Pith. sign in

REVIEW 4 major objections 5 minor 47 references

Proto-LeakNet: Towards Signal-Leak Aware Attribution in Synthetic Human Face Imagery

T0 review · 4 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

Pith's one-line read Proto-LeakNet claims that diffusion models' signal-leak biases in latents form a stable, generator-specific fingerprint that supports closed-set attribution and, via density estimation, separation of unseen generators without retraining.

desk verdict Closed-set attribution is competently done; the open-set generalization claim is unsupported by the paper's own numbers. read the letter →

arxiv 2511.04260 v3 pith:KS2LNLUF submitted 2025-11-06 cs.CV cs.AI

classification cs.CVcs.AI MSC 68T4568T10
keywords deepfakeattributionsignal-leakbiasdiffusionmodelslatentspaceforensicsopen-setrecognitionprototypelearningtemporalattentionkerneldensityestimation
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

Proto-LeakNet aims to show that diffusion image generators leave persistent statistical traces — signal-leaks — in their latent representations, and that these traces can serve as generator fingerprints. The method encodes an image with a Stable Diffusion 2.1 VAE, re-adds a small amount of diffusion noise at three timesteps, and aggregates the resulting latents with a temporal-attention ResNet-18 before classifying via learnable prototypes. Trained only on ten closed-set generators, it reports a Macro AUC of 98.13% and the best scores under progressive post-processing. The paper further claims that a Gaussian density estimate on the learned embeddings separates known from unseen generators without retraining, which would make attribution practical for new models. If true, the framework would give forensic analysts both high accuracy and a built-in explanation of which latent features and diffusion steps drove each decision.

What carries the argument

Signal-leak bias: the residual low-frequency statistical mismatch left in diffusion latents by the generator. The pipeline built on it: (1) partial forward-diffusion re-simulation at timesteps {0,5,10} on the SD2.1 VAE latent; (2) a shared-weights ResNet-18 encoding each zt; (3) temporal attention pooling to form h̄; (4) a prototype-based attribution head with per-feature gating w and LogSumExp scoring; and (5) a Gaussian kernel density estimator (Eq. 15) on the final embeddings for open-set evaluation. The prototypes and attention structure the latent geometry; the KDE is the mechanism on which the open-set generalization claim rests.

What would settle it

Run the paper's open-set protocol on a single new generator that is a fine-tune of Stable Diffusion 3.5 (same family as a closed-set class) and compute the overlap coefficient; an OVL near 1.0 would falsify the claim that unseen generators occupy low-density regions. A second check: sweep the unreported KDE bandwidth σ over two orders of magnitude and test whether the AUROC stays above chance.

Watch

Extended reading notes

Core claim

The paper's central claim is that signal-leak bias — residual low-frequency statistical traces that survive the diffusion noising process — is a stable, generator-specific forensic cue. Proto-LeakNet makes this cue measurable by encoding an image through the SD2.1 VAE, re-noising at timesteps {0,5,10}, encoding each latent with a ResNet-18, and pooling with temporal attention; a prototype head then shapes the embedding via attention-weighted distances. On the closed-set WILD benchmark the method reports a Macro AUC of 98.13% and remains the top scorer under three levels of post-processing. For generalization, the paper fits a Gaussian KDE on the frozen closed-set embeddings and reports AUROC

Load-bearing premise

The load-bearing premise is that a Gaussian KDE fit on the 512-dimensional closed-set embeddings assigns reliably lower density to unseen generators than to known ones; the paper's own open-set metrics (OVL=0.90, AUROC≈57%) indicate this premise is not met.

Editorial extensions

If this is right

  • If signal-leak biases are stable under post-processing, attribution systems can keep identifying the source generator after compression, cropping, blur, or resizing without retraining on those degradations.
  • If the closed-set latent geometry holds, the prototype head can be replaced by a lightweight nearest-prototype or Mahalanobis rule, enabling deployment on edge devices.
  • If the density-based open-set evaluation is reliable, a model trained on today's generators could flag images from future generators as out-of-distribution — exactly what open-world deepfake forensics needs.
  • Because swapping the SD2.1 VAE for SDXL latents yields nearly identical closed-set results, the paper concludes signal-leak behavior is intrinsic to the diffusion process, suggesting the approach generalizes across diffusion pipelines.
  • The three-level interpretability (temporal attention weights, feature gates, prototype responsibilities) gives analysts a traceable reason for each attribution decision.
  • An ablation shows that removing prototypes or attention degrades Macro AUC from 98.13% to as low as 93.09%, indicating both components are load-bearing for the claimed latent geometry.

Reading between the lines

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

  • The forensic fingerprint being learned may actually belong to the SD2.1 VAE encoder rather than to the sampled generator; if so, generators that share the same VAE (e.g., many fine-tuned Stable Diffusion models) would be indistinguishable, and switching VAE families would reset the fingerprints — a limitation the paper itself leaves open.
  • The open-set numbers in Table 3 (OVL=0.90, AUROC≈57%) are far closer to chance than to the abstract's 'strong separability', so a practical 'no retraining' story would need a lower-dimensional or nonparametric density model before it is credible.
  • The asymmetric attention trick (applying temporal attention only to closed-set embeddings) amounts to a domain-specific preprocessing choice; a testable follow-up is whether the same asymmetry yields separation for truly held-out generator families rather than the fixed ten open models in WILD.
  • The method could be reoriented from attributing generators to attributing latent pipelines (encoder/decoder pairs), turning it into a pipeline-forensics tool that answers 'which VAE did this image pass through?'.
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

4 major / 5 minor

Summary. Proto-LeakNet proposes an attribution framework for synthetic face images. It encodes each image with the Stable Diffusion 2.1 VAE, re-applies forward diffusion at timesteps {0,5,10}, maps the resulting latents through a ResNet-18, aggregates the embeddings with learned temporal attention, and trains a feature-weighted prototype head on ten closed-set generators from WILD. The closed-set experiments report Macro AUC 98.13%, the best among the compared methods, and the model remains competitive under three levels of post-processing. The paper's second and more novel contribution is an open-set/representation-level generalization analysis: after training, a Gaussian KDE is fitted on frozen closed-set embeddings and used to score known-versus-unseen samples. The abstract claims 'strong separability between real images and known generators, and between known and unseen ones.'

Significance. The closed-set part is a solid engineering contribution: the architecture is clearly described, the ablations isolate the contributions of prototypes and attention, and the per-class tables are more informative than a single average. If the open-set claim were quantitatively supported, the work would be significant for forensic attribution without retraining. In the submitted form, however, this central claim is not established: the only quantified open-set configurations are near chance (AUROC 56.62–57.24%, OVL 0.89–0.90), and the configuration said to separate the distributions is presented only visually. The paper's headline therefore rests on unreported measurements and an unreported KDE bandwidth.

major comments (4)
  1. [§5.3, Table 3, Abstract] The abstract's claim of 'strong separability ... between known and unseen ones' is contradicted by the only quantitative open-set results in Table 3. With attention off/on for both domains, AUROC is 57.24%/56.62%, EER 0.44/0.45, and OVL 0.89/0.90. An AUROC close to 50% and an OVL close to 1 mean the closed/open score distributions are essentially indistinguishable. The central open-set contribution is therefore unsupported by the evidence actually reported.
  2. [§5.3, Fig. 4(c), Table 3] The asymmetric configuration—attention active only on closed embeddings—is the only setup claimed to produce 'distinct, non-overlapping densities,' yet no AUROC, EER, or OVL is reported for it. Figure 4(c) is a qualitative visualization, not a metric; it has no axes or plotted density quantities, and such plots can create apparent separation even in high-dimensional data. A claim of representation-level generalization requires the same quantitative metrics as Table 3 for this configuration, with error bars across seeds.
  3. [§3.4, Eq. (15)] The Gaussian KDE in Eq. (15) depends on a single isotropic bandwidth σ in the high-dimensional embedding space, but σ is never reported, and no bandwidth-selection procedure or sensitivity analysis is given. In high dimensions the log-likelihood s(h) is extremely sensitive to σ; choosing σ too small or too large can make the closed manifold appear arbitrarily narrow or broad. Without this value, the open-set scores cannot be reproduced, and the visual separation in Fig. 4(c) cannot be distinguished from a bandwidth artifact.
  4. [§5.3, attention protocol] The asymmetric protocol is not a deployable open-set procedure. The method requires knowing in advance whether a test sample belongs to the closed or open domain: attention is applied to closed embeddings and disabled for open embeddings. But deciding whether a sample is known or unseen is exactly the task the open-set evaluation is supposed to solve. Applying different feature extractors to the two populations makes the observed separation an artifact of the evaluation protocol rather than of the learned geometry. All queries should be processed by one fixed function.
minor comments (5)
  1. [Abstract vs. §1] The abstract gives a public code link, while the full-text abstract and §1 state that the codebase 'will be available after acceptance.' Please make these statements consistent.
  2. [Fig. 4] The panels have no axis labels, and it is unclear whether they show KDE curves, 2-D projections, or learned embedding scatter. Add axis labels and specify exactly what is plotted.
  3. [§4.1, Eq. (17)] Macro AUC is defined for multiclass scoring, but the open-set experiment is a binary known-vs-unseen task. Clarify how the open-set AUROC is computed and why Macro AUC is needed in that setting.
  4. [§3.3, Table 6] The choice M=4 prototypes is described as empirical. Report the search range or selection criterion used for M.
  5. [Tables 1–3] No error bars or number of random seeds are reported for any metric. Please add variance estimates, especially for the open-set metrics in Table 3.

Circularity Check

1 steps flagged · score 5.0 of 10

Open-set 'prediction' is partly in-sample: the KDE (Eq. 15) is fitted on the same closed embeddings it later scores, and the abstract's strong-separability claim is quantitatively contradicted by Table 3.

  1. fitted input called prediction [Section 3.4 (Eqs. 15-16); used in Section 5.3 and Table 3]
    "After training, we discard the prototype-based classifier and use only the frozen ResNet18 backbone-based encoder to produce pooled embeddings ... forming the sets Hc ... A Gaussian kernel density estimator (KDE) is fitted on Hc to model the manifold of closed embeddings ... For each sample h, we compute its log-likelihood score s(h)=log pKDE(h)."

    As written, Hc is the set of closed embeddings with no train/val/test restriction, even though Section 4 gives the closed set a 5k/2k/3k split. Eq. 15 is fitted on Hc, and Eq. 16 then scores the same Hc; Eq. 19 and Table 3 compare Pclosed(s) with Popen(s), so the closed side of the separation is the KDE's own training distribution. A density estimate is by construction high near its training points, so the closed log-likelihoods are in-sample and the AUROC/EER/OVL do not measure held-out generalization. The paper never states that Hc is restricted to the training split, making the open-set prediction partly forced by the fitting procedure rather than by a genuinely held-out geometry.

full rationale

The closed-set attribution chain is not circular: the Mahalanobis scoring of Eq. 12 is fitted on training embeddings and evaluated on the official held-out closed test split, and the prototype/temporal-attention training in Eqs. 1-11 is an ordinary supervised objective. The WILD benchmark [25] includes two of the present authors but is an external multi-author dataset, not a self-citation invoked to forbid alternatives. The main circularity concern is the open-set evaluation: Section 3.4 defines Hc as 'closed samples' without excluding the test split, fits Eq. 15 on Hc, then computes s(h) for the same Hc before forming OVL/AUROC. In addition, Table 3 reports only near-chance values (AUROC 57.24/56.62, OVL 0.89/0.90) for the two symmetric attention configurations, while the abstract's 'strong separability' is attributed to an asymmetric configuration with no quantitative metrics and an unreported bandwidth sigma. These are partly fitting/reporting issues rather than equation-level equivalence, so the score is moderate: one in-sample fitted-input step, not a fully circular derivation.

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

The ledger shows the main hand-tuned choices (M, T, KDE σ, attention config) that the open-set claim depends on; none are given with independent justification, and one (attention config) is selected after observing results.

free parameters (4)
  • Number of prototypes per class M = 4
    Empirically chosen; Table 6 shows M=4 best, but this tunes the model to the benchmark.
  • Diffusion timestep set T = {0,5,10}
    Hand-selected as low-noise; no sensitivity analysis for these exact steps.
  • KDE bandwidth σ = not reported
    Controls open-set log-likelihoods in Eq. 15; without a value or selection rule, the open-set results are not reproducible.
  • Attention configuration for open-set evaluation = asymmetric (closed only)
    Chosen post hoc after observing separation in Fig. 4; quantitative metrics for this config are not reported.
assumptions (4)
  • domain assumption Stable Diffusion 2.1 latent space contains persistent generator-specific signal-leak traces.
    Invoked from Everaert et al. [12]; the paper adds no independent verification of this premise beyond citing prior work.
  • standard math Cosine noise schedule with alpha_t^2 + sigma_t^2 = 1 is used for forward diffusion.
    Standard for SD2.1; not derived here.
  • domain assumption ResNet18 embeddings in ~512 dimensions are suitable for Gaussian KDE density estimation.
    Curse of dimensionality; no bandwidth or dimension reduction given.
  • domain assumption WILD benchmark splits are prompt-disjoint and representative of real-world generator distributions.
    Taken from [25]; the paper relies on this for both closed and open evaluation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Proto-LeakNet: Towards Signal-Leak Aware Attribution in Synthetic Human Face Imagery." pith.science (2026). https://pith.science/paper/KS2LNLUF

@misc{pith2026251104260,
  author       = {Pith},
  title        = {Pith review of: Proto-LeakNet: Towards Signal-Leak Aware Attribution in Synthetic Human Face Imagery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KS2LNLUF}},
  note         = {Machine review of arXiv:2511.04260}
}
read the original abstract

The growing sophistication of synthetic image and deepfake generation models has turned source attribution and authenticity verification into a critical challenge for modern computer vision systems. Recent studies suggest that diffusion pipelines unintentionally imprint persistent statistical traces, known as signal-leaks, within their outputs, particularly in latent representations. Building on this observation, we propose Proto-LeakNet, a signal-leak-aware and interpretable attribution framework that integrates Closed-set classification with a density-based Open-set evaluation on the learned embeddings, enabling analysis of unseen generators without retraining. Acting in the latent domain of diffusion models, our method re-simulates partial forward diffusion to expose residual generator-specific cues. A temporal attention encoder aggregates multi-step latent features, while a feature-weighted prototype head structures the embedding space and enables transparent attribution. Trained solely on closed data and achieving a Macro AUC of 98.13\%, Proto-LeakNet learns a latent geometry that remains robust under post-processing, surpassing state-of-the-art methods, and achieves strong separability both between real images and known generators, and between known and unseen ones. The codebase is available at the following link: https://github.com/claudiunderthehood/Proto-LeakNet .

Figures

Figures reproduced from arXiv: 2511.04260 by the authors.

Figure 1
Figure 1. Proto-LeakNet: given an input image x, latent features are extracted from the pretrained Stable Diffusion 2.1 Variational Autoencoder (VAE) in Block 1 (Latent Feature Construction). For each diffusion step t∈ {0, 5, 10}, we extract a zt latent which gets normalized with dimension (4, 32, 32). In Block 2 (Discriminative Backbone), each zt is encoded by a ResNet-18, producing embeddings {ht} that are temporally aggreg… view at source ↗
Figure 2
Figure 2. Interpretability of Proto-LeakNet (t-SNE Evolution). Layer-wise evolution of the embedding space, illustrating [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Histogram of the Top-1 accuracy distributions per class for each method, from raw to step 1-3. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Impact of attention configurations on latent-space separation. (a) Disabling attention for both domains leads [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 5 linked inside Pith

  1. [1]

    Generative adversarial nets.Advances in Neural Information Processing Systems, 27, 2014

    Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets.Advances in Neural Information Processing Systems, 27, 2014

  2. [2]

    Denoising diffusion probabilistic models.Advances in Neural Information Processing Systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in Neural Information Processing Systems, 33:6840–6851, 2020

  3. [3]

    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. 10 Proto-LeakNet: Towards Signal-Leak Aware Attribution in Synthetic Human Face ImageryA PREPRINT

  4. [4]

    Deepfake detection in the aigc era: A survey, benchmarks, and future perspectives.Information Fusion, page 103740, 2025

    Shichuang Xie, Tong Qiao, Sheng Li, Xinpeng Zhang, Jiantao Zhou, and Guorui Feng. Deepfake detection in the aigc era: A survey, benchmarks, and future perspectives.Information Fusion, page 103740, 2025

  5. [5]

    Media forensics and deepfakes: An overview.IEEE journal of selected topics in signal processing, 14(5):910–932, 2020

    Luisa Verdoliva. Media forensics and deepfakes: An overview.IEEE journal of selected topics in signal processing, 14(5):910–932, 2020

  6. [6]

    Deepfake detection by analyzing convolutional traces

    Luca Guarnera, Oliver Giudice, and Sebastiano Battiato. Deepfake detection by analyzing convolutional traces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern recognition workshops, pages 666–667, 2020

  7. [7]

    FakeSpotter: A Simple yet Robust Baseline for Spotting AI-Synthesized Fake Faces

    Run Wang, Felix Juefei-Xu, Lei Ma, Xiaofei Xie, Yihao Huang, Jian Wang, and Yang Liu. FakeSpotter: A Simple yet Robust Baseline for Spotting AI-Synthesized Fake Faces. InProceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI’20, 2021. ISBN 9780999241165

  8. [8]

    Deepfake Attribution: On the Source Identification of Artificially Generated Images.Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 12(3): e1438, 2022

    Brandon Khoo, Raphaël C-W Phan, and Chern-Hong Lim. Deepfake Attribution: On the Source Identification of Artificially Generated Images.Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 12(3): e1438, 2022

Show all 47 references
  1. [9]

    Tiny Autoencoders are Effective Few-Shot Generative Model Detectors

    Luca Bindini, Giulia Bertazzini, Daniele Baracchi, Dasara Shullani, Paolo Frasconi, and Alessandro Piva. Tiny Autoencoders are Effective Few-Shot Generative Model Detectors. In2024 IEEE International Workshop on Information Forensics and Security (WIFS), pages 1–6. IEEE, 2024

  2. [10]

    On the Detection of Synthetic Images Generated by Diffusion Models

    Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the Detection of Synthetic Images Generated by Diffusion Models. InICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages ...

  3. [11]

    BOSC: A Backdoor-Based Framework for Open Set Synthetic Image Attribution.IEEE Transactions on Information Forensics and Security, 20:8043–8058, 2025

    Jun Wang, Benedetta Tondi, and Mauro Barni. BOSC: A Backdoor-Based Framework for Open Set Synthetic Image Attribution.IEEE Transactions on Information Forensics and Security, 20:8043–8058, 2025

  4. [12]

    Exploiting the Signal-Leak Bias in Diffusion Models

    Martin Nicolas Everaert, Athanasios Fitsios, Marco Bocchio, Sami Arpa, Sabine Süsstrunk, and Radhakrishna Achanta. Exploiting the Signal-Leak Bias in Diffusion Models. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4025–4034, 2024

  5. [13]

    Exploring unbiased deepfake detection via token-level shuffling and mixing

    Xinghe Fu, Zhiyuan Yan, Taiping Yao, Shen Chen, and Xi Li. Exploring unbiased deepfake detection via token-level shuffling and mixing. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 3040–3048, 2025

  6. [14]

    Adaptive test-time semantic debiasing for ai-generated image detection

    Yu Cai, Jiahe Tian, Xiaomeng Fu, Jiao Dai, Jizhong Han, and Siwei Lyu. Adaptive test-time semantic debiasing for ai-generated image detection. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 1543–1552, 2025

  7. [15]

    Artificial Fingerprinting for Generative Models: Rooting Deepfake Attribution in Training Data

    Ning Yu, Vladislav Skripniuk, Sahar Abdelnabi, and Mario Fritz. Artificial Fingerprinting for Generative Models: Rooting Deepfake Attribution in Training Data. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 14448–14457, 2021

  8. [16]

    Contrastive Pseudo Learning for Open-World Deepfake Attribution

    Zhimin Sun, Shen Chen, Taiping Yao, Bangjie Yin, Ran Yi, Shouhong Ding, and Lizhuang Ma. Contrastive Pseudo Learning for Open-World Deepfake Attribution. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 20882–20892, 2023

  9. [17]

    Do DeepFake Attribution Models Generalize? InProceedings of the 4th ACM International Workshop on Multimedia AI against Disinformation, pages 45–54, 2025

    Spiros Baxevanakis, Manos Schinas, and Symeon Papadopoulos. Do DeepFake Attribution Models Generalize? InProceedings of the 4th ACM International Workshop on Multimedia AI against Disinformation, pages 45–54, 2025

  10. [18]

    Frequency-Aware Deepfake Detection: Improving Generalizability through Frequency Space Learning

    Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Frequency-Aware Deepfake Detection: Improving Generalizability through Frequency Space Learning. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 5052–5060, 2024

  11. [19]

    Present and Future Generalization of Synthetic Image Detectors

    Pablo Bernabeu-Pérez, Enrique Lopez-Cuena, and Dario Garcia-Gasulla. Present and Future Generalization of Synthetic Image Detectors. InJoint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 3–20. Springer, 2025

  12. [20]

    How to Trace Latent Generative Model Generated Images without Artificial Watermark? InInternational Conference on Machine Learning, pages 51396–51414

    Zhenting Wang, Vikash Sehwag, Chen Chen, Lingjuan Lyu, Dimitris N Metaxas, and Shiqing Ma. How to Trace Latent Generative Model Generated Images without Artificial Watermark? InInternational Conference on Machine Learning, pages 51396–51414. PMLR, 2024

  13. [21]

    Which Model Generated This Image? A Model-Agnostic Approach for Origin Attribution

    Fengyuan Liu, Haochen Luo, Yiming Li, Philip Torr, and Jindong Gu. Which Model Generated This Image? A Model-Agnostic Approach for Origin Attribution. InEuropean Conference on Computer Vision, pages 282–301. Springer, 2024

  14. [22]

    Rethinking the Up- Sampling Operations in CNN-based Generative Network for Generalizable Deepfake Detection

    Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the Up- Sampling Operations in CNN-based Generative Network for Generalizable Deepfake Detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pag...

  15. [23]

    LATTE: Latent Trajectory Embedding for Diffusion-Generated Image Detection.arXiv preprint arXiv:2507.03054, 2025

    Ana Vasilcoiu, Ivona Najdenkoska, Zeno Geradts, and Marcel Worring. LATTE: Latent Trajectory Embedding for Diffusion-Generated Image Detection.arXiv preprint arXiv:2507.03054, 2025

  16. [24]

    Deep Residual Learning for Image Recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770–778, 2016

  17. [25]

    WILD: a new in-the-Wild Image Linkage Dataset for synthetic image attribution

    Pietro Bongini, Sara Mandelli, Andrea Montibeller, Mirko Casu, Orazio Pontorno, Claudio Vittorio Ragaglia, Luca Zanchetta, Mattia Aquilina, Taiba Majid Wani, Luca Guarnera, Benedetta Tondi, Giulia Boato, Paolo Bestagini, Irene Amerini, Francesco De Natale, Sebastiano Battiato,...

  18. [26]

    Adobe.Adobe Firefly, 2023.https://firefly.adobe.com/

  19. [27]

    Improving image generation with better captions.OpenAI Technical Report, 2024

    OpenAI. Improving image generation with better captions.OpenAI Technical Report, 2024. URL https: //cdn.openai.com/papers/dall-e-3.pdf. Accessed: 2025-03-20

  20. [28]

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

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

  21. [29]

    FLUX 1.1 [pro]: Advanced Text-to-Image Generation Model, 2024

    Black Forest Labs. FLUX 1.1 [pro]: Advanced Text-to-Image Generation Model, 2024. URL https:// blackforestlabs.ai/1-1-pro/. Accessed: 2025-03-20

  22. [30]

    Freepik AI Image Generator, 2024

    Freepik. Freepik AI Image Generator, 2024. URL https://docs.freepik.com/api-reference/mystic/ post-mystic. Accessed: 2025-03-20

  23. [31]

    Leonardo AI: AI-Powered Creative Image Generation Platform, 2024

    Leonardo AI. Leonardo AI: AI-Powered Creative Image Generation Platform, 2024. URL https://leonardo. ai. Accessed: 2025-03-20

  24. [32]

    MidJourney: An AI-powered image generation tool, 2024

    MidJourney. MidJourney: An AI-powered image generation tool, 2024. URL https://www.midjourney.com. Accessed: 2025-03-20

  25. [33]

    Stable diffusion 3.5-large

    Stability AI. Stable diffusion 3.5-large. https://huggingface.co/stabilityai/stable-diffusion-3. 5-large, 2025. Accessed: February 28, 2025

  26. [34]

    Using diffusers: Sdxl turbo, 2023

    Hugging Face. Using diffusers: Sdxl turbo, 2023. URL https://huggingface.co/docs/diffusers/en/ using-diffusers/sdxl_turbo. Accessed: 2025-03-11

  27. [35]

    Starry ai, 2023

    Starry AI. Starry ai, 2023. URLhttps://starryai.com/. Accessed: 2025-03-11

  28. [36]

    Zero-shot text-to-image generation.arXiv preprint arXiv:2102.12092, 2021

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea V oss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation.arXiv preprint arXiv:2102.12092, 2021. URL https://arxiv. org/abs/2102.12092

  29. [37]

    DeepAI Text-to-Image Generator, 2024

    DeepAI. DeepAI Text-to-Image Generator, 2024. URL https://deepai.org/machine-learning-model/ text2img. Accessed: 2025-03-20

  30. [38]

    HotPot AI: AI-Powered Image and Text Generation Tools, 2024

    HotPot AI. HotPot AI: AI-Powered Image and Text Generation Tools, 2024. URL https://hotpot.ai/. Accessed: 2025-03-20

  31. [39]

    SANA: Efficient high-resolution image synthesis with linear diffusion transformers.arXiv preprint arXiv:2410.10629, 2024

    Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, and Song Han. SANA: Efficient high-resolution image synthesis with linear diffusion transformers.arXiv preprint arXiv:2410.10629, 2024. URLhttps://arxiv.org/abs/2...

  32. [40]

    Richter, Christopher Pal, and Marc Aubreville

    Pablo Pernias, Dominic Rampas, Mats L. Richter, Christopher Pal, and Marc Aubreville. Würstchen: An efficient architecture for large-scale text-to-image diffusion models. InInternational Conference on Learning Representations (ICLR), 2024

  33. [41]

    Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models, 2023

    Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models, 2023. URLhttps://arxiv.org/abs/2301.13826

  34. [42]

    A style-based generator architecture for generative adversarial networks, 2019

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks, 2019. URLhttps://arxiv.org/abs/1812.04948

  35. [43]

    Analyzing and improving the image quality of stylegan, 2020

    Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan, 2020. URLhttps://arxiv.org/abs/1912.04958

  36. [44]

    Alias-free generative adversarial networks.Advances in neural information processing systems, 34:852–863, 2021

    Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks.Advances in neural information processing systems, 34:852–863, 2021

  37. [45]

    Hunyuan-dit: A powerful multi-resolution diffusion transformer with fine-grained chinese understanding, 2024

    Zhimin Li, Jianwei Zhang, Qin Lin, and et al. Hunyuan-dit: A powerful multi-resolution diffusion transformer with fine-grained chinese understanding, 2024. 12 Proto-LeakNet: Towards Signal-Leak Aware Attribution in Synthetic Human Face ImageryA PREPRINT

  38. [46]

    Deep Learning Face Attributes in the Wild

    Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep Learning Face Attributes in the Wild. In Proceedings of the IEEE International Conference on Computer Vision, pages 3730–3738, 2015

  39. [47]

    Progressive Growing of GANs for Improved Quality, Stability, and Variation

    Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive Growing of GANs for Improved Quality, Stability, and Variation. InInternational Conference on Learning Representations (ICLR), 2018. 13

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.