Pith. sign in

REVIEW 4 major objections 5 minor 42 references

CorrDetail: Visual Detail Enhanced Self-Correction for Face Forgery Detection

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read CorrDetail claims that feeding a face-forgery detector questions containing deliberately wrong detail descriptions trains it to correct hallucinations, and that adding fine-grained visual features plus a fused decision strategy yields…

desk verdict A solid, genuinely new deepfake-detection framework whose strongest claims are undercut by an internal SCVQA description conflict and an untested shortcut risk; worth refereeing but not yet fully convincing. read the letter →

arxiv 2507.05302 v1 pith:5G3RVDEB submitted 2025-07-07 cs.CV cs.AI

classification cs.CVcs.AI
keywords faceforgerydetectiondeepfakevision-languagemodelsself-correctionvisualquestionansweringhallucinationmitigationcross-modalattentiongeneralization
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 is trying to establish that a vision-language model can detect facial deepfakes and explain its decision by naming real forgery details, rather than giving binary labels or hallucinated cues. CorrDetail trains the model on a self-correction visual question-answering dataset in which 70% of questions contain an incorrect forgery detail, 15% contain a correctly rephrased detail, and 15% are unchanged; the model must answer with the authentic detail, learning to override the prompt when the prompt is wrong. A cross-modal module injects fine-grained visual features back into the VLM, and a fusion strategy combines the VLM's reweighted answer with an independent visual classifier. The reported result is 96.93% AUC on low-quality FF++, 99.92% on high-quality FF++, and 72.80% AUC on CelebDF when trained only on FF++, with gains also on DeepFaceGen and WildDeepfake under cross-testing.

What carries the argument

The central mechanism is the self-correction training scheme (SCVQA dataset): with 70% probability the question embeds a wrong forgery detail, with 15% a correct detail, and with 15% no modification, teaching the model to reject the prompt's assertion when it contradicts the image. The other load-bearing pieces are the Cross-model Forgery Detail Enhancement (CFDE) module, which performs cross-modal attention with an extra text prompt as query, image tokens as value, and ViT features as key to sharpen forgery-relevant visual tokens, and the decision-fusion strategy that combines a prompt-enhancement rule with a separately trained visual expert.

What would settle it

Give the trained model pairs of questions on the same forged image, one asserting a wrong detail and one asserting a correct detail, and check whether its answers cite the actual artifact or simply flip with the assertion; if the answer always negates the asserted detail regardless of what is really forged, the self-correction claim collapses.

Watch

Extended reading notes

Core claim

CorrDetail is a framework that turns vision-language model hallucination into a training signal for face forgery detection. On the SCVQA dataset built from FF++, the model is fine-tuned with three question types: 70% with an incorrect detail assertion, 15% with a correct detail paraphrase, and 15% with the original question, so that the correct answer requires the model to verify the asserted detail against the image and correct the error when present. The Cross-model Forgery Detail Enhancement (CFDE) module takes an extra text prompt as the query, the projected image tokens as the value, and ViT-extracted visual features as the key in a cross-modal attention layer, then adds the attention map to the original image tokens in a residual way to preserve forgery-specific nuances. During inference, a prompt enhancement mechanism re-asks the question with an attention-guiding phrase and adjusts the probability when the second answer conflicts with the first, and a dual-expert rule multiplies the VLM's probability with that of a separately trained visual branch. The paper reports that this design achieves the highest accuracy and AUC on FF++ intra-testing, the highest cross-dataset AUC on CelebDF, and beats detectors trained directly on DeepFaceGen when tested there.

Load-bearing premise

The load-bearing premise is that the model learns to identify genuine forgery details from the 70% erroneous questions rather than learning the simpler rule of contradicting whatever detail the question asserts; the paper does not provide direct evidence that this shortcut is avoided.

Editorial extensions

If this is right

  • A multimodal detector can be made to produce interpretable, artifact-level explanations instead of hallucinated ones by training on deliberately wrong plus correct detail assertions.
  • The reported CelebDF and DeepFaceGen results imply that correction-based training transfers to unseen face-swap and AIGC forgeries better than the compared visual and multimodal baselines.
  • The dual-expert fusion implies that hard samples with small faces or dominant backgrounds can be rescued by combining the VLM's probability with a purely visual branch and by re-checking with an attention prompt.
  • The CFDE design implies that preserving fine-grained visual tokens during text-image alignment is a substantive accuracy lever, not just an interpretability aid.

Reading between the lines

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

  • A testable extension is to measure whether the model truly corrects rather than negates: vary the plausibility of the injected wrong detail and see whether the answer tracks the image artifact or merely inverts the prompt.
  • The 70/15/15 split is an arbitrary design choice; sweeping the split would reveal whether the shortcut-avoidance the paper relies on holds across ratios, and whether lower error rates still teach correction.
  • Because the base model is a 7B LLaVA, the framework's cost is high; the same self-correction recipe could be probed on smaller VLMs to see whether the mechanism or the scale carries the gain.
  • The fusion rule operates on two probability estimates, so it could be applied post hoc to other detectors; that would test the generalization of the decision strategy independently of the SCVQA training.
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. The paper presents CorrDetail, a face forgery detection framework that combines a Self-Correction Visual Question Answering (SCVQA) training strategy with a Cross-Model Forgery Detail Enhancement (CFDE) module and a decision-fusion mechanism. The key idea is to train a VLM to correct erroneous forgery details in questions, thereby improving both classification accuracy and interpretability. The framework is evaluated on FF++ (LQ/HQ), CelebDF, DFR, WDF, and DeepFaceGen, reporting state-of-the-art point estimates in most metrics, and the authors provide ablations for each module.

Significance. If the reported results are robust, CorrDetail offers a plausible and useful direction for combining VLM self-correction with visual detail enhancement in deepfake detection. The paper's strengths include evaluation across several standard benchmarks, a modular design that is easy to understand, and ablation studies that isolate the contributions of SCVQA, CFDE, and decision fusion. The qualitative examples in Figures 4 and 6 also help convey the intended interpretability benefit. However, the central claims rest on an under-specified training-data construction and on heuristic fusion choices with no sensitivity analysis or error bars, and the absence of code or dataset release limits reproducibility. These gaps need to be addressed before the SOTA and interpretability claims can be fully assessed.

major comments (4)
  1. [Section 3.1 vs Section 4 (SCVQA Dataset)] The description of the SCVQA dataset is internally contradictory and this is load-bearing for the method. Section 3.1 states that 70% of training questions receive erroneous details, 15% receive genuine details via synonym substitution, and 15% remain unaltered, with the intended behavior being that the model corrects the erroneous detail. Section 4, however, defines the three data formats as (1) 70% '[image, question, incorrect detail description as the answer],' (2) 15% '[image, question, null answer],' and (3) 15% '[image, question, correct detail description as the answer].' These two descriptions cannot both be true: the first says the answer is the incorrect detail, which would train the model to output the wrong detail and invert the self-correction objective, while the second discards the 'unaltered' category and replaces it with an undefined 'null answer.' The authors must specify the exact training targets for each of the three categories and reconcile Section 3.1 with Section 4.
  2. [Section 3.1 and Table 3 (shortcut risk)] The paper acknowledges the shortcut risk in Section 3.1: 'We deliberately refrain from entirely inserting incorrect details, as this could prompt the model to exploit the prompt directly, deriving answers opportunistically.' With 70% of training questions asserting an incorrect detail, a model trained with next-token or cross-entropy loss can achieve low training loss by always responding with the opposite of (or copying) the asserted detail, while ignoring the image content. The presence of 15% genuine and 15% unaltered examples does not by itself prevent this, and no experiment is reported that varies the prompt detail while holding the image fixed, nor is any metric reported for the correctness of the VQA answers. The ablation in Table 3 reports only ACC/AUC, not answer-level accuracy. The authors should add controlled experiments or quantitative analysis showing that the model actually uses visual evidence to correct, rather than exploit, the prompt; otherwise the interpretability claim and part of the classification gain may be due to a prompt prior rather than genuine visual understanding.
  3. [Section 3.3, Tables 1 and 2, Table 3] The fusion decision strategy depends on hyperparameter choices that are not analyzed. The exclusion threshold lambda is set to 0.1 with no sensitivity study, and the piecewise fusion rule is motivated by a single qualitative example (Figure 3) rather than by a quantitative comparison with alternative rules. Moreover, all reported results in Tables 1, 2, and 3 are point estimates from what appears to be a single run, with no error bars, multiple seeds, or significance tests. Since the SOTA claim relies on differences such as 96.93 vs 95.36 AUC on FF++ (LQ) and 72.80 vs 70.15 AUC on CelebDF, the authors should report variance across runs and include a lambda sweep or an alternative fusion baseline to demonstrate that the proposed rule is not overfit to the evaluated datasets.
  4. [Section 4.2, Table 2, Figure 5] The cross-dataset evaluation protocol is ambiguous and must be clarified. The text says 'on CelebDF, we trained CorrDetail on the SCVQA and performed cross-dataset evaluation by testing it on CelebDF,' which is self-contradictory, and then refers to 'the experimental results for the CelebDF are illustrated in Table 2,' although Table 2 reports DFR and WDF. It is also unclear whether the DFR/WDF and DeepFaceGen generalization tests use FF++ (LQ) or FF++ (HQ) as the training source: Table 1 states that CelebDF cross-testing is trained on FF++ (LQ), while Table 2's caption says 'trained on FF++ (HQ).' The authors should specify the exact training dataset, quality level, and protocol for each cross-dataset experiment so that the generalization claims can be verified.
minor comments (5)
  1. [Section 4, SCVQA Dataset] The term 'null answer' in bullet (2) is not defined and does not appear in Section 3.1, which instead says 15% of questions remain unaltered. Please clarify whether these questions have an empty answer, the original DD-VQA answer, or something else.
  2. [Section 1, Contributions] The contribution bullet says 'multi-expert reasoning decisions,' but Section 3.3 implements a dual-expert decision consisting of the VLM branch and one visual branch. Please rename or explain what 'multi-expert' refers to.
  3. [Table 1] The caption states that the highest and second-highest performances are highlighted, but the highlighting is not visible in the submitted version; please ensure the final rendering distinguishes the best and second-best values clearly.
  4. [Figure 4] Figure 4 refers to 'failure cases of GPT-4o mini,' but GPT-4o mini is not otherwise introduced as a baseline. Please state its role, configuration, and how its responses were obtained.
  5. [Reproducibility] The paper does not mention whether the SCVQA dataset, code, or trained models will be released. Since SCVQA is one of the main contributions, a clear data/code availability statement would be important for reproducibility.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: CorrDetail is trained on externally sourced VQA annotations and evaluated on standard held-out benchmarks; the only self-citation is a benchmark reference, not a load-bearing derivation step.

full rationale

CorrDetail's training signal is the SCVQA dataset built from FF++ and DD-VQA-style VQA formats, with target answers described as authentic forgery details (Section 3.1), and evaluation is performed on held-out FF++, CelebDF, WDF, DFR, and DeepFaceGen. No equation in the paper reduces a reported AUC or accuracy to the training labels or to a fitted parameter renamed as a prediction; the hyperparameters (e.g., the 70/15/15 split and lambda = 0.1) are design choices, not outputs claimed as results. The only self-citation is [Bei et al., 2024] for the DeepFaceGen benchmark, used as an external evaluation dataset rather than as a proof of correctness, so it is not load-bearing circularity. The manuscript itself flags a possible shortcut in Section 3.1 ('this could prompt the model to exploit the prompt directly, deriving answers opportunistically'), but that is a robustness concern, not a circularity, because the training target is not defined as the model's own output. Section 4's phrasing 'incorrect detail description as the answer' is internally inconsistent with Section 3.1, but this is a reporting/soundness issue rather than an equivalence between input and prediction. Overall, the derivation chain is self-contained against external benchmarks, and no quoted step exhibits reduction of the claimed result to its own inputs.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on hand-chosen hyperparameters (lambda, mix ratios) and domain assumptions about shortcut learning, dataset integrity, and fusion calibration. No new physical or formal entities are introduced.

free parameters (2)
  • lambda (exclusion threshold) = 0.1
    Used in the prompt enhancement mechanism to decide when to keep the initial prediction vs apply fusion. Chosen by hand, no sensitivity analysis.
  • SCVQA mix ratios = 70%/15%/15%
    Proportions of erroneous detail, null, and correct detail question/answer triples. Hand-chosen to prevent prompt exploitation; no tuning analysis.
assumptions (4)
  • domain assumption The 70/15/15 composition of SCVQA prevents the VLM from learning a shortcut of trusting or always negating the prompt detail.
    The paper relies on this to claim the model learns authentic forgery details; Section 3.1 acknowledges the shortcut risk.
  • domain assumption The FF++ training/test split is standard and the SCVQA triples are built only from the training split.
    No explicit split information is given; intra-test results on FF++ assume no train/test contamination.
  • domain assumption The product fusion of probabilities from the VLM and visual branches yields a well-calibrated decision.
    Equation in Section 3.3 multiplies probabilities without normalization or calibration analysis.
  • standard math Softmax and attention operations in CFDE are computed with compatible dimensions.
    The paper does not specify dimension alignment for the cross-modal attention in Section 3.2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CorrDetail: Visual Detail Enhanced Self-Correction for Face Forgery Detection." pith.science (2026). https://pith.science/paper/5G3RVDEB

@misc{pith2026250705302,
  author       = {Pith},
  title        = {Pith review of: CorrDetail: Visual Detail Enhanced Self-Correction for Face Forgery Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5G3RVDEB}},
  note         = {Machine review of arXiv:2507.05302}
}
read the original abstract

With the swift progression of image generation technology, the widespread emergence of facial deepfakes poses significant challenges to the field of security, thus amplifying the urgent need for effective deepfake detection.Existing techniques for face forgery detection can broadly be categorized into two primary groups: visual-based methods and multimodal approaches. The former often lacks clear explanations for forgery details, while the latter, which merges visual and linguistic modalities, is more prone to the issue of hallucinations.To address these shortcomings, we introduce a visual detail enhanced self-correction framework, designated CorrDetail, for interpretable face forgery detection. CorrDetail is meticulously designed to rectify authentic forgery details when provided with error-guided questioning, with the aim of fostering the ability to uncover forgery details rather than yielding hallucinated responses. Additionally, to bolster the reliability of its findings, a visual fine-grained detail enhancement module is incorporated, supplying CorrDetail with more precise visual forgery details. Ultimately, a fusion decision strategy is devised to further augment the model's discriminative capacity in handling extreme samples, through the integration of visual information compensation and model bias reduction.Experimental results demonstrate that CorrDetail not only achieves state-of-the-art performance compared to the latest methodologies but also excels in accurately identifying forged details, all while exhibiting robust generalization capabilities.

Figures

Figures reproduced from arXiv: 2507.05302 by the authors.

Figure 1
Figure 1. The illustration for CorrDetail, which is designed to cor￾rect authentic forgery details ("eyes look asymmetrical...") when provided with error-guided questioning ("nose look unnatural...") with the assistance of visual fine-grained detail enhancement. industry. However, the capability of AIGC to produce high￾quality content from simple prompts and guidance has simul￾taneously lowered the barriers for malicious acto… view at source ↗
Figure 2
Figure 2. Framework of the proposed CorrDetail. During training, the self-correction Q&A strategy is employed to enhance the forgery clue extraction ability of the VLM, while the visual branch integrates a dual-stream architecture (CLIP + ViT) with the Cross-Model Forgery Detail Enhancement (CFDE) module to provide detailed visual features. In the inference phase, the input image and dual-round prompts are initially processed… view at source ↗
Figure 3
Figure 3. One case for Prompt Enhancement Mechanism handling hard samples (e.g., those with smaller facial proportions). Formally, the extra text prompt Pextra is processed through the same tokenizer and fused with the image tokens Itoken and the image visual feature Ivit (processed through the Vi￾sion Transformer) in a cross-modal attention mechanism. The extra text prompt is used as the query Q, the image tokens as the valu… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Qualitative examples of results on certain samples. For failure cases of GPT-4o mini, CorrDetail achieves correct classifi￾cation results along with corresponding explanations. 4 Experiments In this section, we introduce the experimental setup and en￾vironment, present…
Figure 5
Figure 5. Figure 5: The generalization performance of CorrDetail in comparison to the intra-performance of existing methods. The left radar chart illustrates the detection efficacy across various types of forgery samples, with solid lines representing the comparison techniques and a dashe…
Figure 6
Figure 6. Figure 6: The heatmap visualization of ablation study on CFDE. noting that during this ablation process, no additional innova￾tive modules were incorporated—only prompt guidance was added. Clearly, for some images that 4o mini fails to clas￾sify correctly, our model accurately i…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 38 canonical work pages

  1. [1]

    A large-scale universal evaluation bench- mark for face forgery detection,

    [Bei et al., 2024] Yijun Bei, Hengrui Lou, Jinsong Geng, Erteng Liu, Lechao Cheng, Jie Song, Mingli Song, and Zunlei Feng. A large-scale universal evaluation bench- mark for face forgery detection,

  2. [5]

    Local relation learning for face forgery detection

    [Chen et al., 2021] Shen Chen, Taiping Yao, Yang Chen, Shouhong Ding, Jilin Li, and Rongrong Ji. Local relation learning for face forgery detection. In AAAI, volume 35, pages 1081–1088,

  3. [8]

    Dong, Jin Wang, Jiajun Liang, Hao- qiang Fan, and Renhe Ji

    [Dong et al., 2022] S. Dong, Jin Wang, Jiajun Liang, Hao- qiang Fan, and Renhe Ji. Explaining deepfake detection by analysing image matching. In ECCV,

  4. [9]

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

    [Dosovitskiy, 2020] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv,

  5. [11]

    Large language models cannot self-correct reasoning yet

    [Huang et al., 2024] Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, and Denny Zhou. Large language models cannot self-correct reasoning yet. In ICLR,

  6. [12]

    Deeperforensics-1.0: A large-scale dataset for real-world face forgery detection

    [Jiang et al., 2020] Liming Jiang, Ren Li, Wayne Wu, Chen Qian, and Chen Change Loy. Deeperforensics-1.0: A large-scale dataset for real-world face forgery detection. In CVPR, pages 2889–2898,

  7. [13]

    Clipping the deception: Adapt- ing vision-language models for universal deepfake detec- tion

    [Khan and Dang-Nguyen, 2024] Sohail Ahmed Khan and Duc-Tien Dang-Nguyen. Clipping the deception: Adapt- ing vision-language models for universal deepfake detec- tion. In ICMR, pages 1006–1015,

  8. [14]

    Moe-ffd: Mixture of experts for general- ized and parameter-efficient face forgery detection

    [Kong et al., 2024] Chenqi Kong, Anwei Luo, Peijun Bao, Yi Yu, Haoliang Li, Zengwei Zheng, Shiqi Wang, and Alex C Kot. Moe-ffd: Mixture of experts for general- ized and parameter-efficient face forgery detection. arXiv,

Show all 42 references
  1. [15]

    Face x-ray for more general face forgery detection

    [Li et al., 2020a] Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Baining Guo. Face x-ray for more general face forgery detection. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5000–5009,

  2. [16]

    Blip: Bootstrapping language-image pre- training for unified vision-language understanding and generation

    [Li et al., 2022] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre- training for unified vision-language understanding and generation. In ICML,

  3. [17]

    Let’s verify step by step,

    [Lightman et al., 2023] Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step,

  4. [18]

    [Liu et al., 2023] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee

    Association for Computational Linguistics. [Liu et al., 2023] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning,

  5. [19]

    Generalizing face forgery detection with high- frequency features

    [Luo et al., 2021] Yuchen Luo, Yong Zhang, Junchi Yan, and Wei Liu. Generalizing face forgery detection with high- frequency features. In CVPR, pages 16317–16326,

  6. [20]

    Beyond the prior forgery knowledge: Mining critical clues for general face forgery detection

    [Luo et al., 2023] Anwei Luo, Chenqi Kong, Jiwu Huang, Yongjian Hu, Xiangui Kang, and Alex C Kot. Beyond the prior forgery knowledge: Mining critical clues for general face forgery detection. IEEE Transactions on Information F orensics and Security, 19:1168–1182,

  7. [21]

    Critique abil- ity of large language models,

    [Luo et al., 2024] Liangchen Luo, Zi Lin, Yinxiao Liu, Lei Shu, Yun Zhu, Jingbo Shang, and Lei Meng. Critique abil- ity of large language models,

  8. [22]

    Self-refine: Iterative refinement with self- feedback,

    [Madaan et al., 2023] Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder, Kather- ine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter Clar...

  9. [23]

    Detecting gan-generated imagery using color cues

    [McCloskey and Albright, 2018] Scott McCloskey and Michael Albright. Detecting gan-generated imagery using color cues. ArXiv, abs/1812.08247,

  10. [26]

    Thinking in frequency: Face forgery detection by mining frequency-aware clues

    [Qian et al., 2020] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In ECCV, page 86–103, Berlin, Heidelberg,

  11. [27]

    Springer- Verlag. [Radford et al., 2021] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agar- wal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural...

  12. [28]

    Self-critiquing models for assisting human eval- uators,

    [Saunders et al., 2022] William Saunders, Catherine Yeh, Jeff Wu, Steven Bills, Long Ouyang, Jonathan Ward, and Jan Leike. Self-critiquing models for assisting human eval- uators,

  13. [29]

    Adaptive face forgery detection in cross domain

    [Song et al., 2022] Luchuan Song, Zheng Fang, Xiaodan Li, Xiaoyi Dong, Zhenchao Jin, Yuefeng Chen, and Siwei Lyu. Adaptive face forgery detection in cross domain. In ECCV, pages 467–484. Springer,

  14. [30]

    Stable Diffusion

    [Stability.ai, 2023] Stability.ai. Stable Diffusion. https:// stability.ai/,

  15. [31]

    Domain gen- eral face forgery detection by learning to weight

    [Sun et al., 2021] Ke Sun, Hong Liu, Qixiang Ye, Yue Gao, Jianzhuang Liu, Ling Shao, and Rongrong Ji. Domain gen- eral face forgery detection by learning to weight. In AAAI, volume 35, pages 2638–2646,

  16. [32]

    Dual con- trastive learning for general face forgery detection

    [Sun et al., 2022] Ke Sun, Taiping Yao, Shen Chen, Shouhong Ding, Jilin Li, and Rongrong Ji. Dual con- trastive learning for general face forgery detection. In AAAI, volume 36, pages 2316–2324,

  17. [33]

    Efficient- net: Rethinking model scaling for convolutional neural networks

    [Tan and Le, 2019] Mingxing Tan and Quoc Le. Efficient- net: Rethinking model scaling for convolutional neural networks. In ICML, pages 6105–6114. PMLR,

  18. [34]

    Representative forgery mining for fake face detection

    [Wang and Deng, 2021] Chengrui Wang and Weihong Deng. Representative forgery mining for fake face detection. In CVPR, pages 14923–14932,

  19. [35]

    M2tr: Multi-modal multi-scale transform- ers for deepfake detection

    [Wang et al., 2022] Junke Wang, Zuxuan Wu, Wenhao Ouyang, Xintong Han, Jingjing Chen, Yu-Gang Jiang, and Ser-Nam Li. M2tr: Multi-modal multi-scale transform- ers for deepfake detection. In ICMR, ICMR ’22, page 615–623, New York, NY , USA,

  20. [36]

    [Wu et al., 2023] H

    Association for Computing Machinery. [Wu et al., 2023] H. Wu, J. Zhou, and S. Zhang. General- izable synthetic image detection via language-guided con- trastive learning. arXiv,

  21. [37]

    Visco: Benchmarking fine-grained critique and correction towards self-improvement in visual reasoning,

    [Wu et al., 2024] Xueqing Wu, Yuheng Ding, Bingxuan Li, Pan Lu, Da Yin, Kai-Wei Chang, and Nanyun Peng. Visco: Benchmarking fine-grained critique and correction towards self-improvement in visual reasoning,

  22. [38]

    Large language models are con- trastive reasoners,

    [Yao, 2024] Liang Yao. Large language models are con- trastive reasoners,

  23. [39]

    Common sense reason- ing for deepfake detection

    [Zhang et al., 2025] Yue Zhang, Ben Colman, Xiao Guo, Ali Shahriyari, and Gaurav Bharaj. Common sense reason- ing for deepfake detection. In ECCV, pages 399–415. Springer,

  24. [41]

    Uia-vit: Unsupervised inconsistency- aware method based on vision transformer for face forgery detection

    [Zhuang et al., 2022] Wanyi Zhuang, Qi Chu, Zhentao Tan, Qiankun Liu, Haojie Yuan, Changtao Miao, Zixiang Luo, and Nenghai Yu. Uia-vit: Unsupervised inconsistency- aware method based on vision transformer for face forgery detection. In ECCV, pages 391–407. Springer,

  25. [42]

    Wilddeepfake: A chal- lenging real-world dataset for deepfake detection

    [Zi et al., 2020] Bojia Zi, Minghao Chang, Jingjing Chen, Xingjun Ma, and Yu-Gang Jiang. Wilddeepfake: A chal- lenging real-world dataset for deepfake detection. In ACM MM, pages 2382–2390, 2020

  26. [2017]

    On the detection of synthetic images generated by diffusion models

    [Corvi et al., 2023] Riccardo Corvi, Davide Cozzolino, Gi- ada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. In ICASSP, pages 1–5,

  27. [2018]

    Detecting gan-generated imagery using saturation cues

    [McCloskey and Albright, 2019] Scott McCloskey and Michael Albright. Detecting gan-generated imagery using saturation cues. In 2019 IEEE International Conference on Image Processing (ICIP), pages 4584–4588,

  28. [2019]

    Midjourney

    [Midjourney, 2022] Midjourney. Midjourney. https://www. midjourney.com/home,

  29. [2020]

    Critic: Large language models can self-correct with tool- interactive critiquing,

    [Gou et al., 2024] Zhibin Gou, Zhihong Shao, Yeyun Gong, Yelong Shen, Yujiu Yang, Nan Duan, and Weizhu Chen. Critic: Large language models can self-correct with tool- interactive critiquing,

  30. [2021]

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

    [Choi et al., 2017] Yunjey Choi, Min-Je Choi, Mun Su Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Star- gan: Unified generative adversarial networks for multi- domain image-to-image translation. CVPR, pages 8789– 8797,

  31. [2022]

    Yu, and Lichao Sun

    [Cao et al., 2023] Yihan Cao, Siyu Li, Yixin Liu, Zhiling Yan, Yutong Dai, Philip S. Yu, and Lichao Sun. A compre- hensive survey of ai-generated content (aigc): A history of generative ai from gan to chatgpt. ArXiv, abs/2303.04226,

  32. [2023]

    Antifakeprompt: Prompt-tuned vision- language models are fake image detectors,

    [Chang et al., 2024] You-Ming Chang, Chen Yeh, Wei-Chen Chiu, and Ning Yu. Antifakeprompt: Prompt-tuned vision- language models are fake image detectors,

  33. [2024]

    Large scale gan training for high fidelity natural image synthesis

    [Brock et al., 2018] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. ArXiv, abs/1809.11096,

  34. [2025]

    A survey of large language models,

    [Zhao et al., 2023] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Zhipeng Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jian-Yu...

Pith tools

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