Pith. sign in

REVIEW 3 major objections 4 minor 31 references

Optimization-Free Universal Watermark Forgery with Regenerative Diffusion Models

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a single regeneration pass can transplant a diffusion watermark from one image onto any cover image, with no optimization or model access.

desk verdict Optimization-free forgery works well for UNet-based targets and Gaussian Shading, but the paper's own DiT results refute its 'universal' claim. read the letter →

arxiv 2506.06018 v1 pith:SMCY4J74 submitted 2025-06-06 cs.MM cs.AIcs.CR

Chaoyi Zhu , Zaitang Li , Renyi Yang , Robert Birke , Pin-Yu Chen , Tsung-Yi Ho , Lydia Y. Chen This is my paper
classification cs.MMcs.AIcs.CR
keywords watermarkforgerydiffusionmodelssemanticwatermarkingregenerativeblack-boxattackTree-RingGaussianShadingplug-and-play
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

This paper claims that an attacker who possesses a single watermarked image can plant that watermark onto any cover image using only publicly available generative models, with no optimization, fine-tuning, or access to the target model's parameters or watermarking scheme. The attack, called PnP (Plug-and-Plant), recovers the watermark-bearing latent noise from the target image through a public proxy diffusion model, then uses that latent as the starting noise for a regenerative diffusion model that redraws the cover image under visual and textual guidance. If the claim holds, current semantic watermarks cannot be trusted as proof of provenance: anyone could take one legitimate image from a creator and attribute arbitrary images to that creator within seconds. The paper reports up to 100% detectability and user attribution for Gaussian Shading on several target models, and its appendix shows the transfer degrades sharply for models that are architecturally distant from the proxy.

What carries the argument

The load-bearing object is the estimated watermark latent $\hat{z}^{(w)}_T$, obtained by mapping a target watermarked image into the proxy model's latent space with an encoder $E^P$ and then running inverse DDIM sampling $I_{0\to T}$ with the proxy's noise predictor $U^P$. This latent is then used as the initial noise for the regenerative model's DDIM sampler $G_{T\to 0}$, conditioned on the cover image's visual embeddings $E^R(x_c)$ and, where available, textual prompts $T^R(x_c)$. The argument depends on a shared latent-space structure between the proxy and target diffusion models: if the estimated latent retains the watermark's semantic information after inversion and regeneration, the forged image is detectable and attributable.

What would settle it

Take a TreeRing-watermarked image from a DiT-based target such as PixArt-Σ or FLUX.1, run PnP with the SD2.1 proxy, and measure detection at the paper's 1e-2 false-positive threshold; the appendix reports rates of 0.04 to 0.13, essentially chance, which would falsify the abstract's universality claim. A decisive test would repeat the 24-scenario evaluation while varying only the proxy architecture and latent channels; if detectability tracks proxy-target similarity rather than the watermark scheme, the universal claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that a semantic watermark survives a cross-model transplant: the watermark is not tied to the particular diffusion model that generated the watermarked image, but lives in a latent pattern that can be recovered and re-injected. Stage one estimates the watermark latent $\hat{z}^{(w)}_T$ by encoding the target watermarked image with the proxy encoder and running inverse DDIM sampling; stage two plugs that estimated latent into a regenerative diffusion model's DDIM sampler, conditioned on the cover image's visual and textual features, so the regenerated cover image carries the target watermark and is attributed to the target user. The method is therefore optimization-free, requires only one latent-estimation pass per watermark, and applies to arbitrary cover images.

Load-bearing premise

The attack assumes that the watermark-bearing internal noise pattern recovered by inverting the target image with a proxy diffusion model keeps its information when plugged into a different regenerative diffusion model; if the two models do not share a compatible latent space, the regeneration step erases the watermark.

Editorial extensions

If this is right

  • Any user with one target watermarked image and access to public regenerative models can produce images that detectors attribute to the target user, without knowing the target generator or watermark scheme.
  • Per-image forgery cost drops from roughly 1,190 seconds for the optimization-based Imprint baseline to 3.7 to 29.3 seconds for the PnP backbones.
  • For Gaussian Shading on SDXL and Animagine XL, PnP matches or approaches 100% detection and user attribution across CtrlRegen, StableSR, DiffBIR, SeeSR, CCSR, and HoliSDiP.
  • Super-resolution-based regenerative backbones trade some watermark detectability for noticeably better perceptual quality, producing forged images that look better than the adversarial-perturbation baseline.
  • Because the attack requires no per-image optimization, it is cheap enough to run at scale, strengthening the paper's conclusion that latent-based semantic watermarking is an unreliable governance mechanism.

Reading between the lines

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

  • Beyond the paper's stated results, this implies that provenance systems cannot rely on latent-based semantic watermarks alone when a compatible public proxy model exists; a single leaked or stolen watermarked image becomes a reusable attribution stamp.
  • The paper does not test a DiT-based proxy; replacing SD2.1 with a proxy whose autoencoder and latent dimensionality match the target would be the natural next experiment and would clarify whether the universality claim is architectural or genuinely general.
  • A practical extension would be to use PnP as a watermark-removal audit: forge a watermark onto a known clean image, then check whether the original clean latent is recoverable, which would quantify how much forensic information survives regeneration.
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

3 major / 4 minor

Summary. The paper proposes PnP (Plug-and-Plant), a two-stage, optimization-free watermark forgery attack. Stage 1 estimates the watermark latent by DDIM-inverting a target watermarked image through a public proxy model (SD 2.1). Stage 2 injects this latent into a regenerative diffusion model (super-resolution, face restoration, or watermark-removal backbone) conditioned on the cover image's visual and textual features, producing a forged image said to carry the target watermark. The authors evaluate against Gaussian Shading and TreeRing across four target models (SDXL, PixArt-Σ, FLUX.1, Animagine XL) and three datasets (RealSR, DRealSR, CelebA), comparing with the optimization-based Imprint baseline. They report near-perfect detection and user attribution in the best cases, improved runtime, and a code release.

Significance. If the strong claims held, PnP would be a significant result: it would show that semantic watermarks can be forged by off-the-shelf regeneration without per-image optimization, at practical runtime and with better perceptual quality than adversarial optimization. The paper's strengths include a broad 24-scenario evaluation, direct comparison with the optimization-based Imprint baseline, and a public code release. The Gaussian Shading results on SDXL and Animagine XL are consistent and credible. However, the central 'universal' claim is not supported by the paper's own TreeRing results on DiT-based targets, and the detectability-versus-quality trade-off is acknowledged but not reflected in the abstract. The work is relevant and publishable as a conditional attack, but the claims need substantial revision.

major comments (3)
  1. [Appendix C.1 / Table 4] The 'universal' claim in the Abstract and Section 3.2.3 is contradicted by the paper's own Table 4. For TreeRing, PnP with CtrlRegen reaches detection success rates of only 0.53/0.72/0.52 on PixArt-Σ and 0.13/0.09/0.06 on FLUX.1 across RealSR/DRealSR/CelebA at a 1e-2 false positive rate, with FLUX.1 p-values around 0.21–0.22; these are statistically indistinguishable from clean images under the stated threshold. The assumption in Section 3.2.3 of a 'shared latent-space structure' across diffusion models is therefore empirically violated for DiT-based targets and for FLUX's distinct autoencoder. Even though the Imprint baseline also fails on FLUX.1 TreeRing, the paper's own universality claim is independent of the baseline's behavior. Please restrict the claim to UNet-family targets whose latent spaces are compatible with the proxy, report worst-case numbers as prominently as best-case, and either remove the universality wording or add a formal transferability condition.
  2. [Table 1 / Section 4.2] Even for Gaussian Shading, the attack's success degrades substantially on PixArt-Σ and FLUX.1. For example, CtrlRegen's user attribution rate drops to 0.88–0.91 on PixArt-Σ and 0.95–0.98 on FLUX.1, versus 1.00 on SDXL and Animagine XL. The abstract's 'up to 100% detectability and user attribution' is thus a best-case figure. The paper should report the full range of results and explicitly characterize when attribution accuracy is below the level needed to reliably assign the forged image to the target user. Without such qualification, the statements 'successfully forge the watermark' and 'universal' are misleading.
  3. [Section 4.3 / Fig. 4] The joint claim of 'maintain the best visual perception' is not supported for the method that achieves the highest detectability. CtrlRegen has near-1.00 detection on SDXL/Animagine but is grouped with Imprint as having the worst perceptual quality in Fig. 4 (smallest covered area, low CLIPIQA/MUSIQ/NIQE/MANIQA scores). Conversely, the super-resolution backbones that improve quality, such as StableSR and SeeSR, have materially lower detection rates on PixArt-Σ and FLUX.1 (Table 1). The paper should present detectability and quality as an explicit trade-off, as Section 4.3 already does, and align the abstract and conclusion with that trade-off.
minor comments (4)
  1. [Section 3.2.2 / Eq. (3)] Notation is inconsistent: the text defines the image encoder and text encoder as E_R and T_R, but Eq. (3) writes E(x_c) and T_R(x_c); please use consistent subscripts (e.g., E_R and T_R) throughout.
  2. [Appendix A.1] For TreeRing, the description says a ring pattern of radius 10 with zero-bit watermarking is used; please clarify whether the same ring key/pattern is used for every image and whether this matches the official TreeRing configuration, since this affects the generality of the reported p-values.
  3. [Abstract / Conclusion] The phrase 'maintain the best visual perception' and the conclusion's claim that watermarks 'can be successfully forged from a target image to any cover image' are too broad given Table 4; the concluding paragraph should state the actual conditions (UNet-family targets, certain watermark schemes, and the detectability-quality trade-off).
  4. [Conclusion] The sentence 'Watermarks offer an effective mean to manage the governance' should read 'an effective means'; similar typographical issues (e.g., 'perceptuality') should be corrected before publication.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: PnP's watermark injection is the attack mechanism itself, and its evaluation depends on external detectors and baselines; the universality over-claim is a correctness concern, not a circularity.

full rationale

The paper's derivation chain has no step in which an input is defined in terms of the output or a fitted parameter is relabeled as a prediction. Stage 1 estimates a watermark latent via proxy-model inversion (Eqs. 1-2); Stage 2 injects that latent into an off-the-shelf regenerative model (Eqs. 3-4). Success is then measured by external detectors (TreeRing p-value and Gaussian Shading bit accuracy/attribution), whose thresholds come from [9] or are recomputed by the same statistical procedure, not fitted to PnP's outputs. The attack's core operation—starting regeneration from the estimated watermark latent—is the intended forgery mechanism rather than a hidden circular loop. The paper's own limitations and Appendix C.1 (TreeRing detection rates 0.04–0.13 on PixArt-Σ/FLUX.1) weaken the 'universal' claim, and Section 4.2 concedes UNet-vs-DiT architectural gaps; but these are empirical refutations of an overbroad generalization, not definitional or self-citational circularity. No load-bearing argument reduces to a self-citation: reference [9] is the prior-work baseline and source of thresholds, not the authors' own work. Hence the central derivation is self-contained against external benchmarks and merits a circularity score of 0.

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

The attack introduces no new physical or architectural entities; it reuses existing proxy and regenerative models. The central claims rest on the transferability of latents across model families and the preservation of the watermark during regeneration, both of which are empirically only partially true.

free parameters (3)
  • TreeRing detection threshold for Animagine XL = computed by authors from 5000 watermarked and 5000 clean images (Appendix A.1)
    Used to compute Detection Success Rate for Animagine XL TreeRing results; it is a detector-side threshold fitted by the authors, not an attack constant.
  • DDIM sampling steps and guidance scale = 50 steps, guidance 7.5
    Applied to target, proxy, and regenerative pipelines; chosen by hand in Appendix A.2 and A.4, and they affect both detectability and quality.
  • Gaussian Shading bit accuracy thresholds = 0.5976525 (zero-bit) and 0.6484375 (user attribution)
    Taken from prior work [9]; used only for evaluation, not for the forgery method.
assumptions (3)
  • domain assumption Semantic watermarks are recoverable from a single image via DDIM inversion of the initial latent z_T.
    Invoked in Stage 1 (Section 3.1) to justify extracting the watermark from the target image; inherited from TreeRing [7], Gaussian Shading [8], and the prior attack [9].
  • domain assumption Latent spaces of different diffusion models are sufficiently aligned that a noise latent estimated with proxy SD2.1 transfers to the regenerative model and to detection by the target model.
    Stated in Section 3.2.3 as 'shared latent-space structure'; the paper's own Table 4 shows this fails for TreeRing on PixArt-Σ and FLUX.1.
  • domain assumption Regenerative diffusion models preserve the forged watermark signal under visual-textual conditioning.
    Core mechanism of Stage 2 (Equations 3 and 4); empirically tested across backbones, but no ablation demonstrates the watermark latent is necessary for detection.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Optimization-Free Universal Watermark Forgery with Regenerative Diffusion Models." pith.science (2026). https://pith.science/paper/SMCY4J74

@misc{pith2026250606018,
  author       = {Pith},
  title        = {Pith review of: Optimization-Free Universal Watermark Forgery with Regenerative Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SMCY4J74}},
  note         = {Machine review of arXiv:2506.06018}
}
read the original abstract

Watermarking becomes one of the pivotal solutions to trace and verify the origin of synthetic images generated by artificial intelligence models, but it is not free of risks. Recent studies demonstrate the capability to forge watermarks from a target image onto cover images via adversarial optimization without knowledge of the target generative model and watermark schemes. In this paper, we uncover a greater risk of an optimization-free and universal watermark forgery that harnesses existing regenerative diffusion models. Our proposed forgery attack, PnP (Plug-and-Plant), seamlessly extracts and integrates the target watermark via regenerating the image, without needing any additional optimization routine. It allows for universal watermark forgery that works independently of the target image's origin or the watermarking model used. We explore the watermarked latent extracted from the target image and visual-textual context of cover images as priors to guide sampling of the regenerative process. Extensive evaluation on 24 scenarios of model-data-watermark combinations demonstrates that PnP can successfully forge the watermark (up to 100% detectability and user attribution), and maintain the best visual perception. By bypassing model retraining and enabling adaptability to any image, our approach significantly broadens the scope of forgery attacks, presenting a greater challenge to the security of current watermarking techniques for diffusion models and the authority of watermarking schemes in synthetic data generation and governance.

Figures

Figures reproduced from arXiv: 2506.06018 by the authors.

Figure 1
Figure 1. Comparison of different Watermark forgery methods for diffusion models (DM): (a) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of our watermark forgery approach: [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Example images with forged watermark using the Imprint baseline and our proposed [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Quality of forged watermarks generated by [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Average over￾head required to forge watermark to a cover im￾age. Here, we discuss the feasibility of forgery attacks by measuring the average time required to embed a forged watermark into a single cover image, as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Additional example images with forged watermark using the Imprint baseline and our [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: Additional example images with forged watermark using the Imprint baseline and our [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]
Figure 8
Figure 8. Figure 8: Images generated with the Gaussian Shading watermark using target models. [PITH_FULL_IMAGE:figures/full_fig_p021_8.png]
Figure 9
Figure 9. Figure 9: Images generated with the Tree Ring watermark using target models. [PITH_FULL_IMAGE:figures/full_fig_p022_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 17 canonical work pages

  1. [1]

    Advancements in generative ai: A comprehensive review of gans, gpt, autoencoders, diffusion model, and transformers.IEEE Access, 12:69812–69837, 2024

    Staphord Bengesi, Hoda El-Sayed, MD Kamruzzaman Sarker, Yao Houkpati, John Irungu, and Timothy Oladunni. Advancements in generative ai: A comprehensive review of gans, gpt, autoencoders, diffusion model, and transformers.IEEE Access, 12:69812–69837, 2024

    work page 2024

  2. [2]

    Announcing microsoft copilot, your everyday ai companion

    Yusuf Mehdi. Announcing microsoft copilot, your everyday ai companion. https://blogs.microsoft.com/blog/2023/09/21/ announcing-microsoft-copilot-your-everyday-ai-companion/ , 2023. Accessed: 2025-05-15

    work page 2023

  3. [3]

    Identifying ai-generated im- ages with synthid

    Sven Gowal and Pushmeet Kohli. Identifying ai-generated im- ages with synthid. https://deepmind.google/discover/blog/ identifying-ai-generated-images-with-synthid/ , 2023. Accessed: 2025-05- 15

    work page 2023

  4. [4]

    Robust template matching for affine resistant image watermarks

    Shelby Pereira and Thierry Pun. Robust template matching for affine resistant image watermarks. IEEE Trans. Image Process., 9(6):1123–1129, 2000

    work page 2000

  5. [5]

    Robust image watermarking based on multiband wavelets and empirical mode decomposition.IEEE Trans

    Ning Bi, Qiyu Sun, Daren Huang, Zhihua Yang, and Jiwu Huang. Robust image watermarking based on multiband wavelets and empirical mode decomposition.IEEE Trans. Image Process., 16(8):1956–1966, 2007

    work page 1956

  6. [6]

    Watermark-based attribution of ai-generated content, 2024

    Zhengyuan Jiang, Moyang Guo, Yuepeng Hu, and Neil Zhenqiang Gong. Watermark-based attribution of ai-generated content, 2024

    work page 2024

  7. [7]

    Tree-rings watermarks: Invisible fingerprints for diffusion images

    Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-rings watermarks: Invisible fingerprints for diffusion images. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors,Advances in Neural Information Processing Systems, vol- ume 36, pages 58047–58063. Curran Associates, Inc., 2023

    work page 2023

  8. [8]

    Gaussian shading: Provable performance-lossless image watermarking for diffusion models

    Zijin Yang, Kai Zeng, Kejiang Chen, Han Fang, Weiming Zhang, and Nenghai Yu. Gaussian shading: Provable performance-lossless image watermarking for diffusion models. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12162–12171, 2024

    work page 2024

Show all 31 references
  1. [9]

    Black-box forgery attacks on semantic watermarks for diffusion models

    Andreas Müller, Denis Lukovnikov, Jonas Thietke, Asja Fischer, and Erwin Quiring. Black-box forgery attacks on semantic watermarks for diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2025

    2025

  2. [10]

    Pixart- α: Fast training of diffusion transformer for photorealistic text-to-image synthesis, 2023.URL https://arxiv

    Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart- α: Fast training of diffusion transformer for photorealistic text-to-image synthesis, 2023.URL https://arxiv. org/abs/2310, 426

    2023

  3. [11]

    Chan, and Chen Change Loy

    Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin C.K. Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. 2024

    2024

  4. [12]

    Diffbir: Toward blind image restoration with generative diffusion prior

    Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong. Diffbir: Toward blind image restoration with generative diffusion prior. InEuropean Conference on Computer Vision, pages 430–448. Springer, 2024. 10

    2024

  5. [13]

    Image watermarks are removable using controllable regeneration from clean noise.arXiv preprint arXiv:2410.05470, 2024

    Yepeng Liu, Yiren Song, Hai Ci, Yu Zhang, Haofan Wang, Mike Zheng Shou, and Yuheng Bu. Image watermarks are removable using controllable regeneration from clean noise.arXiv preprint arXiv:2410.05470, 2024

    2024 arXiv

  6. [14]

    Seesr: Towards semantics-aware real-world image super-resolution

    Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 25456–25467, 2024

    2024

  7. [16]

    Holisdip: Image super-resolution via holistic semantics and diffusion prior.arXiv preprint arXiv:2411.18662, 2024

    Li-Yuan Tsao, Hao-Wei Chen, Hao-Wei Chung, Deqing Sun, Chun-Yi Lee, Kelvin CK Chan, and Ming-Hsuan Yang. Holisdip: Image super-resolution via holistic semantics and diffusion prior.arXiv preprint arXiv:2411.18662, 2024

    2024 arXiv

  8. [17]

    Supermark: Robust and training-free image watermarking via diffusion-based super-resolution, 2024

    Runyi Hu, Jie Zhang, Yiming Li, Jiwei Li, Qing Guo, Han Qiu, and Tianwei Zhang. Supermark: Robust and training-free image watermarking via diffusion-based super-resolution, 2024

    2024

  9. [18]

    Toward real-world single image super-resolution: A new benchmark and a new model

    Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. InProceedings of the IEEE/CVF international conference on computer vision, pages 3086–3095, 2019

    2019

  10. [19]

    Component divide-and-conquer for real-world image super-resolution

    Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin. Component divide-and-conquer for real-world image super-resolution. InComputer Vision– ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part VIII 16, pag...

    2020

  11. [20]

    Towards real-world blind face restoration with generative facial prior

    Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9168–9178, 2021

    2021

  12. [21]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis.arXiv preprint arXiv:2307.01952, 2023

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis.arXiv preprint arXiv:2307.01952, 2023

    2023 arXiv

  13. [22]

    Flux.https://github.com/black-forest-labs/flux, 2024

    Black Forest Labs. Flux.https://github.com/black-forest-labs/flux, 2024

    2024

  14. [23]

    Animagine xl 3.0

    Cagliostro Research Lab. Animagine xl 3.0. https://huggingface.co/cagliostrolab/ animagine-xl-3.0, 2024. Accessed: 2025-05-16

    2024

  15. [24]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

    2022

  16. [25]

    Improv- ing the stability of diffusion models for content consistent super-resolution.arXiv preprint arXiv:2401.00877, 2024

    Lingchen Sun, Rongyuan Wu, Zhengqiang Zhang, Hongwei Yong, and Lei Zhang. Improv- ing the stability of diffusion models for content consistent super-resolution.arXiv preprint arXiv:2401.00877, 2024

    2024 arXiv

  17. [26]

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

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4):600– 612, 2004

    2004

  18. [27]

    The unrea- sonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unrea- sonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 586–595, 2018

    2018

  19. [28]

    Image quality assessment: Unifying structure and texture similarity.IEEE transactions on pattern analysis and machine intelligence, 44(5):2567–2581, 2020

    Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity.IEEE transactions on pattern analysis and machine intelligence, 44(5):2567–2581, 2020

    2020

  20. [29]

    Exploring clip for assessing the look and feel of images

    Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. InProceedings of the AAAI conference on artificial intelligence, volume 37, pages 2555–2563, 2023. 11

    2023

  21. [30]

    A feature-enriched completely blind image quality evaluator.IEEE Transactions on Image Processing, 24(8):2579–2591, 2015

    Lin Zhang, Lei Zhang, and Alan C Bovik. A feature-enriched completely blind image quality evaluator.IEEE Transactions on Image Processing, 24(8):2579–2591, 2015

    2015

  22. [31]

    Musiq: Multi-scale image quality transformer

    Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. InProceedings of the IEEE/CVF international conference on computer vision, pages 5148–5157, 2021

    2021

  23. [32]

    Maniqa: Multi-dimension attention network for no-reference image quality assessment

    Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniqa: Multi-dimension attention network for no-reference image quality assessment. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages...

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.