Pith. sign in

REVIEW 5 major objections 7 minor 2 cited by

Evaluating Deepfake Detectors in the Wild

T0 review · 5 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Open-source deepfake detectors that score near-perfectly on standard benchmarks largely fail to distinguish real faces from SimSwap and Inswapper swaps on a new 500,000-image evaluation built to mimic real-world conditions.

desk verdict A genuinely useful 500k-image deepfake benchmark, but the robustness experiments apply transformations only to the fake class, which muddies the 'in the wild' conclusion. read the letter →

arxiv 2507.21905 v2 pith:6DARLJDH submitted 2025-07-29 cs.CV cs.AIcs.LG

Viacheslav Pirogov , Maksim Artemev This is my paper
classification cs.CVcs.AIcs.LG
keywords deepfakedetectionfaceswappingzero-shotbenchmarkgeneralizationSimSwapInswapperJPEGcompressionimageenhancement
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

Most published deepfake detectors are tested on benchmark data made with older generators, so their near-perfect scores may not reflect what happens when fraudsters use current tools. This paper builds a large synthetic test set, over 500,000 images, by swapping faces with SimSwap and Inswapper on faces from CelebA-HQ, LFW, and FairFace, using the default cropping and alignment a typical attacker would use. It then runs six open-source detectors and finds that fewer than half reach an AUC above 60%, with the weakest at 50%, which is chance level. Basic manipulations attackers commonly apply, such as JPEG compression, downscaling, and GPEN enhancement, further degrade several detectors. The upshot is that published AUC figures overstate real-world capability for zero-shot face-swap attacks.

What carries the argument

The load-bearing mechanism is the evaluation pipeline rather than a single theorem or model. It combines two zero-shot face-swap generators, SimSwap at 224 or 512 pixels and Inswapper at 128 pixels, with three real-face datasets, CelebA-HQ, LFW, and FairFace, to produce synthetic fake images under the default face cropping and alignment an attacker would use, then feeds them to each detector through its official preprocessing. Three attacker-style augmentations, JPEG compression at five quality levels, downscaling to 128 pixels, and GPEN restoration, perturb those images, and ROC-AUC is used as the primary score because it is threshold-independent and comparable across detectors with different output scales. The pipeline's job is to expose whether a detector is separating real from fake for the right reasons, such as generic forgery artifacts, or relying on dataset- and generator-specific cues that do not survive simple image transformations.

What would settle it

A concrete check would be to rerun the released pipeline with a different, equally standard face alignment or detector preprocessing: if a detector's AUC jumps from near 50% to above 90% under that change, the reported failures are an artifact of the test setup rather than of deepfake detection. Alternatively, collecting real face-swap images from public apps and showing that detectors score above 90% AUC on them would directly contradict the paper's generalization claim.

Watch

Extended reading notes

Core claim

The paper claims that modern open-source deepfake detectors do not generalize to current zero-shot face-swap attacks, despite near-perfect performance on the benchmarks they were trained or tuned on. The authors construct a test set of more than 500,000 images by swapping faces with SimSwap and Inswapper, using default cropping and alignment, across CelebA-HQ, LFW, and FairFace. Six detectors are run through their official code and weights: FaceForensics++, MAT, M2TR, RECCE, CADDM, and SBI. On the unmodified swaps, fewer than half of the tested detectors reach an ROC-AUC above 60%, and the weakest sits at 50%, no better than chance; SBI is the clear leader at 95.5% AUC on SimSwap and 75.9% on Inswapper. The paper further shows that basic manipulations used in real attacks degrade performance unevenly: downscaling to 128 pixels craters SBI into the 30s, JPEG compression moves some models up and others down, and GPEN enhancement pushes several models below 50% AUC, meaning they label deepfakes as real. The authors conclude that detecting deepfakes remains challenging and that published benchmark results mask the gap.

Load-bearing premise

The evaluation assumes that the way the authors generated the test images, using SimSwap and Inswapper with their default face cropping and alignment on three public face datasets, then running each detector's official preprocessing, matches what real-world deepfake attacks look like to a deployed detector.

Editorial extensions

If this is right

  • Published AUC figures on older benchmarks such as DFDC, FaceForensics++, and Celeb-DF should not be treated as deployment-ready performance for current zero-shot face-swap tools.
  • A single, visually undetectable JPEG re-compression at quality around 75, or a downscale to 128 pixels, can move an otherwise strong detector toward chance, which matters for identity verification and moderation pipelines.
  • No tested detector maintains high ROC-AUC across original, compressed, downscaled, and enhanced inputs; the strongest overall model, SBI, collapses specifically under downscaling.
  • Evaluation protocols for deepfake detectors should include attacker-style augmentations and out-of-distribution face datasets instead of relying on clean benchmark splits.
  • Models that learn generic artifacts, such as blending boundaries, color mismatch, and frequency inconsistency, generalize better than models that memorize training-domain cues, as the paper argues SBI demonstrates.

Reading between the lines

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

  • Because the failures appear at JPEG quality levels a social network or messaging app would apply by default, deployed detectors may be substantially weaker than their benchmark cards suggest; this is an inference about deployment, not a claim the paper directly tests.
  • The same protocol could be extended to video and audio deepfakes, or to diffusion-based face swappers, to see whether the generalization gap widens with newer generators.
  • A detector trained on the released dataset would presumably score high on this benchmark, so the result is evidence about zero-shot generalization of existing detectors, not about the impossibility of detection.
  • The different failure signatures of SimSwap, with 512-pixel blending, and Inswapper, with 128-pixel upsampling, suggest detectors may latch onto generator-specific artifacts; varying generator versions and resolutions in the benchmark would test that directly.
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

5 major / 7 minor

Summary. The paper evaluates six open-source deepfake detectors (FaceForensics++, MAT, M2TR, RECCE, CADDM, and SBI) against a new synthetic benchmark built by swapping faces with SimSwap and Inswapper using CelebA-HQ, LFW, and FairFace as source datasets. It reports ROC-AUC, PR-AUC, logloss, F1, and accuracy, and it probes robustness by applying JPEG compression, downscaling, and GPEN enhancement to the deepfake images. The authors conclude that modern detectors are brittle on out-of-distribution face swaps and that common image manipulations substantially reduce performance; they release the code and a dataset of over 500,000 deepfake images.

Significance. If the technical issues are resolved, this is a potentially valuable empirical contribution: it provides a large released benchmark, runs six detectors from official code, and adds evidence that published AUC figures on standard deepfake benchmarks may overstate out-of-domain robustness. The dataset and code release are concrete strengths, and the core qualitative finding that zero-shot face-swap detectors do not all generalize to new generators is plausible and worth reporting. However, the paper's headline statistic is imprecise, no uncertainty quantification is reported, and the robustness experiments use a one-sided manipulation protocol, so the quantitative conclusions are not yet reliably supported.

major comments (5)
  1. [§4.3–§4.4, Table 2] The quality manipulations are applied only to the synthetic (fake) images, while the real comparison images are not compressed, downscaled, or enhanced. The paper describes the protocol as 'subjecting each image in the synthetic datasets to JPEG compression' (§4.3) and 'downscaled our synthetic dataset images' (§4.4), and Table 2 is therefore a comparison between manipulated fakes and untouched real images. The reported AUC changes conflate detector failure with a one-sided distribution shift: the fake class acquires compression, resolution, or enhancement artifacts that the real class lacks, and a detector can change its AUC simply by exploiting that new covariate difference. The paper's own observation that FF and MAT score all compressed images as fakes is consistent with this artifact. Because real-world deployments such as KYC, social-media uploads, and liveness checks also expose legitimate images to compression and resizing, the matched-quality condition is the relevant one. Please report matched-quality controls in which the same transformations are applied to both classes, or explicitly reframe the experiments as an attacker-only evasion model and justify that threat model.
  2. [Abstract and Table 1] The headline sentence 'in fewer than half of the deepfake detectors tested achieved an AUC score greater than 60%, with the lowest being 50%' is ambiguous and numerically inaccurate. Taking the overall rows in Table 1 as the unit, exactly three of the six detectors exceed 60% AUC for SimSwap (MAT, CADDM, SBI), i.e., half, not fewer than half; the statement is only true for Inswapper (two of six) or for detector–generator combinations in aggregate. Moreover, Table 1 contains values well below 50%—for example FF/LFW/Inswapper at 49.2, MAT/CelebA-HQ/SimSwap at 49.0, and RECCE/CelebA-HQ/SimSwap at 46.9—so 'the lowest being 50%' is false. Define the unit of analysis (detector, detector×dataset, or detector×generator) explicitly and correct the statement to match Table 1.
  3. [§4.2, Table 1] No sample sizes or confidence intervals are reported for any cell. The paper draws categorical conclusions from small AUC differences around the 60% boundary—for example CADDM/Inswapper/FairFace at 56.6 and M2TR/Inswapper/LFW at 53.9—but without per-cell N and uncertainty it is impossible to determine whether these differences are meaningful or whether a given value is statistically distinguishable from 50%. Please report the number of real and fake images per cell and provide confidence intervals or standard errors (e.g., via bootstrapping), especially for values near chance.
  4. [§3.2 and §4] The generalization claim rests on two preprocessing assumptions that are asserted but not verified in the manuscript: that the default SimSwap/Inswapper face-cropping and alignment settings match realistic attacks, and that all six detectors' official preprocessing pipelines are correctly reproduced on the new images. The statement that official repositories, weights, and fixed library versions were used is helpful, but the paper does not show the preprocessed inputs to each detector, give the exact commits used, or discuss how representative this zero-shot still-image protocol is of the 'wild' scenarios named in the title and abstract. Please document the preprocessing per detector, show sample preprocessed inputs, and discuss the scope and limitations of the protocol as a model of real-world attacks.
  5. [§4.2, Table 1] The sentence 'from all selected models only MAT, SBI, and CADDM models are showing high enough quality in detecting deepfakes from SimSwap and Inswapper on all three datasets' is not supported by Table 1. MAT scores 49.0/49.8 on CelebA-HQ for SimSwap/Inswapper, and CADDM scores 56.6 on FairFace for Inswapper, which are not high-quality results. Please correct or substantially soften this claim, since it directly informs the paper's takeaway about which detectors are usable.
minor comments (7)
  1. [§3.1] The paper says FairFace's attributes can be used 'to assess how each detector performs on a separate ethnic, age, and gender subset,' but no demographic-stratified results appear in §4. Either add that analysis, which would substantially strengthen the fairness claim, or explicitly state that it is left to future work.
  2. [§2.2 and Figure 2] GPEN is attributed to (Yin et al., 2023) in the text and figure caption, but the reference list identifies Yin et al. (2023) as DAGP-Face Restorer, while the actual GPEN paper is Yang et al. (2021), which is also cited. The attribution should be corrected.
  3. [§5, final paragraph] The text says 'no single model consistently maintains high performance metrics such as ROC-AUC, across all variations' and then immediately says 'One notable exception to this is the SBI model.' This is self-contradictory; decide whether SBI is an exception and rephrase accordingly.
  4. [Abstract and §1] The sentence beginning 'in fewer than half of the deepfake detectors tested achieved' is grammatically incomplete, and in §1 'the second, method is to take two photos' contains an awkward comma. Both should be cleaned up.
  5. [§3.3.1 and Figure 5] The values 95, 75, 50, 30, and 10 are JPEG quality levels, not 'coefficients'; please use consistent terminology.
  6. [§4.4] The phrase 'the ability of the detectors to discriminate between original and downscaled deepfakes' is imprecise: the actual comparison is between manipulated fake images and untouched real images. State the comparison explicitly.
  7. [Throughout] There are repeated typos and inconsistencies, including 'FaceForentics' instead of 'FaceForensics', 'CelebaHQ' instead of 'CelebA-HQ', and 'it's collection' instead of 'its collection'. A careful proofreading pass is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an external benchmark with no fitted parameters and no load-bearing self-citation chain.

full rationale

The paper's central claim is an empirical evaluation: six open-source deepfake detectors were run, using their official code and weights, on a newly generated set of SimSwap and Inswapper face swaps drawn from CelebA-HQ, LFW, and FairFace, and their ROC-AUC scores were reported. No parameter is fitted to the test data, no detector is trained or fine-tuned on the synthetic dataset, and no 'prediction' is derived from a model whose inputs already contain the output. The reported AUC values are computed from held-out real and fake images with standard classification metrics, so the conclusion that 'detecting deepfakes still remains a challenging task' is a direct measurement rather than a tautology. The robustness experiments in Sections 4.3 and 4.4 apply JPEG compression, downscaling, and GPEN enhancement only to the fake images while leaving real images untouched; this is a legitimate concern about whether the attack model matches real-world conditions, and it could bias the reported AUC comparisons, but it is a soundness or protocol-validity issue rather than circular reasoning. The paper does not invoke a uniqueness theorem, does not rely on self-citations by the present authors as load-bearing evidence, and does not rename a known result as a new derivation. The authors' affiliations and the public GitHub repository do not create circularity because the benchmark results are externally checkable against the released code and data. Accordingly, no specific circular step can be quoted and exhibited, and the appropriate finding is no significant circularity.

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

No numbers are fitted to data in this evaluation; the central result is an empirical measurement, not a derived model. The claim rests on three domain assumptions about representativeness of generators, augmentations, and correct deployment of detector code.

assumptions (3)
  • domain assumption The two generators (SimSwap and Inswapper) and the three face datasets (CelebA-HQ, LFW, FairFace) are representative of real-world deepfake attacks and identities.
    The central claim that detectors fail 'in the wild' assumes the synthetic dataset reflects realistic attack conditions; Section 3.2 motivates this but provides no external validation.
  • domain assumption JPEG compression, downscaling to 128px, and GPEN enhancement are the key real-world image manipulations fraudsters apply.
    Section 3.3 selects these augmentations as representative; other common operations (blur, noise, recapture, laundering through social media) are excluded.
  • domain assumption Official detector implementations with their disclosed weights and preprocessing are deployed correctly in the evaluation pipeline.
    Section 4 states environments match the authors' instructions, but no checks are reported that the detectors' outputs are meaningful on these out-of-distribution inputs (e.g., sanity checks on calibration).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Evaluating Deepfake Detectors in the Wild." pith.science (2026). https://pith.science/paper/6DARLJDH

@misc{pith2026250721905,
  author       = {Pith},
  title        = {Pith review of: Evaluating Deepfake Detectors in the Wild},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6DARLJDH}},
  note         = {Machine review of arXiv:2507.21905}
}
read the original abstract

Deepfakes powered by advanced machine learning models present a significant and evolving threat to identity verification and the authenticity of digital media. Although numerous detectors have been developed to address this problem, their effectiveness has yet to be tested when applied to real-world data. In this work we evaluate modern deepfake detectors, introducing a novel testing procedure designed to mimic real-world scenarios for deepfake detection. Using state-of-the-art deepfake generation methods, we create a comprehensive dataset containing more than 500,000 high-quality deepfake images. Our analysis shows that detecting deepfakes still remains a challenging task. The evaluation shows that in fewer than half of the deepfake detectors tested achieved an AUC score greater than 60%, with the lowest being 50%. We demonstrate that basic image manipulations, such as JPEG compression or image enhancement, can significantly reduce model performance. All code and data are publicly available at https://github.com/SumSubstance/Deepfake-Detectors-in-the-Wild.

Figures

Figures reproduced from arXiv: 2507.21905 by the authors.

Figure 1
Figure 1. Samples from our synthetic deepfake dataset. Images in the top row were processed by the Inswapper (Jia Guo, 2017) (roop) model, while images in the bottom row were generated using SimSwap (Chen et al., 2020). Samples from Inswapper have slightly lower quality because the model operates at a lower resolution of 128. GPEN, as presented in the ”GAN Prior Embedded Network for Blind Face Restoration in the Wild” paper (… view at source ↗
Figure 2
Figure 2. Samples from our synthetic dataset enhanced using GPEN (Yin et al., 2023). Originals on the left; enhanced versions on the right. Top: Inswapper outputs; bottom: SimSwap. We demonstrate that enhancers render low-quality generators’ outputs visually comparable to real samples. iohara & Yamasaki, 2022) uses a unique dataset generated through the blending of pseudo-source and target images derived from individual prist… view at source ↗
Figure 3
Figure 3. Many video frames from DFDC and FaceForentics datasets frequently demonstrate lower fidelity and a higher amount of artifacts due to specific keypoint matching and blending procedures as well as the usage of older deepfake methods. bridge the gap in current deepfake detection capabilities by providing a robust, up-to-date testing framework that reflects advances in deepfake generation technologies. This not only imp… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Examples of images from our synthetic dataset downscaled to 128 with the same aspect ratio. Originals on the left; downscaled versions on the right. Notably, Inswapper (top row) samples are not heavily changed, while the deterioration of the SimSwap (bottom row) sample…
Figure 5
Figure 5. Figure 5: Visual examples of image quality deterioration for different JPEG compression coefficients. Coefficients from left to right: [original, 95, 75, 50, 30, 10] its in-dataset stability. To evaluate out-of-domain perfor￾mance, we incorporated the FairFace dataset, which has…
Figure 6
Figure 6. Figure 6: Normalized histograms of predicted values for ”real” images in selected datasets. Here one can see that all models failed to classify ”real” samples with some incorrectly assigning high probability to a major percentage of dataset samples. (a) CADDM (b) FF (c) M2TR (d)…
Figure 7
Figure 7. Figure 7: AUC-ROC by JPEG compression coefficient on all models, separated by datasets. posed pipeline. The main problem is the abundance of deep￾fake detection methods with unpublished code or weights, preventing comprehensive testing of their efficacy. Another issue is the lac…

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Continuously Evolving Deepfake Detection: An Architecture and Public-Benchmark Evaluation of a Dynamic Detection System

    cs.CV 2026-07 conditional novelty 6.0 of 10

    A continuously refreshed, incentive-driven deepfake detector beats static detectors on in-the-wild benchmarks and improves on post-export AI-generated media.

  2. Visual Language Models as Zero-Shot Deepfake Detectors

    cs.CV 2025-07 conditional novelty 4.0 of 10

    Zero-shot VLMs scored by normalized yes/no token probabilities beat most trained deepfake detectors on a new SimSwap dataset, and a lightly fine-tuned InstructBLIP is near-perfect on DFDC-P.

Reference graph

Works this paper leans on

71 extracted references · 32 canonical work pages · cited by 2 Pith papers

  1. [1]

    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 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    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 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019 . OpenReview.net, 2019. URL https://openreview.net/forum?id=B1xsqj09Fm

    work page 2019

  3. [3]

    and Tzimiropoulos, G

    Bulat, A. and Tzimiropoulos, G. Super-fan: Integrated facial landmark localization and super-resolution of real-world low resolution faces in arbitrary poses with gans. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 109--117, 2017. URL https://api.semanticscholar.org/CorpusID:4319595

    work page 2018

  4. [4]

    Roop unleashed

    C0untFloyd. Roop unleashed. https://github.com/C0untFloyd/roop-unleashed, 2023

    work page 2023

  5. [5]

    End-to-end reconstruction-classification learning for face forgery detection

    Cao, J., Ma, C., Yao, T., Chen, S., Ding, S., and Yang, X. End-to-end reconstruction-classification learning for face forgery detection. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 4103--4112, 2022. doi:10.1109/CVPR52688.2022.00408

    arXiv 2022

  6. [6]

    M., and Zisserman, A

    Cao, Q., Shen, L., Xie, W., Parkhi, O. M., and Zisserman, A. Vggface2: A dataset for recognising faces across pose and age. In 13th IEEE International Conference on Automatic Face & Gesture Recognition, FG 2018, Xi'an, China, May 15-19, 2018 , pp.\ 67--74. IEEE Computer Society, 2018. doi:10.1109/FG.2018.00020. URL https://doi.org/10.1109/FG.2018.00020

    arXiv 2018

  7. [7]

    Simswap: An efficient framework for high fidelity face swapping

    Chen, R., Chen, X., Ni, B., and Ge, Y. Simswap: An efficient framework for high fidelity face swapping. In Proceedings of the 28th ACM International Conference on Multimedia, MM '20, pp.\ 2003–2011, New York, NY, USA, 2020. Association for Computing Machinery. ISBN 9781450379885. doi:10.1145/3394171.3413630. URL https://doi.org/10.1145/3394171.3413630

    arXiv 2003

  8. [8]

    Xception: Deep learning with depthwise separable convolutions

    Chollet, F. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 1251--1258, 2017

    work page 2017

Show all 71 references
  1. [9]

    Arcface: Additive angular margin loss for deep face recognition

    Deng, J., Guo, J., Xue, N., and Zafeiriou, S. Arcface: Additive angular margin loss for deep face recognition. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019 , pp.\ 4690--4699. Computer Vision Foundation / IEEE ,...

    2019

  2. [10]

    Exemplar guided face image super-resolution without facial landmarks

    Dogan, B., Gu, S., and Timofte, R. Exemplar guided face image super-resolution without facial landmarks. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pp.\ 1814--1823, 2019. doi:10.1109/CVPRW.2019.00232

    2019

  3. [11]

    Dolhansky, B., Bitton, J., Pflaum, B., Lu, J., Howes, R., Wang, M., and Ferrer, C. C. The deepfake detection challenge (dfdc) dataset. arXiv preprint arXiv:2006.07397, 2020

    2006 arXiv

  4. [12]

    Implicit identity leakage: The stumbling block to improving deepfake detection generalization

    Dong, S., Wang, J., Ji, R., Liang, J., Fan, H., and Ge, Z. Implicit identity leakage: The stumbling block to improving deepfake detection generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 3994--4004, 2023

    2023

  5. [13]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020

    2010 arXiv

  6. [14]

    Faceswapper.ai

    Faceswapper.ai. Faceswapper.ai. https://faceswapper.ai/swapper. Accessed: 2024-06-17

    2024

  7. [15]

    Gecer, B., Ploumpis, S., Kotsia, I., and Zafeiriou, S. P. Fast-ganfit: Generative adversarial network for high fidelity 3d face reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence, pp.\ 1–1, 2021. ISSN 1939-3539. doi:10.1109/tpami.2021.3084524. URL h...

    2021

  8. [16]

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

    Goodfellow, I. J., Pouget - Abadie, J., Mirza, M., Xu, B., Warde - Farley, D., Ozair, S., Courville, A. C., and Bengio, Y. Generative adversarial nets. In Ghahramani, Z., Welling, M., Cortes, C., Lawrence, N. D., and Weinberger, K. Q. (eds.), Advances in Neural Information Pro...

    2014

  9. [17]

    Google colab

    Google LLC . Google colab. https://colab.research.google.com/. Accessed: 2024-06-17

    2024

  10. [18]

    Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., and Courville, A. C. Improved training of wasserstein gans. In Guyon, I., von Luxburg, U., Bengio, S., Wallach, H. M., Fergus, R., Vishwanathan, S. V. N., and Garnett, R. (eds.), Advances in Neural Information Processing Sy...

    2017

  11. [19]

    Deepfake detection using deep learning methods: A systematic and comprehensive review

    Heidari, A., Jafari Navimipour, N., Dag, H., and Unal, M. Deepfake detection using deep learning methods: A systematic and comprehensive review. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 14 0 (2): 0 e1520, 2024

    2024

  12. [20]

    Denoising diffusion probabilistic models

    Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33: 0 6840--6851, 2020

    2020

  13. [21]

    Feat: Face editing with attention, 2022

    Hou, X., Shen, L., Patashnik, O., Cohen-Or, D., and Huang, H. Feat: Face editing with attention, 2022

    2022

  14. [22]

    B., Ramesh, M., Berg, T., and Learned-Miller, E

    Huang, G. B., Ramesh, M., Berg, T., and Learned-Miller, E. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Technical Report 07-49, University of Massachusetts, Amherst, October 2007

    2007

  15. [23]

    Faking fake news for real fake news detection: Propaganda-loaded training data generation, 2023

    Huang, K.-H., McKeown, K., Nakov, P., Choi, Y., and Ji, H. Faking fake news for real fake news detection: Propaganda-loaded training data generation, 2023

    2023

  16. [24]

    and Belongie, S

    Huang, X. and Belongie, S. J. Arbitrary style transfer in real-time with adaptive instance normalization. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017 , pp.\ 1510--1519. IEEE Computer Society, 2017. doi:10.1109/ICCV.2017.16...

    2017 doi

  17. [25]

    Isola, P., Zhu, J., Zhou, T., and Efros, A. A. Image-to-image translation with conditional adversarial networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017 , pp.\ 5967--5976. IEEE Computer Society, 2017. d...

    2017 doi

  18. [26]

    Jia Guo, J. D. Insightface: 2d and 3d face analysis project. https://github.com/deepinsight/insightface, 2017

    2017

  19. [27]

    and Joo, J

    Karkkainen, K. and Joo, J. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp.\ 1548--1558, 2021

    2021

  20. [28]

    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 International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id=Hk99zCeAb

    2018

  21. [29]

    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 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019 , pp.\ 4401--4410. Computer Vision Foundation / IEEE , ...

    2019

  22. [30]

    Diffface: Diffusion-based face swapping with facial guidance, 2022

    Kim, K., Kim, Y., Cho, S., Seo, J., Nam, J., Lee, K., Kim, S., and Lee, K. Diffface: Diffusion-based face swapping with facial guidance, 2022

    2022

  23. [31]

    Why do facial deepfake detectors fail? In Proceedings of the 2nd Workshop on Security Implications of Deepfakes and Cheapfakes, pp.\ 24--28, 2023

    Le, B., Tariq, S., Abuadbba, A., Moore, K., and Woo, S. Why do facial deepfake detectors fail? In Proceedings of the 2nd Workshop on Security Implications of Deepfakes and Cheapfakes, pp.\ 24--28, 2023

    2023

  24. [32]

    Faceshifter: Towards high fidelity and occlusion aware face swapping

    Li, L., Bao, J., Yang, H., Chen, D., and Wen, F. Faceshifter: Towards high fidelity and occlusion aware face swapping. arXiv preprint arXiv:1912.13457, 2019

    1912 arXiv

  25. [33]

    V., Roy-Chowdhury, A

    Li, S., Khalil, K., Panda, R., Song, C., Krishnamurthy, S. V., Roy-Chowdhury, A. K., and Swami, A. Measurement-driven security analysis of imperceptible impersonation attacks, 2020 a

    2020

  26. [34]

    Learning warped guidance for blind face restoration

    Li, X., Liu, M., Ye, Y., Zuo, W., Lin, L., and Yang, R. Learning warped guidance for blind face restoration. In The European Conference on Computer Vision (ECCV), September 2018

    2018

  27. [35]

    Celeb-df: A large-scale challenging dataset for deepfake forensics

    Li, Y., Yang, X., Sun, P., Qi, H., and Lyu, S. Celeb-df: A large-scale challenging dataset for deepfake forensics. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 3207--3216, 2020 b

    2020

  28. [36]

    Few-shot unsupervised image-to-image translation

    Liu, M., Huang, X., Mallya, A., Karras, T., Aila, T., Lehtinen, J., and Kautz, J. Few-shot unsupervised image-to-image translation. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019 , pp.\ 10550--10559....

    2019

  29. [37]

    Blendgan: Implicitly gan blending for arbitrary stylized face generation

    Liu, M., Li, Q., Qin, Z., Zhang, G., Wan, P., and Zheng, W. Blendgan: Implicitly gan blending for arbitrary stylized face generation. Advances in Neural Information Processing Systems, 34: 0 29710--29722, 2021

    2021

  30. [38]

    Liu, P., Tao, Q., and Zhou, J. T. Evolving from single-modal to multi-modal facial deepfake detection: A survey. arXiv preprint arXiv:2406.06965, 2024

    2024 arXiv

  31. [39]

    Fooocus-inswapper

    machineminded. Fooocus-inswapper. https://github.com/machineminded/Fooocus-inswapper, 2024

    2024

  32. [40]

    Dragondiffusion: Enabling drag-style manipulation on diffusion models, 2023

    Mou, C., Wang, X., Song, J., Shan, Y., and Zhang, J. Dragondiffusion: Enabling drag-style manipulation on diffusion models, 2023

    2023

  33. [41]

    FSGAN : Subject agnostic face swapping and reenactment

    Nirkin, Y., Keller, Y., and Hassner, T. FSGAN : Subject agnostic face swapping and reenactment. In Proceedings of the IEEE International Conference on Computer Vision, pp.\ 7184--7193, 2019

    2019

  34. [42]

    Fsganv2: Improved subject agnostic face swapping and reenactment

    Nirkin, Y., Keller, Y., and Hassner, T. Fsganv2: Improved subject agnostic face swapping and reenactment. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022

    2022

  35. [43]

    A., Jodas, D., da Costa, K

    Passos, L. A., Jodas, D., da Costa, K. A. P., Júnior, L. A. S., Rodrigues, D., Ser, J. D., Camacho, D., and Papa, J. P. A review of deep learning-based approaches for deepfake content detection, 2023

    2023

  36. [44]

    S., RP, L., Jiang, J., et al

    Perov, I., Gao, D., Chervoniy, N., Liu, K., Marangonda, S., Um \'e , C., Dpfks, M., Facenheim, C. S., RP, L., Jiang, J., et al. Deepfacelab: Integrated, flexible and extensible face-swapping framework. arXiv preprint arXiv:2005.05535, 2020

    2005 arXiv

  37. [45]

    Pixble. Pixble. https://pixble.com. Accessed: 2024-06-17

    2024

  38. [46]

    Hierarchical text-conditional image generation with clip latents, 2022

    Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., and Chen, M. Hierarchical text-conditional image generation with clip latents, 2022

    2022

  39. [47]

    Reface. Reface. https://reface.ai/. Accessed: 2024-06-17

    2024

  40. [48]

    Towards the detection of diffusion model deepfakes

    Ricker, J., Damm, S., Holz, T., and Fischer, A. Towards the detection of diffusion model deepfakes. arXiv preprint arXiv:2210.14571, 2022

    2022 arXiv

  41. [49]

    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 Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 10684--10695, 2022

    2022

  42. [50]

    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 Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 1--11, 2019

    2019

  43. [51]

    Deep learning meets liveness detection: Recent advancements and challenges, 2021

    Sabaghi, A., Oghbaie, M., Hashemifard, K., and Akbari, M. Deep learning meets liveness detection: Recent advancements and challenges, 2021

    2021

  44. [52]

    Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S. K. S., Ayan, B. K., Mahdavi, S. S., Lopes, R. G., Salimans, T., Ho, J., Fleet, D. J., and Norouzi, M. Photorealistic text-to-image diffusion models with deep language understanding, 2022

    2022

  45. [53]

    Sangwan, S. Roop. https://github.com/s0md3v/roop, 2023 a

    2023

  46. [54]

    Roop for stablediffusion

    Sangwan, S. Roop for stablediffusion. https://github.com/s0md3v/sd-webui-roop, 2023 b

    2023

  47. [55]

    and Yamasaki, T

    Shiohara, K. and Yamasaki, T. Detecting deepfakes with self-blended images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 18720--18729, 2022

    2022

  48. [56]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., and Ganguli, S. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pp.\ 2256--2265. PMLR, 2015

    2015

  49. [57]

    Swapface

    Swapface. Swapface. https://www.swapface.org/. Accessed: 2024-06-17

    2024

  50. [58]

    and Le, Q

    Tan, M. and Le, Q. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pp.\ 6105--6114. PMLR, 2019

    2019

  51. [59]

    dali lives: museum brings artist back to life with ai

    thedali.org. dali lives: museum brings artist back to life with ai. https://thedali.org/press-room/dali-lives-museum-brings-artists-back-to-life-with-ai/. Accessed: 2024-06-17

    2024

  52. [60]

    N., Kaiser, ., and Polosukhin, I

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, ., and Polosukhin, I. Attention is all you need. Advances in neural information processing systems, 30, 2017

    2017

  53. [61]

    Vidnoz. Vidnoz. https://www.vidnoz.com. Accessed: 2024-06-17

    2024

  54. [62]

    M2tr: Multi-modal multi-scale transformers for deepfake detection

    Wang, J., Wu, Z., Ouyang, W., Han, X., Chen, J., Jiang, Y.-G., and Li, S.-N. M2tr: Multi-modal multi-scale transformers for deepfake detection. In Proceedings of the 2022 international conference on multimedia retrieval, pp.\ 615--623, 2022

    2022

  55. [63]

    High-resolution image synthesis and semantic manipulation with conditional gans

    Wang, T., Liu, M., Zhu, J., Tao, A., Kautz, J., and Catanzaro, B. High-resolution image synthesis and semantic manipulation with conditional gans. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018 , pp.\ 87...

    2018

  56. [64]

    Few-shot classification with feature map reconstruction networks

    Wertheimer, D., Tang, L., and Hariharan, B. Few-shot classification with feature map reconstruction networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 8012--8021, 2021

    2021

  57. [65]

    High-fidelity 3d face generation from natural language descriptions

    Wu, M., Zhu, H., Huang, L., Zhuang, Y., Lu, Y., and Cao, X. High-fidelity 3d face generation from natural language descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 4521--4530, 2023

    2023

  58. [66]

    Gan prior embedded network for blind face restoration in the wild

    Yang, T., Ren, P., Xie, X., and Zhang, L. Gan prior embedded network for blind face restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 672--681, 2021

    2021

  59. [67]

    Y., and Zhao, Y

    Yi, J., Wang, C., Tao, J., Zhang, X., Zhang, C. Y., and Zhao, Y. Audio deepfake detection: A survey, 2023

    2023

  60. [68]

    Dagp-face restorer: Blind face restoration network with domain-aligned generative prior

    Yin, Y., Li, L., Guan, J., and Li, Z. Dagp-face restorer: Blind face restoration network with domain-aligned generative prior. In Proceedings of the 2023 9th International Conference on Computing and Artificial Intelligence, ICCAI '23, pp.\ 222–229, New York, NY, USA, 2023. As...

    2023

  61. [69]

    Multi-attentional deepfake detection

    Zhao, H., Zhou, W., Chen, D., Wei, T., Zhang, W., and Yu, N. Multi-attentional deepfake detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 2185--2194, 2021

    2021

  62. [70]

    Zhou, S., Chan, K., Li, C., and Loy, C. C. Towards robust blind face restoration with codebook lookup transformer. Advances in Neural Information Processing Systems, 35: 0 30599--30611, 2022

    2022

  63. [71]

    V., and Zafarani, R

    Zhou, X., Jain, A., Phoha, V. V., and Zafarani, R. Fake news early detection: An interdisciplinary study, 2020

    2020

Pith tools

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