Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

RAVID: Retrieval-Augmented Visual Detection: A Knowledge-Driven Approach for AI-Generated Image Identification

T0 review · 4 major / 4 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 RAVID claims that retrieval-augmented generation, applied to images for the first time, detects AI-generated pictures at 93.85% average accuracy by retrieving similar reference images and fusing them into a vision-language model.

desk verdict Novel retrieval-augmented detection idea, but the 35.51-point retrieval gap makes the SOTA claim hinge on an unspecified database—peer review yes, citation no. read the letter →

arxiv 2508.03967 v1 pith:GZG3LGYE submitted 2025-08-05 cs.CV cs.CRcs.IR

Mamadou Keita , Wassim Hamidouche , Hessen Bougueffa Eutamene , Abdelmalik Taleb-Ahmed , Abdenour Hadid This is my paper
classification cs.CVcs.CRcs.IR
keywords AI-generatedimagedetectionretrieval-augmentedgenerationvisualRAGvision-languagemodelsCLIPfine-tuninggeneralizationrobustnessUniversalFakeDetect
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 introduces RAVID, a retrieval-augmented framework for AI-generated image detection. Its central claim is that a detector can decide whether an image is synthetic by retrieving similar reference images from a database and feeding the query together with those references into a vision-language model, instead of relying on low-level artifacts or generator fingerprints. On UniversalFakeDetect, which spans 19 generative models, RAVID reports 93.85% average accuracy and, under Gaussian blur and JPEG compression, 80.27% average accuracy versus 63.44% for the strongest comparison method. The authors argue this is the first visual application of RAG to this detection task and that retrieval supplies contextual knowledge that compensates for visual features lost under degradation.

What carries the argument

The central mechanism is the retrieval-augmented decision loop: RAVID CLIP, a CLIP image encoder fine-tuned with category-level prompts, maps the query and all database images into a shared embedding space; a vector database returns the top-k nearest neighbors; and a vision-language model (Qwen-VL, OpenFlamingo, or Gemma3 in the paper's comparisons) fuses the query with the retrieved images to emit the real/fake decision. This loop turns detection into a contextual comparison against known reference images rather than a fixed artifact-classification rule.

What would settle it

Run RAVID on queries from a generator that has no images in the retrieval database, with the database containing only real images. If accuracy falls to the no-retrieval baseline, the retrieval gain is driven by same-generator references rather than by general visual knowledge.

Watch

Extended reading notes

Core claim

RAVID's central claim is that the real/fake boundary can be drawn from retrieved visual context rather than from model-specific fingerprints. The pipeline embeds the query with a fine-tuned CLIP image encoder, fetches the nearest neighbors from a vector database, and concatenates query with neighbors into an enriched input for a vision-language model. The paper reports 93.85% average accuracy on the 19-model UniversalFakeDetect benchmark, 80.27% average accuracy under Gaussian blur and JPEG compression, and 78.81% mean accuracy when trained only on ProGAN and tested on unseen real and synthetic domains. An ablation shows a 35.51 percentage-point gap between 3-shot runs with and without retri

Load-bearing premise

The pipeline assumes the retrieval database holds reference images whose real/fake labels transfer to the query; if the database overlaps the test generators or leaks labels, the reported gains become a label lookup rather than learned detection.

Editorial extensions

If this is right

  • Detection no longer requires knowing a generator's artifacts in advance; the retrieval database gives the vision-language model exemplars to compare against for each query.
  • Under degradation, RAVID keeps 80.27% average accuracy under Gaussian blur and JPEG compression, while the strongest baseline C2P-CLIP drops to 63.44%.
  • Accuracy improves as more retrieved images are supplied, with a 35.51-point gap between 3-shot with and without retrieval, so performance scales with database richness.
  • Trained only on ProGAN images, RAVID generalizes to unseen real and synthetic domains with 78.81% mean accuracy, including high accuracy on localized editing (LTE 99.97%) and latent diffusion (SGXL 98.80%).
  • The framework works with different vision-language models, letting deployment trade off one-shot versus few-shot behavior and retrieval budget.

Reading between the lines

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

  • Because the reported retrieval benefit is large enough to dominate the VLM's own judgment, deployment would need to monitor database composition; the paper does not analyze what happens when the database contains only real images or only unrelated domains.
  • The retrieval-plus-fusion recipe is a general mechanism: the same pipeline could extend to localized edits, face forgery, or camera-model attribution whenever a labeled exemplar set exists.
  • A natural ablation the paper leaves implicit is separating the nearest-neighbor signal from the VLM fusion; measuring RAVID with a nearest-neighbor classifier on RAVID CLIP embeddings alone would isolate how much of the gain is pure retrieval similarity.
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 / 4 minor

Summary. The paper proposes RAVID, a retrieval-augmented framework for AI-generated image detection. A fine-tuned CLIP encoder (RAVID CLIP) embeds the query and a retrieval database; the top-k images are fed together with the query to a VLM (Qwen-VL, OpenFlamingo, or Gemma3) for a final real/fake decision. Experiments on UniversalFakeDetect report 93.85% average accuracy over 19 generators, robustness gains under Gaussian blur and JPEG compression (80.27% vs. 63.44% for C2P-CLIP), and cross-domain generalization when training only on ProGAN. The authors report a 35.51 percentage-point accuracy gap between retrieval and no-retrieval in a 3-shot setting, which they attribute to the value of retrieved context.

Significance. Retrieval-augmented detection is a plausible and relatively underexplored idea, and if the reported numbers are unconfounded, the framework is a meaningful advance over artifact-based detectors: it is evaluated on a broad benchmark, includes robustness tests, and provides an internal ablation quantifying the role of retrieval. The paper also makes productive use of existing VLMs and CLIP encoders rather than training a new detector from scratch. However, the central claim hinges entirely on the contents and separation of the retrieval database, which the manuscript never specifies; until that is resolved, the SOTA numbers cannot be interpreted.

major comments (4)
  1. [Method and experiments (Sections 3–4)] The retrieval database is never specified: its source, size, label structure, and relationship to the UniversalFakeDetect training/test splits are absent. Because the final decision is made by a VLM that sees retrieved images, any overlap or near-duplicate between the database and the test set would reduce the task to test-time label lookup if database labels are known or inferable. This is load-bearing for the headline 93.85% accuracy and must be resolved with a precise protocol (database composition, deduplication, and a leakage test).
  2. [Section 6] The cross-domain experiment claims RAVID is trained only on ProGAN 4-class data and tested on unseen sources, but the retrieval database used in this experiment is not described. If the database contains images from the 'unseen' generators, the claim of generalization to unseen generators collapses. Please specify whether the database is fixed across all experiments, whether it excludes the test generators, and report retrieval accuracy on held-out queries.
  3. [Section 7 (and Section 5 ablation)] The 35.51 percentage-point gap between W/ and W/O retrieval in the 3-shot setting is quoted as evidence for the value of retrieval. As reported, this gap is equally consistent with a simple label-transfer baseline: if the k retrieved images are labeled, a k-nearest-neighbor vote on their labels could attain high accuracy without any learned detection. Please add a label-transfer baseline (e.g., nearest-neighbor classification on RAVID CLIP embeddings) and ablate retrieval performance as a function of k and database composition. Without this, the contribution of the proposed VLM fusion relative to the retrieval set's labels is unidentified.
  4. [Tables 1 and 6] All headline numbers are point estimates with no error bars, number of runs, or statistical significance. Given that the top accuracy in Figure 1 is 93.85% versus 93.79% for C2P-CLIP, the difference is within a plausible range of run-to-run variation. Reporting multiple seeds and confidence intervals is necessary to support the SOTA claim.
minor comments (4)
  1. [Throughout] The name is spelled inconsistently: 'RAVID', 'RA VID', and 'RA VID CLIP' appear interchangeably. Please standardize.
  2. [Abstract and Section 7] Typo 'commun image degradations' should be 'common image degradations'; 'Valeneciennes, Fance' should be 'Valenciennes, France'.
  3. [Figure 1] The figure lists 'RA VID (ours)' but the text uses 'RAVID'; also consider indicating which methods are retrieval-based, as the legend mixes categories.
  4. [Section 5] The VLM comparison and the '3-shot' terminology are not formally defined in the visible text. Define what a 'shot' is in the retrieval setting and how many retrieved images are used for each VLM.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: RAVID's retrieval pipeline is an external-knowledge method, and the 35.51% gap is a component ablation, not a fitted result.

full rationale

The paper's claims do not exhibit a circular derivation. RAVID is a pipeline that (1) embeds a query with a fine-tuned CLIP encoder, (2) retrieves similar images from a vector database, and (3) feeds the query plus retrieved images to a VLM. The retrieved images are inputs to the decision, not fitted parameters. The conclusion's 35.51% performance gap between setups with and without retrieval in the 3-shot setting is an ablation showing that the retrieval component matters; it is not a case where a fitted value is renamed as a prediction. No equation in the manuscript defines the output label as a direct readout of the retrieval database labels, and the paper does not state that the database is built from the test set or contains the test queries. The undisclosed construction of the retrieval database is a legitimate external-validity concern (possible label leakage), but without evidence that the database overlaps the test data or that retrieved labels are directly copied to the query, it does not meet the threshold for circularity. Self-citations [21,22] are prior works on LVLM-based detection and Bi-LORA; they serve as related work/baseline and are not load-bearing for the central RAG claim. Therefore no circular step can be quoted and exhibited under the required standard.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the retrieval database being a legitimate source of transferable labels, on the fine-tuned CLIP encoder finding truly relevant images, and on the VLM combining query and retrieved images in a meaningful way. None of these are proven in the visible text, and the 35.51 point retrieval gap makes the database assumption load-bearing.

free parameters (1)
  • Number of retrieved images k = 3 (highlighted 3-shot setting; optimal not reported)
    The ablation contrasts retrieval versus no retrieval using 3-shot, so k is a user-chosen hyperparameter that directly controls the reported accuracy gain (35.51 points).
assumptions (4)
  • domain assumption Retrieved images provide clean, trustworthy references whose labels transfer to the query.
    The retrieval mechanism assumes similar images in the database carry discriminative signal for the query; without this, the VLM fusion has no extra information to exploit.
  • domain assumption The retrieval database does not overlap test data in a way that leaks labels.
    The preprint does not specify whether database images come from train splits, test domains, or external sources; if overlap exists, accuracy is inflated by test-time label lookup.
  • domain assumption CLIP embeddings fine-tuned with category prompts preserve artificiality cues.
    Retrieval quality depends on the fine-tuned encoder capturing fake/real differences, which is asserted but not independently demonstrated.
  • domain assumption The vision-language model can reason over multiple images to classify realism.
    The final decision is delegated to Qwen-VL or OpenFlamingo, and the multi-image reasoning capability is assumed rather than formally verified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RAVID: Retrieval-Augmented Visual Detection: A Knowledge-Driven Approach for AI-Generated Image Identification." pith.science (2026). https://pith.science/paper/GZG3LGYE

@misc{pith2026250803967,
  author       = {Pith},
  title        = {Pith review of: RAVID: Retrieval-Augmented Visual Detection: A Knowledge-Driven Approach for AI-Generated Image Identification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GZG3LGYE}},
  note         = {Machine review of arXiv:2508.03967}
}
read the original abstract

In this paper, we introduce RAVID, the first framework for AI-generated image detection that leverages visual retrieval-augmented generation (RAG). While RAG methods have shown promise in mitigating factual inaccuracies in foundation models, they have primarily focused on text, leaving visual knowledge underexplored. Meanwhile, existing detection methods, which struggle with generalization and robustness, often rely on low-level artifacts and model-specific features, limiting their adaptability. To address this, RAVID dynamically retrieves relevant images to enhance detection. Our approach utilizes a fine-tuned CLIP image encoder, RAVID CLIP, enhanced with category-related prompts to improve representation learning. We further integrate a vision-language model (VLM) to fuse retrieved images with the query, enriching the input and improving accuracy. Given a query image, RAVID generates an embedding using RAVID CLIP, retrieves the most relevant images from a database, and combines these with the query image to form an enriched input for a VLM (e.g., Qwen-VL or Openflamingo). Experiments on the UniversalFakeDetect benchmark, which covers 19 generative models, show that RAVID achieves state-of-the-art performance with an average accuracy of 93.85%. RAVID also outperforms traditional methods in terms of robustness, maintaining high accuracy even under image degradations such as Gaussian blur and JPEG compression. Specifically, RAVID achieves an average accuracy of 80.27% under degradation conditions, compared to 63.44% for the state-of-the-art model C2P-CLIP, demonstrating consistent improvements in both Gaussian blur and JPEG compression scenarios. The code will be publicly available upon acceptance.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. SPARK-IL: Spectral Retrieval-Augmented RAG for Knowledge-driven Deepfake Detection via Incremental Learning

    cs.CV 2026-04 unverdicted novelty 6.0 of 10

    SPARK-IL reaches 94.6% mean accuracy on deepfake detection across 19 generators by fusing multi-band spectral embeddings from ViT and RGB paths, retrieving nearest signatures for majority voting, and using incremental...

Reference graph

Works this paper leans on

55 extracted references · 40 canonical work pages · cited by 1 Pith paper

  1. [1]

    Create with firefly generative ai

    Adobe. Create with firefly generative ai. https://www. adobe.com/products/firefly.html , 2023. Ac- cessed: 2024-10-10. 3

    work page 2023

  2. [2]

    Openflamingo: An open- source framework for training large autoregressive vision- language models

    Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An open- source framework for training large autoregressive vision- language models. arXiv preprint arXiv:2308.01390 , 2023. 2, 7

    arXiv 2023

  3. [3]

    Large scale gan training for high fidelity natural image synthesis

    Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. In International Conference on Learning Representations. 3, 6

    work page

  4. [4]

    What makes fake images detectable? understanding proper- ties that generalize

    Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding proper- ties that generalize. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVI 16, pages 103–120. Springer, 2020. 7

    work page 2020

  5. [5]

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

    You-Ming Chang, Chen Yeh, Wei-Chen Chiu, and Ning Yu. Antifakeprompt: Prompt-tuned vision-language models are fake image detectors. CoRR, 2023. 3, 6

    work page 2023

  6. [6]

    Learning to see in the dark

    Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In Proceedings of the IEEE con- ference on computer vision and pattern recognition , pages 3291–3300, 2018. 6

    work page 2018

  7. [7]

    Photographic image syn- thesis with cascaded refinement networks

    Qifeng Chen and Vladlen Koltun. Photographic image syn- thesis with cascaded refinement networks. In Proceedings of the IEEE international conference on computer vision, pages 1511–1520, 2017. 6

    work page 2017

  8. [8]

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

    Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified genera- tive adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8789–8797,

Show all 55 references
  1. [9]

    Raising the bar of ai-generated image detection with clip

    Davide Cozzolino, Giovanni Poggi, Riccardo Corvi, Matthias Nießner, and Luisa Verdoliva. Raising the bar of ai-generated image detection with clip. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 4356–4366. IEEE, 2024. 2

    2024

  2. [10]

    Second-order attention network for single im- age super-resolution

    Tao Dai, Jianrui Cai, Yongbing Zhang, Shu-Tao Xia, and Lei Zhang. Second-order attention network for single im- age super-resolution. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 11065–11074, 2019. 6

    2019

  3. [11]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural informa- tion processing systems, 34:8780–8794, 2021. 2, 3, 6

    2021

  4. [12]

    Generative adversarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 3

    2014

  5. [13]

    Ffaa: Multimodal large language model based explainable open-world face forgery analysis assistant

    Zhengchao Huang, Bin Xia, Zicheng Lin, Zhun Mou, and Wenming Yang. Ffaa: Multimodal large language model based explainable open-world face forgery analysis assistant. arXiv preprint arXiv:2408.10072, 2024. 3

    2024 arXiv

  6. [14]

    Synthetic face discrimination via learned image compression

    Sofia Iliopoulou, Panagiotis Tsinganos, Dimitris Ampeliotis, and Athanassios Skodras. Synthetic face discrimination via learned image compression. Algorithms, 17(9):375, 2024. 3

    2024

  7. [15]

    Videorag: Retrieval-augmented generation over video corpus

    Soyeong Jeong, Kangsan Kim, Jinheon Baek, and Sung Ju Hwang. Videorag: Retrieval-augmented generation over video corpus. arXiv preprint arXiv:2501.05874, 2025. 3

    2025 arXiv

  8. [16]

    Bihpf: Bilateral high- pass filters for robust deepfake detection

    Yonghyun Jeong, Doyeon Kim, Seungjai Min, Seongho Joe, Youngjune Gwon, and Jongwon Choi. Bihpf: Bilateral high- pass filters for robust deepfake detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Com- puter Vision, pages 48–57, 2022. 3

    2022

  9. [17]

    Frepgan: robust deepfake detection using frequency- level perturbations

    Yonghyun Jeong, Doyeon Kim, Youngmin Ro, and Jongwon Choi. Frepgan: robust deepfake detection using frequency- level perturbations. In Proceedings of the AAAI conference on artificial intelligence, pages 1060–1068, 2022

    2022

  10. [18]

    Fingerprintnet: Synthesized fin- gerprints for generated image detection

    Yonghyun Jeong, Doyeon Kim, Youngmin Ro, Pyounggeon Kim, and Jongwon Choi. Fingerprintnet: Synthesized fin- gerprints for generated image detection. In European Con- ference on Computer Vision , pages 76–94. Springer, 2022. 3

    2022

  11. [19]

    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. In International Conference on Learning Rep- resentations, 2018. 2, 3, 6

    2018

  12. [20]

    A style-based generator architecture for generative adversarial networks

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 4401–4410, 2019. 2, 3, 6

    2019

  13. [21]

    Harness- ing the power of large vision language models for synthetic image detection

    Mamadou Keita, Wassim Hamidouche, Hassen Bougueffa, Abdenour Hadid, and Abdelmalik Taleb-Ahmed. Harness- ing the power of large vision language models for synthetic image detection. arXiv preprint arXiv:2404.02726, 2024. 6

    2024 arXiv

  14. [22]

    Bi-lora: A vision-language approach for synthetic image detection.Expert Systems, 42(2):e13829,

    Mamadou Keita, Wassim Hamidouche, Hessen Bougu- effa Eutamene, Abdelmalik Taleb-Ahmed, David Camacho, and Abdenour Hadid. Bi-lora: A vision-language approach for synthetic image detection.Expert Systems, 42(2):e13829,

  15. [23]

    Leveraging rep- resentations from intermediate encoder-blocks for synthetic image detection

    Christos Koutlis and Symeon Papadopoulos. Leveraging rep- resentations from intermediate encoder-blocks for synthetic image detection. In European Conference on Computer Vi- sion, pages 394–411. Springer, 2024. 3, 7

    2024

  16. [24]

    Retrieval-augmented generation for knowledge-intensive nlp tasks

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K¨uttler, Mike Lewis, Wen-tau Yih, Tim Rockt ¨aschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing syst...

    2020

  17. [25]

    Diverse image synthesis from semantic layouts via conditional imle

    Ke Li, Tianhao Zhang, and Jitendra Malik. Diverse image synthesis from semantic layouts via conditional imle. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4220–4229, 2019. 6 14

    2019

  18. [26]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892–34916, 2023. 3

    2023

  19. [27]

    Forgery-aware adaptive transformer for generalizable synthetic image detection

    Huan Liu, Zichang Tan, Chuangchuang Tan, Yunchao Wei, Jingdong Wang, and Yao Zhao. Forgery-aware adaptive transformer for generalizable synthetic image detection. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 10770–10780, 2024. 7

    2024

  20. [28]

    Midjourney v5

    MidJourney. Midjourney v5. https : / / www . midjourney.com, 2023. Accessed: 2024-10-10. 3

    2023

  21. [29]

    Clip- cap: Clip prefix for image captioning

    Ron Mokady, Amir Hertz, and Amit H Bermano. Clip- cap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 4

    2021 arXiv

  22. [30]

    Detecting gan generated fake images using co-occurrence matrices

    Lakshmanan Nataraj, Tajuddin Manhar Mohammed, Shiv- kumar Chandrasekaran, Arjuna Flenner, Jawadul H Bappy, Amit K Roy-Chowdhury, and BS Manjunath. Detecting gan generated fake images using co-occurrence matrices. arXiv preprint arXiv:1903.06836, 2019. 7

    1903 arXiv

  23. [31]

    Glide: Towards photorealis- tic image generation and editing with text-guided diffusion models

    Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealis- tic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, p...

    2022

  24. [32]

    Towards uni- versal fake image detectors that generalize across genera- tive models

    Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards uni- versal fake image detectors that generalize across genera- tive models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24480– 24489, 2023. 6, 7

    2023

  25. [33]

    Dall-e 3

    OpenAI. Dall-e 3. https://openai.com/dall-e-3,

  26. [34]

    Gaugan: semantic image synthesis with spatially adap- tive normalization

    Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Gaugan: semantic image synthesis with spatially adap- tive normalization. In ACM SIGGRAPH 2019 Real-Time Live!, New York, NY , USA, 2019. Association for Comput- ing Machinery. 3, 6

    2019

  27. [35]

    Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M ¨uller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3

    2023 arXiv

  28. [36]

    Thinking in frequency: Face forgery detection by min- ing frequency-aware clues

    Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by min- ing frequency-aware clues. In European conference on com- puter vision, pages 86–103. Springer, 2020. 7

    2020

  29. [37]

    Zero-shot text-to-image generation

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea V oss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International confer- ence on machine learning, pages 8821–8831. Pmlr, 2021. 2, 6

    2021

  30. [38]

    Videorag: Retrieval-augmented gen- eration with extreme long-context videos

    Xubin Ren, Lingrui Xu, Long Xia, Shuaiqiang Wang, Dawei Yin, and Chao Huang. Videorag: Retrieval-augmented gen- eration with extreme long-context videos. arXiv preprint arXiv:2502.01549, 2025. 3

    2025 arXiv

  31. [39]

    Beyond text: Optimizing rag with multimodal inputs for industrial applications

    Monica Riedler and Stefan Langer. Beyond text: Optimizing rag with multimodal inputs for industrial applications. arXiv preprint arXiv:2410.21943, 2024. 3

    2024 arXiv

  32. [40]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 2, 3, 6

    2022

  33. [41]

    Faceforen- sics++: Learning to detect manipulated facial images

    Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Chris- tian Riess, Justus Thies, and Matthias Nießner. Faceforen- sics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1–11, 2019. 6

    2019

  34. [42]

    Photorealistic text-to-image diffusion models with deep language understanding

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information...

    2022

  35. [43]

    Stylegan- xl: Scaling stylegan to large diverse datasets

    Axel Sauer, Katja Schwarz, and Andreas Geiger. Stylegan- xl: Scaling stylegan to large diverse datasets. In ACM SIG- GRAPH 2022 conference proceedings, pages 1–10, 2022. 3

    2022

  36. [44]

    Deep image fingerprint: To- wards low budget synthetic image detection and model lin- eage analysis

    Sergey Sinitsa and Ohad Fried. Deep image fingerprint: To- wards low budget synthetic image detection and model lin- eage analysis. In Proceedings of the IEEE/CVF Winter Con- ference on Applications of Computer Vision , pages 4067– 4076, 2024. 2

    2024

  37. [45]

    Learning on gradients: Generalized arti- facts representation for gan-generated images detection

    Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, and Yunchao Wei. Learning on gradients: Generalized arti- facts representation for gan-generated images detection. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 12105–12114, 2023. 7

    2023

  38. [46]

    C2p-clip: Inject- ing category common prompt in clip to enhance generaliza- tion in deepfake detection.arXiv preprint arXiv:2408.09647,

    Chuangchuang Tan, Renshuai Tao, Huan Liu, Guanghua Gu, Baoyuan Wu, Yao Zhao, and Yunchao Wei. C2p-clip: Inject- ing category common prompt in clip to enhance generaliza- tion in deepfake detection.arXiv preprint arXiv:2408.09647,

    arXiv

  39. [47]

    Frequency-aware deepfake de- tection: Improving generalizability through frequency space domain learning

    Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Frequency-aware deepfake de- tection: Improving generalizability through frequency space domain learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5052–5060, 2024. 7

    2024

  40. [48]

    Rethinking the up-sampling op- erations 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 op- erations in cnn-based generative network for generalizable deepfake detection. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition,...

    2024

  41. [49]

    Milvus documentation

    Milvus Team. Milvus documentation. https : / / milvus.io/docs/fr, 2025. Accessed: 2025-03-08. 5

    2025

  42. [50]

    Cnn-generated images are surprisingly easy to spot

    Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8695–8704, 2020. 6, 7

    2020

  43. [51]

    Dire for diffusion-generated image detection

    Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 22445–22455, 2023. 2, 3, 6 15

    2023

  44. [52]

    Visrag: Vision-based retrieval-augmented generation on multi-modality documents

    Shi Yu, Chaoyue Tang, Bokai Xu, Junbo Cui, Junhao Ran, Yukun Yan, Zhenghao Liu, Shuo Wang, Xu Han, Zhiyuan Liu, et al. Visrag: Vision-based retrieval-augmented generation on multi-modality documents. arXiv preprint arXiv:2410.10594, 2024. 3

    2024 arXiv

  45. [53]

    Detecting and simulating artifacts in gan fake images

    Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. In2019 IEEE in- ternational workshop on information forensics and security (WIFS), pages 1–6. IEEE, 2019. 7

    2019

  46. [54]

    Unpaired image-to-image translation using cycle- consistent adversarial networks

    Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle- consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision , pages 2223– 2232, 2017. 3, 6 16

    2017

  47. [2023]

    Accessed: 2024-10-10. 3

    2024

Pith tools

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