Pith. sign in

REVIEW 3 major objections 5 minor 38 references

Is Perturbation-Based Image Protection Disruptive to Image Editing?

T0 review · 3 major / 5 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 Perturbation-based image protection does not fully disrupt diffusion editing; protected images often follow text prompts as well as or better than originals.

desk verdict Useful empirical study that measures prompt alignment but not content preservation, so its central claim overreaches. read the letter →

arxiv 2506.04394 v2 pith:6DTCUXGT submitted 2025-06-04 cs.CV

Qiuyu Tang , Bonor Ayambem , Mooi Choo Chuah , Aparna Bharati This is my paper
classification cs.CV
keywords diffusionmodelsadversarialperturbationsimageprotectioneditingtext-to-imagealignmentstylemimicryCLIP-SPAC-S++
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 tests whether the most common family of image-protection techniques, adding small adversarial perturbations to an image before it is posted, can stop a malicious user from editing it with Stable Diffusion. Drawing on image-to-image editing and style-transfer experiments over natural photos and artworks, it argues that this kind of protection does not fully work: in most conditions, an edit generated from a protected image follows the user's text prompt at least as closely as an edit from the unprotected original, and often more closely. The authors trace the effect to the diffusion process itself, where protection noise blends with the noise that diffusion already adds and removes, giving the prompt more room to rewrite content. If correct, the finding matters because it says widely used protections can leave artists and photographers exposed to mimicry even when the perturbations look successful.

What carries the argument

Two linked objects carry the argument. The measurement object is the Percentage Change in Image-Text Alignment (ITA), defined as the change in CLIP-S or PAC-S++ score between edits from protected and unprotected images divided by the unprotected score; it is used to decide whether protection disrupts or improves prompt adherence. The explanatory object is the diffusion model's two Markov chains, a forward process that adds Gaussian noise to the latent and a reverse denoising process conditioned on the text prompt, because adversarial perturbations act as extra noise that accumulates through the forward chain and gives the prompt more latitude to replace content during denoising.

What would settle it

Present paired protected-versus-original edits to human raters and ask which better matches the prompt and looks realistic; if raters do not prefer protected-input edits in conditions where CLIP-S and PAC-S++ show large positive changes, the central claim collapses.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that perturbation-based protection methods (PhotoGuard, Mist, and Glaze) do not reliably produce the fully protected outcome of a noisy, irrelevant image; instead they often make the edited output align better with the text prompt. Across five seeds on Flickr8k, edits from PhotoGuard-protected images had positive Percentage Change in Image-Text Alignment relative to original-image edits in the majority of cases (CLIP-S 54.23% and 61.54%, PAC-S++ 61.82% and 67.79% for close and far captions), and the stylization experiments showed the same trend for natural images and artworks. The paper also shows that replacing the adversarial perturbation with plain Gaussian noise reproduces the positive Percentage Change, and offers a mechanism: the perturbation adds extra noise to the diffusion latents at every timestep, and the reverse denoising process, guided by the text prompt, replaces those noisier regions with prompt-aligned content. It concludes that current perturbation-based methods are not a sufficient solution against inference-based diffusion editing.

Load-bearing premise

The conclusion rests on CLIP-S and PAC-S++ being unbiased measures of whether an edited image is a desirable exploiter output; if these scores reward noisier or more synthetic images for reasons unrelated to semantic alignment, the positive Percentage Change would not show that protection helps the attacker.

Editorial extensions

If this is right

  • Image-to-image editing of protected Flickr8k photos follows the close and far captions as well as or better than editing of originals in the majority of trials, so protection does not consistently create the noisy, unrelated outputs its design targets.
  • Style-mimicry edits of protected natural images and artworks also show positive Percentage Change in the majority of cases (e.g., 77.54% with PhotoGuard on Flickr1024 by CLIP-S), meaning the cloaks can improve stylization instead of breaking it.
  • Because Gaussian noise with standard deviation 5 produces the same trend, the effect is not specific to the carefully optimized PhotoGuard perturbations; any small input noise is predicted to boost prompt adherence.
  • The paper's conclusion implies that protection methods should be judged by whether edited outputs still align with exploitation prompts, not only by whether perturbations are visible or by attack-specific loss values.

Reading between the lines

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

  • If the noise-mechanism explanation is right, the effect should grow with perturbation budget and with diffusion strength; running PhotoGuard at multiple epsilon values and measuring ITA would test that prediction directly.
  • The same mechanism suggests a possible inversion: deliberately adding tuned noise before a legitimate edit could be used as a cheap prompt-alignment enhancement, although the paper does not make this recommendation.
  • The Gaussian-noise result raises the possibility that any imperceptible input-space perturbation, including methods not tested here, faces the same fundamental limit against inference-based diffusion editing.
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 / 5 minor

Summary. The paper evaluates whether perturbation-based image protection methods (PhotoGuard, Mist, Glaze) actually disrupt diffusion-based image editing. Using Stable Diffusion v1.5, the authors measure image-text alignment (CLIP-S and PAC-S++) between generated edits and the prompting text for protected versus original images, across natural scene images (Flickr8k, Flickr1024) and artworks (WikiArt), with five generation seeds. They report that in most configurations the protected-image edits attain equal or higher image-text alignment than edits from clean images, and they reproduce the trend with Gaussian noise as a control. They offer a qualitative theoretical explanation based on noise accumulation during diffusion sampling and conclude that perturbation-based protection does not fully prevent successful editing.

Significance. The question is timely and important for the growing literature on protecting images from diffusion-based manipulation. The experimental design has real strengths: five seeds, multiple datasets and domains, two alignment metrics, and a Gaussian-noise control that strengthens the causal interpretation of the trend. The paper also promises code/data release. If the headline conclusion is confirmed with a fuller evaluation, the work would be a useful cautionary result for the image-protection community. However, the central claim currently outruns the measured quantities because content preservation relative to the original image—an explicit part of the paper's own success criterion—is never measured in the quantitative analysis. The significance of the result is therefore conditional on an additional evaluation, which is feasible within the manuscript's scope.

major comments (3)
  1. [§1, Eq. (1), Tables 1–2] The paper's own definition of successful exploitation requires the edited output to 'share some content with the reference' (Sec. 1), yet every quantitative result (Eq. 1, Tables 1–2, Figs. 3 and 8) measures only image-text alignment via CLIP-S/PAC-S++. No content-similarity metric relative to the original input is reported. Without this, a positive Percentage Change in ITA is also consistent with the protection perturbation destroying the original content and the model regenerating a generic image that matches the prompt alone. The Gaussian-noise control in Fig. 8 and the authors' own explanation in Fig. 7—that protected latents are noisier and text influence is stronger—are exactly what one would expect under this alternative interpretation. I therefore do not regard the central claim that protections 'do not fully disrupt editing' as established until content preservation is measured per condition, for example with LPIPS/SSIM or a human pairwise study.
  2. [§4 (Image Stylization), Table 2] For stylization, the ITA metric is computed against the style prompt only (e.g., 'change the style to Cubism'). Such a score rewards any output that contains Cubist elements, including a generic Cubist scene that shares no content with the input photograph. The validation in Fig. 6 only shows that CLIP-S and PAC-S++ can distinguish text-to-image generations of different styles; it does not show that these metrics capture successful style transfer on top of preserved content. The stylization rows in Table 2 therefore do not measure the stated success criterion. Please add a content-preservation measure (e.g., perceptual similarity to the input, or a user study asking whether the original subject is retained while the style is changed) and report ITA conditional on content being preserved.
  3. [§3.1, Fig. 3, Tables 1–2] No error bars, confidence intervals, or paired significance tests are reported anywhere, and the main text reports only aggregate percentages of images with non-negative Actual Change. Several of these percentages are close to chance (e.g., Glaze CLIP-S 54.12% and Mist CLIP-S 54.15% in Table 2), so the claim that 'most scenarios' show enhanced alignment is not strongly supported without a characterization of variance across the five seeds and across images. Please provide per-seed and per-image distributions, effect sizes, and a paired test (e.g., Wilcoxon signed-rank) on the ITA differences rather than only aggregated percentages.
minor comments (5)
  1. [Theoretical Explanation, Eq. (3)] The conditional distribution is written as q(z_t | z_0) = N(z_t; sqrt(α_t) z_{t-1}, (1 - α_t) I), but conditioning on z_0 should give mean sqrt(α_t) z_0, not z_{t-1}; as written this is not the standard forward-process marginal and confuses the derivation.
  2. [Theoretical Explanation, after Eq. (3)] The notation α_t is overloaded: 'α_t = 1 - β_t, α_t = Q_t_{s=1} α_s' uses the same symbol for both a per-step value and the cumulative product. Please use an overbar, e.g., \bar{α}_t, for the cumulative product.
  3. [§3.1] The number of Flickr8k images actually used in the experiments is not stated, only the dataset size. Please report the sample sizes for each condition, and likewise for the Flickr1024 and WikiArt subsets used for stylization.
  4. [Fig. 3 and Fig. 8] The y-axis label says 'Percent Change' but it is unclear whether the plotted values are means over images, medians, or some other aggregate; please clarify the aggregation and include error bars or per-seed markers.
  5. [Table 2] The PAC-S++ 'Actual Change ≥ 0' percentages for Mist and Glaze are identical (56.90%) even though they are computed on the same WikiArt set for different protection methods; this may be true by coincidence, but it merits a brief note or a check for a copy/paste error.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: central result is empirical with independent Gaussian control; self-citations are contextual and not load-bearing.

full rationale

The paper's claim that perturbation-based protections can enhance prompt alignment is an empirical measurement, not a derivation from its own inputs. Equation 1 defines Percentage Change against the clean-image edit's ITA, and the reported positive changes are data-dependent; nothing in the definition forces them to be positive. The 'theoretical explanation' in Section 4 uses standard LDM definitions (Eqs. 2-4) and a qualitative noise-accumulation story, but no parameters are fitted to the outputs, and the Gaussian-noise control (Fig. 8) provides an independent baseline test. The self-citations ([19], [27]) appear only in related-work lists or as metric context; the central experiments use externally developed metrics (CLIP-S, PAC-S++) and public datasets, so no load-bearing argument reduces to the authors' own prior work. The skeptic's concern that content preservation is never measured is a real construct-validity gap: the paper's own definition of successful exploitation (Sec. 1) requires that the output 'shares some content with the reference and the generated edits align with the guidance prompts,' but Eq. 1 and Tables 1-2 quantify only image-text alignment. However, that is an omitted variable, not a circular reduction: the metric is not defined in terms of the conclusion, and the paper does not claim to measure content preservation. Therefore no self-definitional, fitted-input, or self-citation-circular step is present.

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

The paper introduces no invented entities and no free parameters. Its central claim rests on standard diffusion equations, on the validity of text-image alignment metrics as proxies for exploit success, and on the heuristic assumption that protection perturbations behave like diffusion noise in the latent space.

assumptions (3)
  • standard math The forward diffusion process adds Gaussian noise and the reverse process removes noise conditioned on the text prompt (Eqs. 2-4).
    Standard latent diffusion formulation from Rombach et al. [3], used as the basis for interpreting perturbation effects.
  • domain assumption Image-text alignment metrics CLIP-S and PAC-S++ are valid measures of edit success, i.e., higher scores mean closer adherence to the prompt.
    The entire quantitative evaluation relies on these scores; the paper validates style understanding indirectly via Fig. 6, but does not validate the protected versus unprotected comparison directly.
  • ad hoc to paper Protection perturbations accumulate in the latent space during diffusion sampling and are equivalent to adding noise, thereby increasing prompt influence.
    This is the paper's 'theoretical explanation'; it is asserted heuristically and is not derived from the LDM equations, nor quantified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Is Perturbation-Based Image Protection Disruptive to Image Editing?." pith.science (2026). https://pith.science/paper/6DTCUXGT

@misc{pith2026250604394,
  author       = {Pith},
  title        = {Pith review of: Is Perturbation-Based Image Protection Disruptive to Image Editing?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6DTCUXGT}},
  note         = {Machine review of arXiv:2506.04394}
}
read the original abstract

The remarkable image generation capabilities of state-of-the-art diffusion models, such as Stable Diffusion, can also be misused to spread misinformation and plagiarize copyrighted materials. To mitigate the potential risks associated with image editing, current image protection methods rely on adding imperceptible perturbations to images to obstruct diffusion-based editing. A fully successful protection for an image implies that the output of editing attempts is an undesirable, noisy image which is completely unrelated to the reference image. In our experiments with various perturbation-based image protection methods across multiple domains (natural scene images and artworks) and editing tasks (image-to-image generation and style editing), we discover that such protection does not achieve this goal completely. In most scenarios, diffusion-based editing of protected images generates a desirable output image which adheres precisely to the guidance prompt. Our findings suggest that adding noise to images may paradoxically increase their association with given text prompts during the generation process, leading to unintended consequences such as better resultant edits. Hence, we argue that perturbation-based methods may not provide a sufficient solution for robust image protection against diffusion-based editing.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 30 canonical work pages

  1. [1]

    Is Perturbation-Based Image Protection Disruptive to Image Editing?

    INTRODUCTION Generative models [1, 2, 3, 4, 5, 6] have become increasingly popular and have been enthusiastically adopted within the AI research and creative arts communities. Among various gen- erative techniques, diffusion-based models [3, 4, 5, 6, 7] are widely favored for their high-quality image generation and straightforward usability. However, thei...

    work page Pith review arXiv

  2. [2]

    Evaluation of Protection Methods : Investigating the effectiveness of current protection methods against inference-based editing by analyzing the alignment between generated content and guidance prompts, pro- viding insights into their usefulness and robustness

    work page

  3. [3]

    Comprehensive Domain Analysis: Experiments con- sidering natural scene and artwork images for potential exploitation through content and style-based editing

    work page

  4. [4]

    Theoretical Explanation: We provide a theoretical ex- planation for the above observations

    work page

  5. [5]

    They achieve this via a forward process where noise is added to an image and a reverse process which denoises them

    RELATED WORK Diffusion-based Models - Models such as [3, 7] have demon- strated superior image generation performance by faithfully adhering to user input and ensuring realism in the generated images. They achieve this via a forward process where noise is added to an image and a reverse process which denoises them. Stable Diffusion (SD) [3], a popular mod...

    work page

  6. [6]

    change the style to [V]

    IMPACT OF PROTECTION ON IMAGE EDITING To evaluate the performance of image protection against inference-based editing via pre-trained Stable Diffusion [3], we conduct experiments on three state-of-the-art protection methods that employ meticulously crafted adversarial pertur- bations, PhotoGuard [15], Mist [20], and Glaze [16]. Photo- Guard protects natur...

    work page

  7. [7]

    RESULTS & DISCUSSION Our analysis for the image-to-image editing task includes four experimental settings based on the ITA scores for close and far captions as shown in Fig. 3. The positive Percent- age Changes indicate that protected images achieve stronger alignment with text prompts than original images, highlight- ing the influence of adversarial prot...

    work page

  8. [8]

    Due to the inherent properties of diffusion-based models, these perturbations introduce additional flexibility when editing noise pixels, further reinforcing adherence to prompts

    CONCLUSION Our experiments show that perturbations can inadvertently enhance prompt alignment instead of fully disrupting image editing. Due to the inherent properties of diffusion-based models, these perturbations introduce additional flexibility when editing noise pixels, further reinforcing adherence to prompts. Consequently, current protection methods...

    work page

Show all 38 references
  1. [9]

    Generative adversarial nets,

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio, “Generative adversarial nets,” NeurIPS, vol. 27, 2014

    2014

  2. [10]

    A style-based generator architecture for generative adversarial networks,

    Tero Karras, Samuli Laine, and Timo Aila, “A style-based generator architecture for generative adversarial networks,” in CVPR, 2019, pp. 4401–4410

    2019

  3. [11]

    High-resolution image synthesis with latent diffusion models,

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer, “High-resolution image synthesis with latent diffusion models,” in CVPR, 2022, pp. 10684– 10695

    2022

  4. [12]

    Dreambooth: Fine tuning text-to-image diffusion models for subject-driven gen- eration,

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman, “Dreambooth: Fine tuning text-to-image diffusion models for subject-driven gen- eration,” in CVPR, 2023, pp. 22500–22510

    2023

  5. [13]

    An image is worth one word: Personalizing text-to-image generation using textual inversion,

    Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or, “An image is worth one word: Personalizing text-to-image generation using textual inversion,” arXiv:2208.01618, 2022

    2022 arXiv

  6. [14]

    Multi-concept customization of text- to-image diffusion,

    Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shecht- man, and Jun-Yan Zhu, “Multi-concept customization of text- to-image diffusion,” in CVPR, 2023, pp. 1931–1941

    2023

  7. [15]

    Sdedit: Guided image synthesis and editing with stochastic differential equations,

    Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon, “Sdedit: Guided image synthesis and editing with stochastic differential equations,” arXiv:2108.01073, 2021

    2021 arXiv

  8. [16]

    An a.i.-generated picture won an art prize. artists aren’t happy.,

    Kevin Roose, “An a.i.-generated picture won an art prize. artists aren’t happy.,” The New York Times, 2022

    2022

  9. [17]

    Are a.i.-generated pictures art?,

    Natalie Proulx, “Are a.i.-generated pictures art?,” The New York Times, 2022

    2022

  10. [18]

    Clipscore: A reference-free evaluation metric for image captioning,

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi, “Clipscore: A reference-free evaluation metric for image captioning,” arXiv:2104.08718, 2021

    2021 arXiv

  11. [19]

    Positive-augmented contrastive learning for image and video captioning evaluation,

    Sara Sarto, Manuele Barraco, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara, “Positive-augmented contrastive learning for image and video captioning evaluation,” inCVPR, 2023, pp. 6914–6924

    2023

  12. [20]

    Positive-augmented contrastive learning for vision-and-language evaluation and training,

    Sara Sarto, Nicholas Moratelli, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara, “Positive-augmented contrastive learning for vision-and-language evaluation and training,” arXiv:2410.07336, 2024

    2024 arXiv

  13. [21]

    Laion-5b: An open large-scale dataset for training next gen- eration image-text models,

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al., “Laion-5b: An open large-scale dataset for training next gen- eration image-text models,” NeurIPS, vol. 35, 2022

    2022

  14. [22]

    Safe latent diffusion: Mitigating inappro- priate degeneration in diffusion models,

    Patrick Schramowski, Manuel Brack, Bj ¨orn Deiseroth, and Kristian Kersting, “Safe latent diffusion: Mitigating inappro- priate degeneration in diffusion models,” in CVPR, 2023, pp. 22522–22531

    2023

  15. [23]

    Raising the cost of malicious ai- powered image editing,

    Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry, “Raising the cost of malicious ai- powered image editing,” arXiv:2302.06588, 2023

    2023 arXiv

  16. [24]

    Glaze: Protecting artists from style mimicry by text-to-image models,

    Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y . Zhao, “Glaze: Protecting artists from style mimicry by text-to-image models,” in USENIX Security 23, 2023, pp. 2187–2204

    2023

  17. [25]

    Anti-dreambooth: Protecting users from personalized text-to-image synthesis,

    Thanh Van Le, Hao Phung, Thuan Hoang Nguyen, Quan Dao, Ngoc N Tran, and Anh Tran, “Anti-dreambooth: Protecting users from personalized text-to-image synthesis,” in ICCV, 2023, pp. 2116–2127

    2023

  18. [26]

    Metacloak: Preventing unautho- rized subject-driven text-to-image diffusion-based synthesis via meta-learning,

    Yixin Liu, Chenrui Fan, Yutong Dai, Xun Chen, Pan Zhou, and Lichao Sun, “Metacloak: Preventing unautho- rized subject-driven text-to-image diffusion-based synthesis via meta-learning,” in CVPR, 2024, pp. 24219–24228

    2024

  19. [27]

    Watermarks vs. perturba- tions for preventing AI-based style editing,

    Qiuyu Tang and Aparna Bharati, “Watermarks vs. perturba- tions for preventing AI-based style editing,” in The 1st Work- shop on GenAI Watermarking, 2025

    2025

  20. [28]

    Mist: Towards improved adversarial examples for diffusion models,

    Chumeng Liang and Xiaoyu Wu, “Mist: Towards improved adversarial examples for diffusion models,”arXiv:2305.12683, 2023

    2023 arXiv

  21. [29]

    Adversarial example does good: Preventing paint- ing imitation from diffusion models via adversarial examples,

    Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yim- ing Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan, “Adversarial example does good: Preventing paint- ing imitation from diffusion models via adversarial examples,” arXiv:2302.04578, 2023

    2023 arXiv

  22. [30]

    Adversarial perturbations cannot reliably protect artists from generative ai,

    Robert H ¨onig, Javier Rando, Nicholas Carlini, and Florian Tram`er, “Adversarial perturbations cannot reliably protect artists from generative ai,” arXiv:2406.12027, 2024

    2024 arXiv

  23. [31]

    Fram- ing image description as a ranking task: Data, models and eval- uation metrics,

    Micah Hodosh, Peter Young, and Julia Hockenmaier, “Fram- ing image description as a ranking task: Data, models and eval- uation metrics,” Journal of Artificial Intelligence Research , vol. 47, pp. 853–899, 2013

    2013

  24. [32]

    Claude 3.5 sonnet,

    Anthropic, “Claude 3.5 sonnet,” https://www. anthropic.com/claude, 2024

    2024

  25. [33]

    Universal sentence encoder for english,

    Daniel Cer, Yinfei Yang, Sheng-yi Kong, Nan Hua, Nicole Limtiaco, Rhomni St John, Noah Constant, Mario Guajardo- Cespedes, Steve Yuan, Chris Tar, et al., “Universal sentence encoder for english,” in EMNLP, 2018, pp. 169–174

    2018

  26. [34]

    No-reference image quality assessment in the spatial domain,

    Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik, “No-reference image quality assessment in the spatial domain,” IEEE Transactions on Image Processing, vol. 21, no. 12, pp. 4695–4708, 2012

    2012

  27. [35]

    Exploring saliency bias in manipulation de- tection,

    Joshua Krinsky, Alan Bettis, Qiuyu Tang, Daniel Moreira, and Aparna Bharati, “Exploring saliency bias in manipulation de- tection,” in 2024 IEEE ICIP, 2024, pp. 3257–3263

    2024

  28. [36]

    Learning trans- ferable visual models from natural language supervision,

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al., “Learning trans- ferable visual models from natural language supervision,” in ICML, 2021, pp. 8748–8763

    2021

  29. [37]

    Flickr1024: A large-scale dataset for stereo image super-resolution,

    Yingqian Wang, Longguang Wang, Jungang Yang, Wei An, and Yulan Guo, “Flickr1024: A large-scale dataset for stereo image super-resolution,” in ICCV Workshops, Oct 2019, pp. 3852–3857

    2019

  30. [38]

    Improved artgan for conditional synthesis of natural image and artwork,

    Wei Ren Tan, Chee Seng Chan, Hernan Aguirre, and Kiyoshi Tanaka, “Improved artgan for conditional synthesis of natural image and artwork,” IEEE Transactions on Image Processing, vol. 28, no. 1, pp. 394–409, 2019

    2019

Pith tools

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