the warped garment can be inspected and corrected before final rendering, giving practitioners an explicit control point for fit errors.
  • The reported gains indicate that direct image conditioning, rather than text descriptions, carries the fine-grained details such as logos, wrinkles, and lace that previous diffusion try-on models lose.
  • On the two public benchmarks, the method sets new reference numbers for paired fidelity, so future virtual try-on papers will be compared against a warp-then-render diffusion baseline.
  • The design suggests the same two-stage split can be transferred to related garment-conditioned generation tasks, such as outfit transfer or virtual dressing at scale.

Reading between the lines

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

  • Editorial inference: the most informative experiment the paper does not run is removing or fusing the two stages; a single-stage model with identical conditioning would isolate whether the improvement comes from the warping stage or simply from the extra capacity and conditioning.
  • Editorial inference: because Stage I emits an explicit warped garment image, the framework could be lifted to video try-on by tracking the warp across frames and sharing Stage I results, though the paper itself does not demonstrate this.
  • Editorial inference: the unpaired fidelity scores are less uniformly favorable (KIDu 1.71 versus 1.70 on DressCode and 1.36 versus 0.49 on VITON-HD), suggesting the headline advantage is strongest for paired structural fidelity rather than for the overall distribution of generated images.
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

6 major / 4 minor

Summary. The manuscript proposes DiffFit, a two-stage latent diffusion pipeline for image-based virtual try-on. Stage I uses a cross-attention warping UNet to deform the garment to the target pose, and Stage II uses a conditional latent diffusion UNet that concatenates person, original garment, and warped garment latents and injects warped-garment embeddings via cross-attention. Experiments on DressCode and VITON-HD report SSIM, LPIPS, FIDp/KIDp, and FIDu/KIDu, alongside a user study and ablations of the projection layer and garment conditioning. The abstract claims superior performance over existing state-of-the-art methods in both quantitative metrics and perceptual evaluations.

Significance. Virtual try-on is a practically important and active research area, and the idea of explicitly decoupling geometric warping from texture synthesis within a diffusion framework is plausible and potentially useful. The paper evaluates on two standard benchmarks and includes ablations of two internal design choices. However, the central state-of-the-art claim is not supported by the evidence as presented: the reported numbers are internally inconsistent, the comparison set omits key recent diffusion-based baselines, and the two-stage architecture itself is never ablated. The paper does not provide code or checkpoints, so the empirical claims are the only evidence for its contribution. If the numerical inconsistencies were resolved and the missing comparisons supplied, the core idea might merit further attention, but in its current form the main claim is not established.

major comments (6)
  1. [Section 4.4, Table 2 vs Section 4.6, Table 4] The abstract and Table 2 report DiffFit's KIDp on VITON-HD as 1.08, while Table 4, under the same configuration labeled 'Ours', reports KIDp as 1.56, and the surrounding text explicitly states 'KIDp scores (6.62 and 1.56, respectively)'. Both entries describe the same method, dataset, and evaluation protocol, so at least one number is incorrect. If 1.56 is the correct value, DiffFit no longer ties LaDI-VTON on KIDp and the claim of the best KIDp is false. Because this value is cited in the abstract as a headline result, the inconsistency is load-bearing and must be resolved by rerunning or correcting the evaluation before any SOTA claim can be considered.
  2. [Section 4.4, Tables 1 and 2; Abstract] The results as reported do not support the abstract's claim of 'superior performance over existing state-of-the-art methods'. On DressCode, DCI-VTON achieves a lower LPIPS (0.0301 vs 0.040), and on VITON-HD, DCI-VTON achieves a substantially lower KIDu (0.49 vs 1.36). The text acknowledges these cases but describes them as 'competitive' or 'closely comparable'; this is not sufficient for a global superiority claim when the method loses on two of the six reported metrics on each dataset. The comparison also reports no error bars, standard deviations, or multiple-seed runs, so small differences (e.g., FIDp 6.62 vs 6.66 on VITON-HD) cannot be distinguished from noise.
  3. [Section 2.1 and Section 4.4, Tables 1-2] StableVITON [20] and OOTDiffusion [22] are discussed in the related work as relevant diffusion-based virtual try-on methods, but neither appears in the quantitative comparison tables. Since the paper's central claim is that DiffFit achieves state-of-the-art performance, omitting these recent, widely cited baselines makes the comparison set incomplete. The tables need to be expanded to include them, or the claim of state-of-the-art performance must be restricted to the methods actually evaluated.
  4. [Section 3.1 and Section 4.6] The core architectural premise of the paper is that a two-stage design that separates geometry-aware warping from texture refinement is beneficial, but no ablation tests this premise. The ablations in Section 4.6 vary only the projection layer and the presence of the original-garment condition; they do not remove or merge the two stages. A single-stage variant that integrates the same conditioning without an explicit warping stage is needed to justify the claim that decoupling reduces task complexity and improves stability, as stated in the introduction and Section 3.1.
  5. [Section 3.2 and Eq. (3)] The training target of Stage I is never specified. The loss in Eq. (3) is the standard denoising objective applied to some latent variable z_t, but the text says the output of Stage I is a warped garment image \tilde{x}_c. The paper does not state which image is used to obtain the ground-truth warped garment, how the latent z_t is related to that image, or how the denoised latent is decoded into \tilde{x}_c. Without these details, the warping stage cannot be reproduced or meaningfully evaluated, and the claim that Stage I performs 'fine-grained deformation and pose adaptation' is not verifiable.
  6. [Section 4.5, Table 3] The user study reports preference percentages above 60% for DiffFit across all criteria and datasets, but the paper provides no statistical testing, confidence intervals, or inter-rater agreement measures. The text calls these results 'significant', yet no significance test is described. With 1,000 responses across multiple criteria and baselines, a binomial or chi-square test would be straightforward to add; until then, the perceptual-superiority claim is not quantitatively supported.
minor comments (4)
  1. [Abstract and Section 4.1] The abstract and Section 4.1 state that DiffFit is evaluated on three benchmarks including IGPair, but Tables 1-4 report results only for DressCode and VITON-HD, and no IGPair results appear anywhere in the experimental section.
  2. [Throughout] The method name is spelled inconsistently as 'IMAGEDressing' and 'IMAGDressing' (e.g., Section 2.1 versus Figure 4); the authors should standardize the spelling to match the cited paper.
  3. [Section 3.1] The binary mask x_m is described as 'indicating the body region to be replaced', which is ambiguous: in standard virtual try-on, the mask usually denotes the region to be preserved or the region to be inpainted. The intended semantics should be stated precisely.
  4. [Section 3.2] The text says the token sequence [t_p; t_c; t_m] is injected into the denoising UNet via cross-attention, but Eq. (3) writes the loss as conditioning on t_p, t_c, t_m without specifying how the concatenated token sequence is passed to each cross-attention layer; a precise notation for the cross-attention injection would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DiffFit is an empirical system paper whose claims are evaluated on external benchmarks, with no load-bearing reduction of predictions to fitted inputs or self-citation.

full rationale

DiffFit is a two-stage latent diffusion framework for virtual try-on. The paper's central claims are quantitative comparisons on standard benchmarks (VITON-HD, DressCode, IGPair) using external metrics such as FID, KID, SSIM, and LPIPS. These are measured on held-out test splits, not derived from fitted parameters or from the paper's own definitions. The two-stage architecture is motivated by prior work and by the authors' stated design goals, but the paper does not claim to derive the architecture from first principles; it is an empirical proposal. I checked for the specific circularity patterns: there is no self-definitional construction (no quantity is defined in terms of the quantity it is said to predict); no fitted input is renamed as a prediction (all reported numbers are evaluation metrics); no load-bearing self-citation appears (the author, Xiang Xu, does not overlap with the cited Shen et al. works, so references to IMAGDressing, IMAGGarment, etc., are external citations, not self-citations); no uniqueness theorem is imported to force a choice; no ansatz is smuggled in via citation as if externally justified; and no known empirical pattern is merely renamed as a new contribution. The main weaknesses are empirical: the two-stage design is not ablated against a single-stage variant, and the VITON-HD KIDp value is inconsistent between Table 2 (1.08) and Table 4 (1.56), which undermines the reliability of the SOTA claim. However, an internal numerical inconsistency or an untested architectural premise is a correctness/evidence concern, not circularity. The derivation chain does not reduce to its inputs, so the circularity score is 0.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim is empirical and depends on standard diffusion assumptions and the representativeness of the benchmarks; no novel theoretical entities or fitted constants are introduced.

free parameters (4)
  • guidance_scale = 7.5
    Inference-time guidance weight selected by hand; not swept or justified in the paper.
  • num_sampling_steps = 50
    Number of UniPC sampling steps at inference, chosen without ablation.
  • learning_rate = 5e-05
    AdamW learning rate for both stages, stated without a sweep.
  • batch_size = 6
    Training batch size on a single RTX 4090, stated without justification.
assumptions (3)
  • standard math Latent diffusion forward noising and reverse denoising equations (Eq. 1-2) are valid and the pre-trained autoencoder represents images faithfully.
    Invoked in Section 2.2 and used as the training objective in Eqs. 3-4.
  • domain assumption The paired datasets (VITON-HD, DressCode, IGPair) correctly represent the virtual try-on task and are free of labeling errors.
    Used in Section 4.1 as evaluation and training data; no data cleaning or quality check is reported.
  • domain assumption SSIM, LPIPS, FID, and KID are valid proxies for visual quality and garment alignment.
    Used in Section 4.2 as the sole quantitative evidence for superiority.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DiffFit: Disentangled Garment Warping and Texture Refinement for Virtual Try-On." pith.science (2026). https://pith.science/paper/6QVDTURD

@misc{pith2026250623295,
  author       = {Pith},
  title        = {Pith review of: DiffFit: Disentangled Garment Warping and Texture Refinement for Virtual Try-On},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6QVDTURD}},
  note         = {Machine review of arXiv:2506.23295}
}
read the original abstract

Virtual try-on (VTON) aims to synthesize realistic images of a person wearing a target garment, with broad applications in e-commerce and digital fashion. While recent advances in latent diffusion models have substantially improved visual quality, existing approaches still struggle with preserving fine-grained garment details, achieving precise garment-body alignment, maintaining inference efficiency, and generalizing to diverse poses and clothing styles. To address these challenges, we propose DiffFit, a novel two-stage latent diffusion framework for high-fidelity virtual try-on. DiffFit adopts a progressive generation strategy: the first stage performs geometry-aware garment warping, aligning the garment with the target body through fine-grained deformation and pose adaptation. The second stage refines texture fidelity via a cross-modal conditional diffusion model that integrates the warped garment, the original garment appearance, and the target person image for high-quality rendering. By decoupling geometric alignment and appearance refinement, DiffFit effectively reduces task complexity and enhances both generation stability and visual realism. It excels in preserving garment-specific attributes such as textures, wrinkles, and lighting, while ensuring accurate alignment with the human body. Extensive experiments on large-scale VTON benchmarks demonstrate that DiffFit achieves superior performance over existing state-of-the-art methods in both quantitative metrics and perceptual evaluations.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 23 canonical work pages

  1. [20]

    In: Proc

    Kim, J., Gu, G., Park, M., Park, S., Choo, J.: Stableviton: Learning semantic cor- respondence with latent diffusion model for virtual try-on. In: Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., pp. 8176–8185 (2024)

    work page 2024

  2. [22]

    Preprint at urlhttps://arxiv.org/abs/2403.01779 (2024)

    Xu, Y., Gu, T., Chen, W., Chen, C.: OOTDiffusion: Outfitting Fusion based Latent Diffusion for Controllable Virtual Try-On. Preprint at urlhttps://arxiv.org/abs/2403.01779 (2024)

    arXiv 2024

  3. [1]

    In: Proc

    Wang, C., Zhang, L., Lu, H., Yang, J.: Toward characteristic-preserving image- based virtual try-on network. In: Proc. Eur. Conf. Comput. Vis. (2018)

    work page 2018

  4. [2]

    In: Proc

    Han, X., Wu, Z., Wu, Z., Yu, R., Davis, L.S.: Viton: An image-based virtual try-on network. In: Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (2018)

    work page 2018

  5. [3]

    In: Proc

    Shen, F., Jiang, X., He, X., Ye, H., Wang, C., Du, X., Li, Z., Tang, J.: Imagdressing-v1: Customizable virtual dressing. In: Proc. AAAI Conf. Artif. Intell. (2025)

    work page 2025

  6. [4]

    Preprint at urlhttps://arxiv.org/abs/2404.09512 (2024)

    Chen, W., Gu, T., Xu, Y., Chen, C.: Magic Clothing: Controllable Garment- Driven Image Synthesis. Preprint at urlhttps://arxiv.org/abs/2404.09512 (2024)

    arXiv 2024

  7. [5]

    Preprint at urlhttps://arxiv.org/abs/2506.01949 (2025)

    Shen, F., Du, X., Gao, Y., Yu, J., Cao, Y., Lei, X., Tang, J.: IMAGHarmony: Con- trollable Image Editing with Consistent Object Quantity and Layout. Preprint at urlhttps://arxiv.org/abs/2506.01949 (2025)

    arXiv 2025

  8. [6]

    In: Proc

    Bai, S., Zhou, H., Li, Z., Zhou, C., Yang, H.: Single stage virtual try-on via deformable attention flows. In: Proc. Eur. Conf. Comput. Vis. (2022)

    work page 2022

Show all 33 references
  1. [7]

    In: Proc

    Ge, Y., Song, Y., Zhang, R., Ge, C., Liu, W., Luo, P.: Parser-free virtual try-on via distilling appearance flows. In: Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (2021)

    2021

  2. [8]

    In: Proc

    Lee, S., Gu, G., Park, S., Choi, S., Choo, J.: High-resolution virtual try-on with misalignment and occlusion-handled conditions. In: Proc. Eur. Conf. Comput. Vis. (2022)

    2022

  3. [9]

    In: Proc

    Morelli, D., Cornia, M., Cucchiara, R.: Semantic-aware virtual try-on network. In: Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Workshops (2022)

    2022

  4. [10]

    In: Proc

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (2022)

    2022

  5. [11]

    Shen, F., Tang, J.: Imagpose: A unified conditional framework for pose-guided person generation. Adv. Neural Inf. Process. Syst. 37, 6246–6266 (2024)

    2024

  6. [12]

    Preprint at urlhttps://arxiv.org/abs/2502.09533 (2025)

    Shen, F., Wang, C., Gao, J., Guo, Q., Dang, J., Tang, J., Chua, T.-S.: Long-Term TalkingFace Generation via Motion-Prior Conditional Diffusion Model. Preprint at urlhttps://arxiv.org/abs/2502.09533 (2025)

    2025 arXiv

  7. [13]

    Preprint at 14 urlhttps://arxiv.org/abs/2504.13176 (2025)

    Shen, F., Yu, J., Wang, C., Jiang, X., Du, X., Tang, J.: IMAGGarment-1: Fine- Grained Garment Generation for Controllable Fashion Design. Preprint at 14 urlhttps://arxiv.org/abs/2504.13176 (2025)

    2025 arXiv

  8. [14]

    In: Proc

    Choi, S., Park, S., Lee, M., Choo, J.: Viton-hd: High-resolution virtual try-on via misalignment-aware normalization. In: Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (2021)

    2021

  9. [15]

    In: IEEE Signal Process

    Creswell, A., White, T., Dumoulin, V., Arulkumaran, K., Sengupta, B., Bharath, A.A.: Generative adversarial networks: An overview. In: IEEE Signal Process. Mag. (2018)

    2018

  10. [16]

    In: Proc

    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: Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (2023)

    2023

  11. [17]

    In: Proc

    Yang, H., Zhang, R., Guo, X., Liu, W., Zuo, W., Luo, P.: Towards photo- realistic virtual try-on by adaptively generating-preserving image content. In: Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., pp. 7850–7859 (2020)

    2020

  12. [18]

    In: Proc

    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: Proc. 31st ACM Int. Conf. Multimedia (2023)

    2023

  13. [19]

    In: Proc

    Morelli, D., Baldrati, A., Cartella, G., Cornia, M., Bertini, M., Cucchiara, R.: Ladi-vton: Latent diffusion textual-inversion enhanced virtual try-on. In: Proc. 31st ACM Int. Conf. Multimedia (2023)

    2023

  14. [21]

    In: Proc

    Zhu, L., Yang, D., Zhu, T., Reda, F., Chan, W., Saharia, C., Norouzi, M., Kemelmacher-Shlizerman, I.: Tryondiffusion: A tale of two unets. In: Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (2023)

    2023

  15. [23]

    Preprint at urlhttps://arxiv.org/abs/2310.06313 (2023)

    Shen, F., Ye, H., Zhang, J., Wang, C., Han, X., Yang, W.: Advancing pose-guided image synthesis with progressive conditional diffusion models. Preprint at urlhttps://arxiv.org/abs/2310.06313 (2023)

    2023 arXiv

  16. [24]

    In: Proc

    Shen, F., Ye, H., Liu, S., Zhang, J., Wang, C., Han, X., Wei, Y.: Boosting con- sistency in story visualization with rich-contextual conditional diffusion models. In: Proc. AAAI Conf. Artif. Intell., vol. 39, pp. 6785–6794 (2025) 15

    2025

  17. [25]

    Preprint at urlhttps://arxiv.org/abs/2503.00740 (2025)

    Gao, J., Sun, Y., Shen, F., Jiang, X., Xing, Z., Chen, K., Zhao, C.: Faceshot: Bring any character into life. Preprint at urlhttps://arxiv.org/abs/2503.00740 (2025)

    2025 arXiv

  18. [26]

    In: 2024 IEEE Int

    Gao, B., Ren, J., Shen, F., Wei, M., Huang, Z.: Exploring warping-guided features via adaptive latent diffusion model for virtual try-on. In: 2024 IEEE Int. Conf. Multimedia Expo (ICME), pp. 1–6 (2024)

    2024

  19. [27]

    Preprint at urlhttps://arxiv.org/abs/2405.17082 (2024)

    Wang, C., Tian, K., Guan, Y., Zhang, J., Jiang, Z., Shen, F., Han, X., Gu, Q., Yang, W.: Ensembling Diffusion Models via Adaptive Feature Aggregation. Preprint at urlhttps://arxiv.org/abs/2405.17082 (2024)

    2024 arXiv

  20. [28]

    Li, D., Li, J., Hoi, S.C.H.: Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. In: Adv. Neural Inf. Process. Syst. (2023)

    2023

  21. [29]

    In: Proc

    Morelli, D., Fincato, M., Cornia, M., Landi, F., Cesari, F., Cucchiara, R.: Dress code: High-resolution multi-category virtual try-on. In: Proc. Eur. Conf. Comput. Vis. (ECCV) (2022)

    2022

  22. [30]

    IEEE Transactions on Image Processing 13(4), 600–612 (2004) https://doi.org/10.1109/TIP.2003.819861

    Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assess- ment: From error visibility to structural similarity. IEEE Transactions on Image Processing 13(4), 600–612 (2004) https://doi.org/10.1109/TIP.2003.819861

    2004

  23. [31]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp

    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 (CVPR), pp. 586–595 (2018). https://doi.org/10.1109/CVPR.2018.00068

    2018

  24. [32]

    In: Advances in Neural Information Processing Systems (NeurIPS), vol

    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. In: Advances in Neural Information Processing Systems (NeurIPS), vol. 30, pp. 6626–6637 (2017)

    2017

  25. [33]

    In: International Conference on Learning Representations (ICLR) (2018) 16

    Bi´ nkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying mmd Gans. In: International Conference on Learning Representations (ICLR) (2018) 16

    2018

Pith tools

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