Pith. sign in

REVIEW 2 major objections 5 minor 1 cited by

Stay-Positive: A Case for Ignoring Real Image Features in Fake Image Detection

T0 review · 2 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read The paper claims that fake-image detectors become more robust and general when their final layer is forced to ignore real-image features.

desk verdict Simple, cheap, and often helpful retraining trick, but the missing no-positivity control means the paper never proves its central mechanism. read the letter →

arxiv 2502.07778 v2 pith:IUJ6D474 submitted 2025-02-11 cs.CV

Anirudh Sundara Rajan , Yong Jae Lee This is my paper · ORCID
classification cs.CV
keywords fakeimagedetectionspuriouscorrelationslast-layerretrainingnon-negativeweightslatentdiffusionmodelsforensicsgeneralizationpost-processingrobustness
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 argues that a fake-image detector should decide based only on artifacts introduced by the generator, and that any feature the network associates with real images is at best unnecessary and at worst a spurious correlation that hurts generalization. It proposes Stay-Positive, a last-layer retraining procedure that constrains the final linear layer's weights to be non-negative, so the detector can only increase its fake score from features that are present, never decrease it using real-image patterns. On detectors trained for latent diffusion images, this retraining improves robustness to WEBP compression and resizing, lifts average precision on newer generators such as FLUX and aMUSEd, and makes partially inpainted real images detectable. The gain matters because real training images often carry unknown post-processing artifacts, causing ordinary detectors to misclassify post-processed fakes as real.

What carries the argument

Stay-Positive is the central mechanism: a two-stage procedure that first trains an ordinary detector, then retrains only the final linear layer on the frozen backbone with the projection step $w \leftarrow \max(w,0)$ after each SGD update, so every weight is non-negative. The argument rests on the decomposition of the pre-sigmoid score into a real score $\sum_{w_i<0} w_i h_i$ and a fake score $\sum_{w_i>0} w_i h_i$, where $h=\gamma(g_\phi(x))$ is the non-negative ReLU feature vector; forcing all weights non-negative zeroes out the real score by construction. The last-layer-only retraining is what prevents the backbone from learning to use the absence of real features as a fake cue, which the authors observe when the constraint is applied to the full network.

What would settle it

Take a detector trained on LSUN-inclusive real data, apply Stay-Positive, and measure average precision on WEBP-compressed fake images; if a version trained on COCO-only real data still beats the Stay-Positive version by a large margin, the non-negativity constraint has not eliminated the real-side spurious correlation as claimed. Alternatively, inspect the final-layer feature activations on real images: if any feature that is active on real images receives a positive weight in the retrained layer, the asserted clean separation between real and fake features is violated.

Watch

Extended reading notes

Core claim

The central claim is that the sign of a final-layer weight identifies whether a feature signals real or fake content: because the features are ReLU-activated and non-negative, a negative weight can only lower the pre-sigmoid score and thus marks a real-image feature, while a positive weight marks a fake-image feature. Based on this, the authors show that retraining only the last layer with the constraint that all weights stay non-negative -- freezing the backbone and projecting weights to $\max(w,0)$ after each update -- produces detectors that ignore real-side features and consequently suffer less from spurious correlations. The resulting detectors match or beat their base versions on already-easy generators and substantially improve on FLUX and aMUSEd, remain accurate when fake images are WEBP-compressed or downscaled, and detect partially inpainted real images at high average precision.

Load-bearing premise

The method assumes that the sign of a final-layer weight cleanly separates real-image features from fake-image features, so that erasing all negative weights truly removes real-image influence; in reality a negative weight can serve other purposes, and the paper's own appendix shows the model can learn the absence of a real feature as a fake cue.

Editorial extensions

If this is right

  • Detectors retrained with Stay-Positive are more robust to WEBP compression and downsizing without any knowledge of which specific artifact caused the spurious correlation.
  • Generalization to newer generators in the same family (FLUX, aMUSEd) improves sharply, suggesting part of the cross-generator gap comes from real-side spurious features rather than from novel fake artifacts.
  • Partially inpainted real images, which are mostly real, become detectable because the detector no longer leans on real-image features.
  • The procedure is cheap: last-layer retraining converges in about 15 epochs and improves existing detectors without changing the backbone.

Reading between the lines

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

  • If the sign-based separation is trustworthy, the same last-layer constraint could serve as a diagnostic: decomposing a detector's score into real and fake components would reveal which features drive failures on new generators.
  • The paper's appendix shows the model can learn the absence of a real feature as a fake cue, so a natural extension is a full-network training objective that explicitly penalizes that negation behavior; the authors hint at this but do not implement it.
  • One could test the core mechanism directly by ablating the clean-data condition: train the same base detector on a real set with no compression artifacts, then compare its generalization to the Stay-Positive retrained version; the paper's own Figure 1 suggests the clean-data baseline may still fall short, hinting that last-layer retraining does not fully remove real-side influence.
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

2 major / 5 minor

Summary. The paper proposes Stay-Positive, a last-layer retraining procedure for fake image detectors. The authors argue that in a detector with a ReLU feature head and sigmoid output, negative final-layer weights correspond to features associated with real images. Stay-Positive retrains the final linear layer while projecting weights to non-negative values, forcing the detector to rely only on positive (fake) features. The method is applied to two LDM detectors (Corvi and Rajan) and evaluated on robustness to WEBP compression and resizing, generalization to unseen generators (FLUX, aMUSEd, etc.), and detection of partially inpainted real images. The paper reports consistent improvements on several benchmarks, including large gains on FLUX and on inpainted images, and includes ablations of clamping without retraining and full-network retraining.

Significance. If the causal claim is established, the contribution is practically significant: a simple, cheap last-layer procedure that improves existing detectors without full retraining, with a clear and testable hypothesis about spurious correlations. The experimental coverage is broad (WebP, downsizing, GenImage, UFD, GANs, inpainting), and the paper is candid about limitations. However, the central attribution of the improvements to the non-negative constraint is not currently supported by the ablations, because the natural control—last-layer retraining without the projection—is missing. The reported gains could plausibly come from linear probing or initialization effects. This is fixable within the manuscript's scope, so the paper merits revision rather than rejection.

major comments (2)
  1. [Algorithm 1; Section 5.5, Table 4] The causal claim that the Stay-Positive constraint (projection w←max(w,0)) drives the improvements is not established. Algorithm 1 also (a) re-initializes the head at w=0, (b) freezes the backbone, and (c) retrains the head with binary cross-entropy. Table 4 ablates clamping without retraining and full-network retraining, but it does not include the natural control: run Algorithm 1 exactly as written while omitting the projection step. Without that control, the improvements in Tables 1, 2, and 5 could be due to linear probing of frozen features, the zero initialization, or the restricted hypothesis class, rather than the non-negativity constraint. Please add this baseline and report it alongside Table 4.
  2. [Section 3.2 and Appendix A.6] The paper's interpretation that weight sign cleanly identifies whether a feature is 'real' or 'fake' is an assumption that the authors themselves show can fail. Appendix A.6 describes a toy network that turns a real-image feature (WEBP presence) into a fake-detection cue by learning a transformed feature that activates when WEBP is absent. In such a network, a positive final-layer weight can correspond to a feature derived from the real distribution, and a negative weight can correspond to the presence of a real artifact. This does not invalidate the empirical results, but it weakens the paper's central narrative (title and Section 4) that Stay-Positive makes the detector 'focus purely on fake artifacts.' The authors should either provide evidence that the sign-to-semantics mapping holds for the actual detectors or temper the mechanistic claims.
minor comments (5)
  1. [Tables 1, 2, 5, 6] Please report confidence intervals or standard deviations for the AP/accuracy values, or at least indicate the variance across the two trained networks mentioned in Appendix A.1.
  2. [Section 5.2.1] The LAION-Aesthetics reference appears as ' ?' and needs a proper citation.
  3. [Tables 1, 2, 4] The table headers 'A VG' (Tables 1, 2, and 4) and 'W UERSTCHEN A MUSE D' (Table 1) appear to contain spacing typos; please fix.
  4. [Section 5.6] Please proofread for typos, e.g., 'senstitvity' in Section 5.6.
  5. [Reproducibility] The project page link provides implementation details but no code or checkpoints; making code available would strengthen reproducibility.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the Stay-Positive constraint is an architectural/optimization intervention, and the reported gains are evaluated on external held-out generators and public benchmarks rather than derived from the method's own definitions.

full rationale

The paper's core derivation in Section 3.2 is a mathematical identity, not a circular argument: for a ReLU-feature linear head, f_theta(x)=sigma(sum_{w_i>0} w_i h_i + sum_{w_i<0} w_i h_i + b), so a negative final-layer weight lowers the pre-sigmoid score when its feature is active. The paper explicitly states this as an assumption ('we assume... a ReLU activation... and a sigmoid activation') and uses it only to justify the projection w <- max(w,0) in Algorithm 1. The algorithm is not fitted to the evaluation data; it retrains the last layer on the original training distribution and is then tested on unseen generators (FLUX, aMUSEd, VQDM, GLIDE, ADM, DALL-E), post-processed images, and public benchmarks (GenImage, UFD). The self-references to Corvi et al. (2023) and Rajan et al. (2025) are normal: those are the base detectors being improved, and the comparison against external baselines such as AEROBLADE, UFD, ClipDet, and DRCT provides independent support. The reader's concern about the missing control (last-layer retraining without the non-negative projection) is a legitimate experimental-completeness issue, but it is not circularity: omitting a control does not make the reported predictions equivalent to the method's inputs. The paper's own limitations and Appendix A.6 concede that the model can still use 'absence of a real feature' as a fake cue and that fake-specific spurious correlations (e.g., upsampling) remain; these admissions weaken the claimed mechanism but do not make the empirical comparison circular. No step in the derivation chain reduces by construction to a fitted parameter or to a self-citation.

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

The method rests on two interpretive assumptions about the final-layer weights, and on the requirement that the network has a ReLU before the last layer. No new physical or algorithmic entities are introduced.

assumptions (3)
  • domain assumption The detector's last-layer input is passed through a ReLU activation, making all feature values non-negative.
    Used in Section 2.2 and Section 3.2 to interpret weight signs; if a detector lacks ReLU before the final layer, the method does not apply.
  • ad hoc to paper A negative weight in the final layer indicates the feature is associated with real images, and a positive weight with fake images.
    Central hypothesis of Section 3.2; not proven and may be false for trained networks.
  • domain assumption The fake distribution consists of images from a known generator family, so it is meaningful to talk about 'fake artifacts' common to that family.
    The method is evaluated on LDM-family and GAN-family detectors, but the assumption is implicit throughout.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Stay-Positive: A Case for Ignoring Real Image Features in Fake Image Detection." pith.science (2026). https://pith.science/paper/IUJ6D474

@misc{pith2026250207778,
  author       = {Pith},
  title        = {Pith review of: Stay-Positive: A Case for Ignoring Real Image Features in Fake Image Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IUJ6D474}},
  note         = {Machine review of arXiv:2502.07778}
}
read the original abstract

Detecting AI generated images is a challenging yet essential task. A primary difficulty arises from the detectors tendency to rely on spurious patterns, such as compression artifacts, which can influence its decisions. These issues often stem from specific patterns that the detector associates with the real data distribution, making it difficult to isolate the actual generative traces. We argue that an image should be classified as fake if and only if it contains artifacts introduced by the generative model. Based on this premise, we propose Stay Positive, an algorithm designed to constrain the detectors focus to generative artifacts while disregarding those associated with real data. Experimental results demonstrate that detectors trained with Stay Positive exhibit reduced susceptibility to spurious correlations, leading to improved generalization and robustness to post processing. Additionally, unlike detectors that associate artifacts with real images, those that focus purely on fake artifacts are better at detecting inpainted real images.

Figures

Figures reproduced from arXiv: 2502.07778 by the authors.

Figure 1
Figure 1. Sensitivity to WEBP Compression. Using the LSUN dataset, which contains WEBP compressed images, as part of the real distribution makes the network highly vulnerable to WEBP compression. confuses WEBP-compressed fake images with real images, and if this confusion is driven by the inclusion of LSUN images in the real distribution. To do so, we adopt the ex￾perimental setup proposed by Rajan et al. (2025), where fake i… view at source ↗
Figure 2
Figure 2. Image Quality-Based Spurious Features. Corvi out￾puts a higher real score for Flux reconstructions compared to LDM reconstructions demonstrating the spurious nature of these real fea￾tures. Fake Score reduces due to the use of a different generator. presence of certain fine details with real images. However, this hypothesis does not hold, as Dai et al. (2023) also show that 16-channel autoencoder-based models, such … view at source ↗
Figure 3
Figure 3. Our key idea involves 2 steps. (1) We first train a fake image detector in the standard way without any modifications. This detector focuses on both real and fake features. (2) We re-train the last layer of the network such that it only focuses on the fake features to make a decision. a monotonically increasing sigmoid function, therefore we can infer that the final score, represented as wT h, for an ideal detector … view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Improved Robustness to WEBP Compression. Com￾pared to the original Corvi and Rajan, our detectors Corvi⊕ and Rajan⊕ show increased robustness towards WEBP Compression. 5.1. Mitigating Post-Processing based Spurious Correlations 5.1.1. COMPRESSION-BASED ARTIFACTS Both C…
Figure 5
Figure 5. Figure 5: Improved Robustness to Downsizing. Compared to the original Corvi, our Corvi⊕ shows increased robustness towards downsampling [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Example of an image which has been recursively in￾painted. The first/second row shows the inpainted image and the inpainted region, respectively. 5.4. Improved Detection of Partially Inpainted Images Prior experiments consider fake images that are completely generated.…
Figure 7
Figure 7. Figure 7: Vulnerability to Spurious Fake Features. Our method Corvi⊕ is not able to mitigate spurious correlations pertaining to the fake distribution, where just like the original Corvi, it continues to associate upsampled images with the fake distribution. real-distribution fe…
Figure 8
Figure 8. Figure 8: Example of different kinds of real images that we consider. This dataset ensures a wide range of testing scenarios, from standard real-world distributions to highly diverse artistic and CG (but not neural network generated) domains. An example of the images used can be…
Figure 9
Figure 9. Figure 9: Distribution of different kinds of real images. The violin plots of Corvi⊕ (left) and Rajan⊕ (right) show that the test set used in our experiments accounts for a wide variety of real image types. use in the main paper is representative of various different types of re…
Figure 10
Figure 10. Figure 10: Robustness to Common Post-Processing Artifacts. We observe similar trends across various corruptions (JPEG compression, additive noise, and low-pass filtering) between the original detectors and our improved versions. Note that these perturbations were part of the tra…
Figure 11
Figure 11. Figure 11: Robustness to Resizing (SynthBuster) We can observe that our detector, Corvi⊕ shows improved robustness to downsizing compared to the original Corvi [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]
Figure 12
Figure 12. Figure 12: Spurious Fake Features. In this neural network, the circles demonstrate neurons, some neurons are bigger in size for demonstration purpose. For such a case, the detector can associate the absence of a spurious real image artifact, such as WEBP compression with fake im…
Figure 13
Figure 13. Figure 13: Improved Robustness to WEBP compression (GAN case). Compared to the original GAN-Baseline, our model displays an improved robustness to WEBP compression. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_13.png]
Figure 14
Figure 14. Figure 14: Distribution of different kinds of real images. We observe that most types of real images are assigned a similar value of fakeness. An exception is the GTA-based images which has a relatively higher score indicating the presence of spurious fake features. 21 [PITH_FU…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Few-Shot Synthetic Image Attribution: Identifying Unseen Generators with Limited Samples

    cs.CV 2025-09 conditional novelty 5.0 of 10

    OmniDFA performs few-shot, open-set attribution of AI-generated images, identifying the source generator from just five support samples across 45 known and unseen generators.

Reference graph

Works this paper leans on

70 extracted references · 48 canonical work pages · cited by 1 Pith paper

  1. [1]

    https://www.midjourney.com/

    Midjourney. https://www.midjourney.com/

    work page

  2. [2]

    Which face is real? https://www.whichfaceisreal.com/

    work page

  3. [3]

    https://www.wikiart.org/

    Wikiart. https://www.wikiart.org/

    work page

  4. [4]

    Large scale GAN training for high fidelity natural image synthesis

    Brock, A., Donahue, J., and Simonyan, K. Large scale GAN training for high fidelity natural image synthesis. In ICLR, 2019

    work page 2019

  5. [5]

    Brooks, T., Holynski, A., and Efros, A. A. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023

    work page 2023

  6. [6]

    Lossless Image Compression through Super-Resolution

    Cao, S., Wu, C.-Y., and Kr \"a henb \"u hl, P. Lossless image compression through super-resolution. arXiv preprint arXiv:2004.02872, 2020

    work page Pith review arXiv 2004

  7. [7]

    What makes fake images detectable? understanding properties that generalize

    Chai, L., Bau, D., Lim, S.-N., and Isola, P. What makes fake images detectable? understanding properties that generalize. In ECCV, 2020

    work page 2020

  8. [8]

    Drct: Diffusion reconstruction contrastive training towards universal detection of diffusion generated images

    Chen, B., Zeng, J., Yang, J., and Yang, R. Drct: Diffusion reconstruction contrastive training towards universal detection of diffusion generated images. In ICML, 2024 a

    work page 2024

Show all 70 references
  1. [9]

    Pixart-\ alpha\ : Fast training of diffusion transformer for photorealistic text-to-image synthesis

    Chen, J., YU, J., GE, C., Yao, L., Xie, E., Wang, Z., Kwok, J., Luo, P., Lu, H., and Li, Z. Pixart-\ alpha\ : Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR, 2024 b

    2024

  2. [10]

    and Koltun, V

    Chen, Q. and Koltun, V. Photographic image synthesis with cascaded refinement networks. In ICCV, 2017

    2017

  3. [11]

    Stargan: Unified generative adversarial networks for multi-domain image-to-image translation

    Choi, Y., Choi, M., Kim, M., Ha, J.-W., Kim, S., and Choo, J. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In CVPR, 2018

    2018

  4. [12]

    How stable is stable diffusion under recursive inpainting (rip)? arXiv preprint arXiv:2407.09549, 2024

    Conde, J., Gonz \'a lez, M., Mart \' nez, G., Moral, F., Merino-G \'o mez, E., and Reviriego, P. How stable is stable diffusion under recursive inpainting (rip)? arXiv preprint arXiv:2407.09549, 2024

    2024 arXiv

  5. [13]

    On the detection of synthetic images generated by diffusion models

    Corvi, R., Cozzolino, D., Zingarini, G., Poggi, G., Nagano, K., and Verdoliva, L. On the detection of synthetic images generated by diffusion models. In ICASSP, 2023

    2023

  6. [14]

    Raising the Bar of AI-generated Image Detection with CLIP

    Cozzolino, D., Poggi, G., Corvi, R., Nießner, M., and Verdoliva, L. Raising the Bar of AI-generated Image Detection with CLIP . In CVPRW, 2024

    2024

  7. [15]

    Zero-shot detection of ai-generated images

    Cozzolino, D., Poggi, G., Nie ner, M., and Verdoliva, L. Zero-shot detection of ai-generated images. In ECCV, 2025

    2025

  8. [16]

    Second-order attention network for single image super-resolution

    Dai, T., Cai, J., Zhang, Y., Xia, S.-T., and Zhang, L. Second-order attention network for single image super-resolution. In CVPR, 2019

    2019

  9. [17]

    S., Wang, J., Wang, R., Zhang, P., Vandenhende, S., Wang, X., Dubey, A., Yu, M., Kadian, A., Radenovic, F., Mahajan, D., Li, K., Zhao, Y., Petrovic, V., Singh, M

    Dai, X., Hou, J., Ma, C.-Y., Tsai, S. S., Wang, J., Wang, R., Zhang, P., Vandenhende, S., Wang, X., Dubey, A., Yu, M., Kadian, A., Radenovic, F., Mahajan, D., Li, K., Zhao, Y., Petrovic, V., Singh, M. K., Motwani, S., Wen, Y., Song, Y., Sumbaly, R., Ramanathan, V., He, Z., Vaj...

    2023

  10. [18]

    RedCaps: Web-curated image-text data created by the people, for the people

    Desai, K., Kaul, G., Aysola, Z., and Johnson, J. RedCaps: Web-curated image-text data created by the people, for the people . In NeurIPS Datasets and Benchmarks, 2021

    2021

  11. [19]

    and Nichol, A

    Dhariwal, P. and Nichol, A. Q. Diffusion models beat GAN s on image synthesis. In NeurIPS, 2021

    2021

  12. [20]

    Scaling rectified flow transformers for high-resolution image synthesis

    Esser, P., Kulal, S., Blattmann, A., Entezari, R., M \"u ller, J., Saini, H., Levi, Y., Lorenz, D., Sauer, A., Boesel, F., et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024

    2024

  13. [21]

    Dreamsim: Learning new dimensions of human visual similarity using synthetic data

    Fu, S., Tamir, N., Sundaram, S., Chai, L., Zhang, R., Dekel, T., and Isola, P. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. arXiv preprint arXiv:2306.09344, 2023

    2023 arXiv

  14. [22]

    J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y

    Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y. Generative adversarial nets. In NeurIPS, 2014

    2014

  15. [23]

    Are gan generated images easy to detect? a critical analysis of the state-of-the-art

    Gragnaniello, D., Cozzolino, D., Marra, F., Poggi, G., and Verdoliva, L. Are gan generated images easy to detect? a critical analysis of the state-of-the-art. In ICME, 2021

    2021

  16. [24]

    Fake or jpeg? revealing common biases in generated image detection datasets

    Grommelt, P., Weiss, L., Pfreundt, F.-J., and Keuper, J. Fake or jpeg? revealing common biases in generated image detection datasets. arXiv preprint arXiv:2403.17608, 2024

    2024 arXiv

  17. [25]

    Vector quantized diffusion model for text-to-image synthesis

    Gu, S., Chen, D., Bao, J., Wen, F., Zhang, B., Chen, D., Yuan, L., and Guo, B. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022

    2022

  18. [26]

    Deep residual learning for image recognition

    He, K., Zhang, X., Ren, S., and Sun, J. Deep residual learning for image recognition. In CVPR, 2016

    2016

  19. [27]

    Progressive growing of GAN s for improved quality, stability, and variation

    Karras, T., Aila, T., Laine, S., and Lehtinen, J. Progressive growing of GAN s for improved quality, stability, and variation. In ICLR, 2018

    2018

  20. [28]

    A style-based generator architecture for generative adversarial networks

    Karras, T., Laine, S., and Aila, T. A style-based generator architecture for generative adversarial networks. In CVPR, 2019

    2019

  21. [29]

    Analyzing and improving the image quality of stylegan

    Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., and Aila, T. Analyzing and improving the image quality of stylegan. In CVPR, 2020

    2020

  22. [30]

    Kirichenko, P., Izmailov, P., and Wilson, A. G. Last layer re-training is sufficient for robustness to spurious correlations. In ICLR, 2023

    2023

  23. [31]

    Labs, B. F. Flux1. https://blackforestlabs.ai/

  24. [32]

    Playground v2

    Li, D., Kamko, A., Akhgari, E., Sabet, A., Xu, L., and Doshi, S. Playground v2. 5: Three insights towards enhancing aesthetic quality in text-to-image generation. arXiv preprint arXiv:2402.17245, 2024

    2024 arXiv

  25. [33]

    Diverse image synthesis from semantic layouts via conditional imle

    Li, K., Zhang, T., and Malik, J. Diverse image synthesis from semantic layouts via conditional imle. In ICCV, 2019

    2019

  26. [34]

    L., and Dollár, P

    Lin, T.-Y., Maire, M., Belongie, S., Bourdev, L., Girshick, R., Hays, J., Perona, P., Ramanan, D., Zitnick, C. L., and Dollár, P. Microsoft coco: Common objects in context, 2015

    2015

  27. [35]

    T., Ben-Hamu, H., Nickel, M., and Le, M

    Lipman, Y., Chen, R. T., Ben-Hamu, H., Nickel, M., and Le, M. Flow matching for generative modeling. In ICLR, 2022

    2022

  28. [36]

    Z., Haghgoo, B., Chen, A

    Liu, E. Z., Haghgoo, B., Chen, A. S., Raghunathan, A., Koh, P. W., Sagawa, S., Liang, P., and Finn, C. Just train twice: Improving group robustness without training group information. In ICML, 2021

    2021

  29. [37]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Liu, X., Gong, C., and qiang liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In ICLR, 2023

    2023

  30. [38]

    A convnet for the 2020s

    Liu, Z., Mao, H., Wu, C.-Y., Feichtenhofer, C., Darrell, T., and Xie, S. A convnet for the 2020s. In CVPR, 2022

    2022

  31. [39]

    Latent consistency models: Synthesizing high-resolution images with few-step inference

    Luo, S., Tan, Y., Huang, L., Li, J., and Zhao, H. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023

    2023 arXiv

  32. [40]

    MindSpore. wukong. https://xihe.mindspore.cn/modelzoo/wukong.2022

    2022

  33. [41]

    Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., Mcgrew, B., Sutskever, I., and Chen, M

    Nichol, A. Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., Mcgrew, B., Sutskever, I., and Chen, M. GLIDE : Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022

    2022

  34. [42]

    Deconvolution and checkerboard artifacts

    Odena, A., Dumoulin, V., and Olah, C. Deconvolution and checkerboard artifacts. Distill, 1 0 (10): 0 e3, 2016

    2016

  35. [43]

    Ojha, U., Li, Y., and Lee, Y. J. Towards universal fake image detectors that generalize across generative models. In CVPR, 2023

    2023

  36. [44]

    Semantic image synthesis with spatially-adaptive normalization

    Park, T., Liu, M.-Y., Wang, T.-C., and Zhu, J.-Y. Semantic image synthesis with spatially-adaptive normalization. In CVPR, 2019

    2019

  37. [45]

    Identifying synthetic faces through gan inversion and biometric traits analysis

    Pasquini, C., Laiti, F., Lobba, D., Ambrosi, G., Boato, G., and De Natale, F. Identifying synthetic faces through gan inversion and biometric traits analysis. Applied Sciences, 2023

    2023

  38. [46]

    amused: An open muse reproduction

    Patil, S., Berman, W., Rombach, R., and von Platen, P. amused: An open muse reproduction. arXiv preprint arXiv:2401.01808, 2024

    2024 arXiv

  39. [47]

    Wuerstchen: Efficient pretraining of text-to-image models

    Pernias, P., Rampas, D., and Aubreville, M. Wuerstchen: Efficient pretraining of text-to-image models. 2023

    2023

  40. [48]

    SDXL : Improving latent diffusion models for high-resolution image synthesis

    Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., M \"u ller, J., Penna, J., and Rombach, R. SDXL : Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024

    2024

  41. [49]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. Learning transferable visual models from natural language supervision. In ICML, 2021

    2021

  42. [50]

    S., Ojha, U., Schloesser, J., and Lee, Y

    Rajan, A. S., Ojha, U., Schloesser, J., and Lee, Y. J. Aligned datasets improve detection of latent diffusion-generated images. In ICLR, 2025

    2025

  43. [51]

    Zero-shot text-to-image generation

    Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., and Sutskever, I. Zero-shot text-to-image generation. In ICML, 2021

    2021

  44. [52]

    Kandinsky: an improved text-to-image synthesis with image prior and latent diffusion

    Razzhigaev, A., Shakhmatov, A., Maltseva, A., Arkhipkin, V., Pavlov, I., Ryabov, I., Kuts, A., Panchenko, A., Kuznetsov, A., and Dimitrov, D. Kandinsky: an improved text-to-image synthesis with image prior and latent diffusion. arXiv preprint arXiv:2310.03502, 2023

    arXiv

  45. [53]

    Aeroblade: Training-free detection of latent diffusion images using autoencoder reconstruction error

    Ricker, J., Lukovnikov, D., and Fischer, A. Aeroblade: Training-free detection of latent diffusion images using autoencoder reconstruction error. In CVPR, 2024

    2024

  46. [54]

    High-resolution image synthesis with latent diffusion models

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-resolution image synthesis with latent diffusion models. In CVPR, 2022

    2022

  47. [55]

    Faceforensics++: Learning to detect manipulated facial images

    Rossler, A., Cozzolino, D., Verdoliva, L., Riess, C., Thies, J., and Nie ner, M. Faceforensics++: Learning to detect manipulated facial images. In ICCV, 2019

    2019

  48. [56]

    W., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S

    Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C. W., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S. R., Crowson, K., Schmidt, L., Kaczmarczyk, R., and Jitsev, J. LAION -5b: An open large-scale dataset for training n...

    2022

  49. [57]

    Zerofake: Zero-shot detection of fake images generated and edited by text-to-image generation models

    Sha, Z., Tan, Y., Li, M., Backes, M., and Zhang, Y. Zerofake: Zero-shot detection of fake images generated and edited by text-to-image generation models. In ACM SIGSAC Conference on Computer and Communications Security, 2024

    2024

  50. [58]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., and Ganguli, S. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015

    2015

  51. [59]

    Denoising diffusion implicit models

    Song, J., Meng, C., and Ermon, S. Denoising diffusion implicit models. In ICLR, 2021

    2021

  52. [60]

    and Ermon, S

    Song, Y. and Ermon, S. Generative modeling by estimating gradients of the data distribution. In NeurIPS, 2019

    2019

  53. [61]

    Score-based generative modeling in latent space

    Vahdat, A., Kreis, K., and Kautz, J. Score-based generative modeling in latent space. NeurIPS, 2021

    2021

  54. [62]

    Wang, S.-Y., Wang, O., Zhang, R., Owens, A., and Efros, A. A. Cnn-generated images are surprisingly easy to spot... for now. In CVPR, 2020

    2020

  55. [63]

    Dire for diffusion-generated image detection

    Wang, Z., Bao, J., Zhou, W., Wang, W., Hu, H., Chen, H., and Li, H. Dire for diffusion-generated image detection. In ICCV, 2023

    2023

  56. [64]

    Gan inversion: A survey

    Xia, W., Zhang, Y., Yang, Y., Xue, J.-H., Zhou, B., and Yang, M.-H. Gan inversion: A survey. IEEE transactions on pattern analysis and machine intelligence, 2022

    2022

  57. [65]

    Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop, 2016

    Yu, F., Seff, A., Zhang, Y., Song, S., Funkhouser, T., and Xiao, J. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop, 2016

    2016

  58. [66]

    Detecting and simulating artifacts in gan fake images

    Zhang, X., Karaman, S., and Chang, S.-F. Detecting and simulating artifacts in gan fake images. In WIFS, 2019

    2019

  59. [67]

    Zhu, J.-Y., Park, T., Isola, P., and Efros, A. A. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, 2017

    2017

  60. [68]

    Gendet: Towards good generalizations for ai-generated image detection

    Zhu, M., Chen, H., Huang, M., Li, W., Hu, H., Hu, J., and Wang, Y. Gendet: Towards good generalizations for ai-generated image detection. arXiv preprint arXiv:2312.08880, 2023

    2023 arXiv

  61. [69]

    Genimage: A million-scale benchmark for detecting ai-generated image

    Zhu, M., Chen, H., Yan, Q., Huang, X., Lin, G., Li, W., Tu, Z., Hu, H., Hu, J., and Wang, Y. Genimage: A million-scale benchmark for detecting ai-generated image. In NeurIPS, 2024

    2024

  62. [70]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.