Pith. sign in

REVIEW 3 major objections 3 minor 88 references

RAIDX: A Retrieval-Augmented Generation and GRPO Reinforcement Learning Framework for Explainable Deepfake Detection

T0 review · 3 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read RAIDX claims a deepfake detector that retrieves similar images, then explains its verdicts in text and heatmaps.

desk verdict The RAG+GRPO combination is new, but the paper's own SID-Set numbers disagree by ~6 points across tables, and the explainability claims are not backed by the reward design. read the letter →

arxiv 2508.04524 v1 pith:Q2Q5IGFA submitted 2025-08-06 cs.CV cs.AI

Tianxiao Li , Zhenglin Huang , Haiquan Wen , Yiwei He , Shuchang Lyu , Baoyuan Wu , Guangliang Cheng This is my paper · ORCID
classification cs.CVcs.AI
keywords deepfakedetectionvision-languagemodelretrieval-augmentedgenerationGRPOreinforcementlearningmultimodalexplainabilitysaliencymapAI-generatedimage
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

RAIDX is a deepfake-detection framework built on a vision-language model that does two things before it answers. It retrieves the most similar labeled training images, summarizes how many are real and fake, and appends that summary to the prompt; then, after training with GRPO, it must emit a reasoning block and a final REAL/FAKE answer. The authors claim this is the first use of retrieval-augmented generation for deepfake detection and the first use of GRPO there, and that the combination reaches the highest average accuracy on the AntifakePrompt benchmark and beats SIDA-13B on SID-Set while also producing fine-grained text explanations and saliency maps with no mask or text annotations. The reason to care is that explainability in this field has meant either classification-only detectors or coarse summaries trained on expensive manual labels; RAIDX claims to get detailed justifications without those labels. The paper also states its scope limits: tampered and partially manipulated images are excluded, and its own case studies show saliency maps can drift.

What carries the argument

The load-bearing mechanism is the combination of a retrieval-augmented prompt and a GRPO-trained LoRA policy. The RAG module (a FAISS index over ViT features) turns the label distribution of similar training images into one sentence of context, giving the LLM a cheap soft prior while asking it to judge the image itself. GRPO's rule-based reward, $r_i = r_{\text{acc},i} + r_{\text{fmt},i}$, is what drives the model to produce structured chain-of-thought outputs without any human explanation labels. Saliency maps come from attention rollout, the product of per-layer attention matrices $\tilde{A} = A^{(1)} A^{(2)} \cdots A^{(L)}$, which highlights which image patches the ViT attended to; the pa

What would settle it

Take a fixed set of AI-generated images with hand-annotated forgery masks and compute the intersection-over-union (or hit rate) between RAIDX's attention-rollout saliency maps and the masks, comparing against the supervised-fine-tuning baseline. If the IoU is near zero while classification accuracy stays high, and if blind human raters cannot confirm that the artifacts named in the explanations are actually present, then the explainability claim is unsupported even though detection may be correct.

Watch

Extended reading notes

Core claim

RAIDX couples a frozen vision-language model (Qwen2.5-VL) with a ViT encoder and lightweight low-rank adapters (LoRA). A FAISS index over ViT embeddings retrieves the $k$ nearest training images for a query and inserts the count of REAL versus FAKE neighbors into the prompt. Training then uses GRPO to optimize, per sampled output, $r_i = r_{\text{acc},i} + r_{\text{fmt},i}$, where $r_{\text{acc}}$ is one if the prediction is correct and $r_{\text{fmt}}$ is one if the output follows the required format; advantages are normalized within each group. The saliency map is produced by attention rollout across ViT layers, not by a learned localizer. On SID-Set, RAIDX reports higher Accuracy and F1 t

Load-bearing premise

The load-bearing premise is that rewarding only label correctness and output formatting is enough to make the model's text explanations and saliency maps faithfully describe why an image is fake; nothing in the training reward checks explanation truthfulness or visual grounding.

Editorial extensions

If this is right

  • If the accuracy claims hold, retrieval-augmented context is a viable low-cost booster for VLM-based forgery detection, since the index is built once from training labels and adds no per-image annotation.
  • If the explanation claims hold, fine-grained rationales can be elicited without mask or text labels, removing a major bottleneck for interpretable detectors.
  • The GRPO ablation result implies that rule-based reinforcement learning, not supervised instruction tuning, is what closes most of the accuracy gap in this task, a pattern shared with recent reasoning models.
  • The reported robustness under JPEG compression, Gaussian blur, resizing, and brightness noise suggests the pipeline can hold up in realistic social-media conditions.
  • The paper's stated limitations imply the current system is a whole-image detector only: tampered and partially manipulated images are out of scope, with video and tampering listed as future extensions.

Reading between the lines

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

  • Because the reward never checks whether the explanation names real artifacts, the expert-rated explanation gain is not direct evidence of factual grounding; a dedicated test would compare RAIDX's artifact mentions against ground-truth forgery regions.
  • The RAG module feeds only label counts, not visual patches, from retrieved images, so its benefit may be a class-prior effect rather than artifact knowledge; ablating with deliberately mislabeled retrieved images would separate the two.
  • Attention rollout is not trained by the GRPO reward, so saliency drift and classification accuracy can diverge; the paper's own failure examples make this a testable concern.
  • The framework suggests a general recipe for perceptual forensics: any binary detector on a frozen VLM can be made to emit structured rationales with the same reward scheme, a concrete extension to video or tampered-image detection.
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

3 major / 3 minor

Summary. The paper introduces RAIDX, a framework for explainable deepfake image detection that combines retrieval-augmented generation (RAG) with Group Relative Policy Optimization (GRPO). A ViT encoder extracts query features, a FAISS index retrieves the top-k most similar training images, and the label distribution of those retrieved images is packaged into the prompt as “reference information.” A Qwen2.5-VL model with trainable LoRA adapters is optimized by GRPO using a reward that is the sum of a binary classification-correctness term and a format-adherence term. The authors claim state-of-the-art detection accuracy on the SID-Set benchmark and on 18 unseen-generation datasets, and further claim that GRPO enables the model to produce fine-grained textual explanations and saliency maps without manual annotations. The paper reports expert-rated explanation-quality scores, robustness experiments under image perturbations, and ablations for the RAG and GRPO components.

Significance. If the results are validated, RAIDX would be a noteworthy first combination of RAG and GRPO for deepfake detection, with the practical appeal of avoiding manual mask and text annotations. The generalization experiments across unseen generative models and the robustness evaluations are useful additions to the literature, and the authors state that code and models will be released. However, the primary detection claim is currently unverifiable because the paper's own ablation table reports a large accuracy discrepancy for the same benchmark. In addition, the explanation claim rests on a reward signal that does not reward explanation quality, and the reported human evaluation lacks a rubric and inter-rater reliability measures. These issues prevent the paper from being accepted in its present form.

major comments (3)
  1. [§4.1, Table 1; §4.5, Table 7; also Table 5] The central detection claim is internally inconsistent. Section 4.1 (Table 1) reports RAIDX Real ACC 98.5 / Fake ACC 99.4 on SID-Set, Section 4.4 (Table 5) reports SID-Set Test1 ACC 98.95, but Section 4.5 (Table 7) reports that on the SID-Set benchmark, “holding all data splits, model architecture, and hyperparameters constant,” the full RAIDX model obtains 93.07% accuracy, with Base+GRPO at 87.45%. A gap of roughly 6 percentage points between 93.07% and 98.95% is not explained by metric choice or by the “Real/Fake” versus “Test1” distinction. Either Table 1/5 or Table 7 is wrong, or the “same benchmark” actually uses a different split. Because the SOTA detection claim and the claimed +30.33 GRPO gain both depend on these numbers, the authors must reconcile them, report the exact evaluation protocol used in each table, and rerun the affected experiments.
  2. [§3.2, Eq. (2); §4.3; §4.6] The claimed explanation contribution is not supported by the optimization objective or the evaluation. The reward in Eq. (2) is r_i = r_acc + r_fmt, with no term for explanation factual grounding, artifact localization, or saliency accuracy. It is therefore unclear how GRPO can “autonomously generate fine-grained textual explanations and saliency maps” that are grounded in specific image artifacts. Section 4.3 evaluates explanations with 100 images scored by 10 experts, but no scoring rubric, no inter-rater reliability statistic, and no quantitative localization metric is reported. Section 4.6 explicitly documents “saliency map drift” and “failure explanation” cases, and the conclusion lists these as failure modes. As written, the experimental evidence is insufficient to substantiate the explainability claim; the authors should either add an explanation/saliency reward with a concrete, v
  3. [§3.1, Figure 3; §4.5, Table 6] The RAG module's contribution is potentially confounded by the label-distribution prior. The prompt states “Among the {k} reference images most similar to the current image, {real_count} are labeled as REAL, and {fake_count} are labeled as FAKE,” which is a strong supervised prior computed from training labels. The paper attributes accuracy gains to retrieval-grounded knowledge, but it is not clear whether the improvement comes from visual artifact matching or simply from the class-posterior information in the retrieved counts. The “Static Prompt” ablation in Table 6 also uses the same label-count summary, so it does not isolate the effect of dynamic retrieval versus the prior. To support the RAG claim, the authors should compare against a control that uses the same label distribution but with randomly selected reference images, or that removes the label-count summary and only provides v
minor comments (3)
  1. [Various] There are typos in the manuscript, e.g., “sc ene” in Figure 2, and “SGXL” in Table 2 where “SDXL” is likely meant. The reference list uses inconsistent formatting (“Arxiv” versus “arXiv”).
  2. [§3.2, §4.5] GRPO hyperparameters (ε, β, learning rate, group size G, number of training steps, LoRA rank) are not given, despite the statement that “hyperparameters are held constant.” Without these, the ablation is not reproducible.
  3. [§4.3] The explanation evaluation would benefit from reporting the expert scoring rubric, inter-annotator agreement (e.g., Krippendorff’s alpha), and the exact instructions given to the experts. The current one-sentence description is not sufficiently detailed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: RAIDX's RAG and GRPO contributions are not forced by construction; the reported internal inconsistency is a correctness issue, not a circularity.

full rationale

RAIDX's derivation chain is self-contained rather than circular. The RAG module (Sec. 3.1) retrieves top-k images from a FAISS index over training images and supplies the label counts of those retrievals; this is a legitimate supervised retrieval signal built from ground-truth labels, not from the model's own predictions or test labels, so the accuracy contribution is not self-referential. The GRPO reward (Eq. 2, r_i = r_acc,i + r_fmt,i) directly optimizes classification accuracy and output format; the paper's explanation-quality results are an empirical side effect rather than an equation-level identity, and the absence of an explanation-grounding reward is a supportability concern, not circularity. The SID-Set benchmark and SIDA-13B baseline originate from the authors' prior work [27], but the paper also evaluates on the external AntifakePrompt benchmark and reports higher average accuracy than external baselines (Table 2), giving the central SOTA claim independent support. The internal inconsistency between Table 1/5 (RAIDX SID-Set ~98.5-99.4) and Table 7's ablation (93.07 for full RAIDX on the same benchmark) is a serious verifiability/correctness issue, but it is not a circularity because no equation or construction forces the reported numbers to equal the training objective. Therefore score 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The main uncharged costs are the unstated retrieval count and GRPO hyperparameters, plus three behavioral assumptions about retrieval informativeness, attention faithfulness, and reward-to-explanation transfer.

free parameters (3)
  • RAG top-k retrieval count
    The prompt summarizes the labels of the k most similar training images; the paper never reports the value of k used in Tables 1 through 7.
  • GRPO and LoRA hyperparameters
    Epsilon, beta, group size G, LoRA rank, and learning rate in Eq. 4 are not reported, so the training cannot be reproduced exactly.
  • Reward weights = r_acc = 1, r_fmt = 1
    The reward in Eq. 2 weights accuracy and format equally by hand; no sensitivity analysis is provided.
assumptions (4)
  • domain assumption FAISS retrieval over ViT embeddings retrieves images whose labels are informative about the query's authenticity.
    The RAG module feeds a histogram of retrieved labels into the prompt and attributes accuracy gains to it; if neighbors are uninformative or if test or near-duplicate images enter the index, the detection claim collapses. Section 3.1.
  • domain assumption Attention rollout from the [CLS] token over ViT layers produces saliency maps that localize fake regions.
    Saliency maps are computed by Eq. 1 with no localization supervision or metric; the explainability claim rests on this assumed faithfulness. Section 3.1.
  • ad hoc to paper Binary reward r_acc + r_fmt is sufficient for the LLM to develop grounded, artifact-specific explanations.
    GRPO's only training signal is classification correctness plus output format; no mechanism or evidence connects this to truthful visual explanations. Section 3.2.
  • domain assumption SID-Set Real and Synthetic subsets are a suitable benchmark and the Tampered subset can be excluded.
    The paper excludes Tampered images by fiat and uses a benchmark created by overlapping authors; this scopes the claim to fully synthetic images. Section 4.1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RAIDX: A Retrieval-Augmented Generation and GRPO Reinforcement Learning Framework for Explainable Deepfake Detection." pith.science (2026). https://pith.science/paper/Q2Q5IGFA

@misc{pith2026250804524,
  author       = {Pith},
  title        = {Pith review of: RAIDX: A Retrieval-Augmented Generation and GRPO Reinforcement Learning Framework for Explainable Deepfake Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q2Q5IGFA}},
  note         = {Machine review of arXiv:2508.04524}
}
read the original abstract

The rapid advancement of AI-generation models has enabled the creation of hyperrealistic imagery, posing ethical risks through widespread misinformation. Current deepfake detection methods, categorized as face specific detectors or general AI-generated detectors, lack transparency by framing detection as a classification task without explaining decisions. While several LLM-based approaches offer explainability, they suffer from coarse-grained analyses and dependency on labor-intensive annotations. This paper introduces RAIDX (Retrieval-Augmented Image Deepfake Detection and Explainability), a novel deepfake detection framework integrating Retrieval-Augmented Generation (RAG) and Group Relative Policy Optimization (GRPO) to enhance detection accuracy and decision explainability. Specifically, RAIDX leverages RAG to incorporate external knowledge for improved detection accuracy and employs GRPO to autonomously generate fine-grained textual explanations and saliency maps, eliminating the need for extensive manual annotations. Experiments on multiple benchmarks demonstrate RAIDX's effectiveness in identifying real or fake, and providing interpretable rationales in both textual descriptions and saliency maps, achieving state-of-the-art detection performance while advancing transparency in deepfake identification. RAIDX represents the first unified framework to synergize RAG and GRPO, addressing critical gaps in accuracy and explainability. Our code and models will be publicly available.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

88 extracted references · 68 canonical work pages

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report.Arxiv(2023)

    2023

  2. [2]

    Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen. 2018. Mesonet: a compact facial video forgery detection network. InWIFS

    2018

  3. [3]

    Stability AI. 2022. Stable Diffusion v2.0. https://huggingface.co/stabilityai/stable- diffusion-2. Accessed: 2025-04-12

    2022

  4. [4]

    Stability AI. 2023. DeepFloyd IF. https://github.com/deep-floyd/IF. Accessed: 2025-04-12

    2023

  5. [5]

    Xiuli Bi, Bo Liu, Fan Yang, Bin Xiao, Weisheng Li, Gao Huang, and Pamela C. Cosman. 2023. Detecting Generated Images by Real Images Only.Arxiv(2023)

    2023

  6. [6]

    Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang. 2022. End-to-End Reconstruction-Classification Learning for Face Forgery Detection. InCVPR

    2022

  7. [7]

    You-Ming Chang, Chen Yeh, Wei-Chen Chiu, and Ning Yu. 2023. AntifakePrompt: Prompt-Tuned Vision-Language Models are Fake Image Detectors.Arxiv(2023)

    2023

  8. [8]

    Liang Chen, Lei Li, Haozhe Zhao, Yifan Song, and Vinci. 2025. R1-V: Reinforcing Super Generalization Ability in Vision-Language Models with Less Than $3. https://github.com/Deep-Agent/R1-V. Accessed: 2025-02-02

    2025

Show all 88 references
  1. [9]

    Liang Chen, Yong Zhang, Yibing Song, Lingqiao Liu, and Jue Wang. 2022. Self- supervised Learning of Adversarial Example: Towards Good Generalizations for Deepfake Detection.Arxiv(2022)

    2022

  2. [10]

    François Chollet. 2017. Xception: Deep Learning with Depthwise Separable Convolutions.Arxiv(2017)

    2017

  3. [11]

    Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah

  4. [12]

    Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. 2023. InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning. arXiv:2305.06500 [cs.CV] https://arxiv.org/abs/2305.06500

    2023 arXiv

  5. [13]

    Alan Dao and Dinh Bach Vu. 2025. AlphaMaze: Enhancing Large Language Models’ Spatial Intelligence via GRPO.Arxiv(2025)

    2025

  6. [14]

    DeepSeek-AI. 2024. DeepSeek-V3 Technical Report.Arxiv(2024)

    2024

  7. [15]

    DeepSeek-AI. 2025. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning.Arxiv(2025)

    2025

  8. [16]

    Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. 2020. The DeepFake Detection Challenge (DFDC) Dataset.Arxiv(2020)

    2020

  9. [17]

    Chengbo Dong, Xinru Chen, Ruohan Hu, Juan Cao, and Xirong Li. 2023. MVSS- Net: Multi-View Multi-Scale Supervised Networks for Image Manipulation De- tection.T-PAMI(2023)

    2023

  10. [18]

    Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Johnson, Gergely Szilvasy, Pierre-Emmanuel Mazaré, Maria Lomeli, Lucas Hosseini, and Hervé Jégou. 2025. The Faiss library.Arxiv(2025)

    2025

  11. [19]

    Junxian Duan, Yuang Ai, Jipeng Liu, Shenyuan Huang, Huaibo Huang, Jie Cao, and Ran He. 2024. Test-time Forgery Detection with Spatial-Frequency Prompt Learning.ICCV(2024)

    2024

  12. [20]

    Joel Frank, Thorsten Eisenhofer, Lea Schönherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. 2020. Leveraging frequency analysis for deep fake image recognition. InICML. 3247–3258

    2020

  13. [21]

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Ji- awei Sun, Meng Wang, and Haofen Wang. 2024. Retrieval-Augmented Generation for Large Language Models: A Survey.Arxiv(2024)

    2024

  14. [22]

    Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde- Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio

    Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde- Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative Adversarial Networks.Arxiv(2014)

    2014

  15. [23]

    Xiao Guo, Xiaohong Liu, Zhiyuan Ren, Steven Grosz, Iacopo Masi, and Xiaoming Liu. 2023. Hierarchical Fine-Grained Image Forgery Detection and Localization. InCVPR

    2023

  16. [24]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2015. Deep Residual Learning for Image Recognition.Arxiv(2015)

    2015

  17. [25]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-Rank Adaptation of Large Language Models. InICLR

    2022

  18. [26]

    Wenxuan Huang, Bohan Jia, Zijie Zhai, Shaosheng Cao, Zheyu Ye, Fei Zhao, Zhe Xu, Yao Hu, and Shaohui Lin. 2025. Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models.Arxiv(2025)

    2025

  19. [27]

    Zhenglin Huang, Jinwei Hu, Xiangtai Li, Yiwei He, Xingyu Zhao, Bei Peng, Baoyuan Wu, Xiaowei Huang, and Guangliang Cheng. 2025. SIDA: Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model. InCVPR

    2025

  20. [28]

    Zhenglin Huang, Tianxiao Li, Xiangtai Li, Haiquan Wen, Yiwei He, Jiangning Zhang, Hao Fei, Xi Yang, Xiaowei Huang, Bei Peng, and Guangliang Cheng. 2025. So-Fake: Benchmarking and Explaining Social Media Image Forgery Detection. Arxiv(2025)

    2025

  21. [29]

    Yan Ju, Shan Jia, Lipeng Ke, Hongfei Xue, Koki Nagano, and Siwei Lyu. 2022. Fusing Global and Local Features for Generalized AI-Synthesized Image Detection. InICIP

    2022

  22. [30]

    Hengrui Kang, Siwei Wen, Zichen Wen, Junyan Ye, Weijia Li, Peilin Feng, Baichuan Zhou, Bin Wang, Dahua Lin, Linfeng Zhang, and Conghui He. 2025. LEGION: Learning to Ground and Explain for Synthetic Image Detection.Arxiv (2025)

    2025

  23. [31]

    Tero Karras, Samuli Laine, and Timo Aila. 2021. A Style-Based Generator Archi- tecture for Generative Adversarial Networks.PAMI(2021)

    2021

  24. [32]

    Rohit Kundu, Hao Xiong, Vishal Mohanty, Athula Balachandran, and Amit K Roy-Chowdhury. 2024. Towards a Universal Synthetic Video Detector: From Face or Background Manipulations to Fully AI-Generated Content.Arxiv(2024)

    2024

  25. [33]

    Le and Simon S

    Binh M. Le and Simon S. Woo. 2023. Quality-Agnostic Deepfake Detection with Intra-model Collaborative Learning.Arxiv(2023)

    2023

  26. [34]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2021. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.Arxiv(2021)

    2021

  27. [35]

    Patrick S. H. Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rock- täschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-Augmented Genera- tion for Knowledge-Intensive NLP Tasks. InNeurIPS

    2020

  28. [36]

    Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi

  29. [37]

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. InICML

    2022

  30. [38]

    Jingchun Lian, Lingyu Liu, Yaxiong Wang, Yujiao Wu, Li Zhu, and Zhedong Zheng. 2024. A Large-scale Interpretable Multi-modality Benchmark for Facial Image Forgery Localization.Arxiv(2024)

    2024

  31. [39]

    Lawrence Zitnick, and Piotr Dollár

    Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dollár

  32. [40]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual instruc- tion tuning.NeurIPS(2023)

    2023

  33. [41]

    Huan Liu, Zichang Tan, Chuangchuang Tan, Yunchao Wei, Yao Zhao, and Jing- dong Wang. 2023. Forgery-aware Adaptive Transformer for Generalizable Syn- thetic Image Detection.Arxiv(2023)

    2023

  34. [42]

    Jiawei Liu, Fanrui Zhang, Jiaying Zhu, Esther Sun, Qiang Zhang, and Zheng-Jun Zha. 2025. ForgeryGPT: Multimodal Large Language Model For Explainable Image Forgery Detection and Localization.Arxiv(2025)

    2025

  35. [43]

    Xiaohong Liu, Yaojie Liu, Jun Chen, and Xiaoming Liu. 2022. PSCC-Net: Progres- sive Spatio-Channel Correlation Network for Image Manipulation Detection and Localization.T-CSVT(2022)

    2022

  36. [44]

    Zhengzhe Liu, Xiaojuan Qi, and Philip H. S. Torr. 2020. Global Texture Enhance- ment for Fake Face Detection in the Wild. InCVPR

    2020

  37. [45]

    Yongdong Luo, Xiawu Zheng, Xiao Yang, Guilin Li, Haojia Lin, Jinfa Huang, Jiayi Ji, Fei Chao, Jiebo Luo, and Rongrong Ji. 2024. Video-RAG: Visually-aligned Retrieval-Augmented Long Video Comprehension.ArXiv(2024)

    2024

  38. [46]

    Midjourney

    Inc. Midjourney. 2023. Midjourney v5. https://www.midjourney.com/. Accessed: 2025-04-12

    2023

  39. [47]

    Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. 2022. GLIDE: Towards Photoreal- istic Image Generation and Editing with Text-Guided Diffusion Models.Arxiv (2022)

    2022

  40. [48]

    Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. 2023. Towards Universal Fake Image Detectors that Generalize Across Generative Models. InCVPR

    2023

  41. [49]

    OpenAI. 2023. DALL·E 3. https://openai.com/dall-e. Accessed: 2025-04-12

    2023

  42. [50]

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. 2023. SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis.Arxiv(2023)

    2023

  43. [51]

    Jingyuan Qi, Zhiyang Xu, Rulin Shao, Yang Chen, Di Jin, Yu Cheng, Qifan Wang, and Lifu Huang. 2024. RoRA-VLM: Robust Retrieval-Augmented Vision Language Models.ArXiv(2024)

    2024

  44. [52]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. InICML

    2021

  45. [53]

    Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al . 2018. Improving language understanding by generative pre-training

    2018

  46. [54]

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen

  47. [55]

    Jonas Ricker, Simon Damm, Thorsten Holz, and Asja Fischer. 2024. Towards the Detection of Diffusion Model Deepfakes.Arxiv(2024)

    2024

  48. [56]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-Resolution Image Synthesis with Latent Diffusion Models. InCVPR

    2022

  49. [57]

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov

  50. [58]

    Zeyang Sha, Zheng Li, Ning Yu, and Yang Zhang. 2023. DE-FAKE: Detection and Attribution of Fake Images Generated by Text-to-Image Generation Models. Arxiv(2023)

    2023

  51. [59]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models.Arxiv(2024)

    2024

  52. [60]

    Haozhan Shen, Zilun Zhang, Kangjia Zhao, Qianqian Zhang, Ruochen Xu, and Tiancheng Zhao. 2025. VLM-R1: A stable and generalizable R1-style Large Vision- Language Model. https://github.com/om-ai-lab/VLM-R1. Accessed: 2025-02-15

    2025

  53. [61]

    Ensheng Shi, Yanlin Wang, Wei Tao, Lun Du, Hongyu Zhang, Shi Han, Dongmei Zhang, and Hongbin Sun. 2022. RACE: Retrieval-Augmented Commit Message Generation.Arxiv(2022)

    2022

  54. [62]

    Xincheng Shuai, Henghui Ding, Xingjun Ma, Rongcheng Tu, Yu-Gang Jiang, and Dacheng Tao. 2024. A Survey of Multimodal-Guided Image Editing with Text-to-Image Diffusion Models.Arxiv(2024)

    2024

  55. [63]

    Zhihao Sun, Haoran Jiang, Haoran Chen, Yixin Cao, Xipeng Qiu, Zuxuan Wu, and Yu-Gang Jiang. 2024. ForgerySleuth: Empowering Multimodal Large Language Models for Image Manipulation Detection.Arxiv(2024)

    2024

  56. [64]

    Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. 2024. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. InCVPR

    2024

  57. [65]

    Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, and Yunchao Wei. 2023. Learning on gradients: Generalized artifacts representation for gan-generated images detection. InCVPR

    2023

  58. [66]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lam- ple. 2023. LLaMA: Open and Efficient Foundation ...

    2023

  59. [67]

    Chengrui Wang and Weihong Deng. 2021. Representative Forgery Mining for Fake Face Detection.Arxiv(2021)

    2021

  60. [68]

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. 2024. Qwen2-VL: Enhancing Vision-Language Mode...

    2024

  61. [69]

    Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. 2020. CNN-generated images are surprisingly easy to spot... for now. In CVPR

    2020

  62. [70]

    Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A. Efros. 2020. CNN-generated images are surprisingly easy to spot... for now. arXiv:1912.11035 [cs.CV] https://arxiv.org/abs/1912.11035

    2020 arXiv

  63. [71]

    Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. 2024. CogVLM: Visual Expert for Pretrained Language Models.Arxiv(2024)

    2024

  64. [72]

    Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. 2023. DIRE for Diffusion-Generated Image Detection. Arxiv(2023)

    2023

  65. [73]

    Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau

    Zijie J. Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. 2023. DiffusionDB: A Large-scale Prompt Gallery Dataset for Text-to-Image Generative Models.Arxiv(2023)

    2023

  66. [74]

    Haiquan Wen, Yiwei He, Zhenglin Huang, Tianxiao Li, Zihan Yu, Xingru Huang, Lu Qi, Baoyuan Wu, Xiangtai Li, and Guangliang Cheng. 2025. BusterX: MLLM- Powered AI-Generated Video Forgery Detection and Explanation.Arxiv(2025)

    2025

  67. [75]

    Haiquan Wen, Tianxiao Li, Zhenglin Huang, Yiwei He, and Guangliang Cheng

  68. [76]

    Haiwei Wu, Jiantao Zhou, and Shile Zhang. 2023. Generalizable Synthetic Image Detection via Language-guided Contrastive Learning.Arxiv(2023)

    2023

  69. [77]

    Jin Xu, Zhifang Guo, Jinzheng He, Hangrui Hu, Ting He, Shuai Bai, Keqin Chen, Jialin Wang, Yang Fan, Kai Dang, Bin Zhang, Xiong Wang, Yunfei Chu, and Junyang Lin. 2025. Qwen2.5-Omni Technical Report.Arxiv(2025)

    2025

  70. [78]

    Zhiyuan Yan, Taiping Yao, Shen Chen, Yandan Zhao, Xinghe Fu, Junwei Zhu, Donghao Luo, Chengjie Wang, Shouhong Ding, Yunsheng Wu, and Li Yuan. 2024. DF40: Toward Next-Generation Deepfake Detection. InNeurIPS

    2024

  71. [79]

    Hockenmaier

    Peter Young, Alice Lai, Micah Hodosh, and J. Hockenmaier. 2014. From image descriptions to visual denotations: New similarity metrics for semantic infer- ence over event descriptions.Transactions of the Association for Computational Linguistics2 (2014), 67–78

    2014

  72. [80]

    Fangneng Zhan, Yingchen Yu, Rongliang Wu, Jiahui Zhang, Shijian Lu, Lingjie Liu, Adam Kortylewski, Christian Theobalt, and Eric P. Xing. 2023. Multimodal Image Synthesis and Editing: The Generative AI Era.PAMI(2023)

    2023

  73. [81]

    Fengji Zhang, Bei Chen, Yue Zhang, Jacky Keung, Jin Liu, Daoguang Zan, Yi Mao, Jian-Guang Lou, and Weizhu Chen. 2023. RepoCoder: Repository-Level Code Completion Through Iterative Retrieval and Generation.Arxiv(2023)

    2023

  74. [82]

    Yinglin Zheng, Jianmin Bao, Dong Chen, Ming Zeng, and Fang Wen. 2021. Ex- ploring temporal coherence for more general video face forgery detection. In CVPR

    2021

  75. [2015]

    Microsoft COCO: Common Objects in Context.Arxiv(2015)

    2015

  76. [2017]

    Proximal policy optimization algorithms.Arxiv(2017)

    2017

  77. [2022]

    MM ’25, October 27–31, 2025, Dublin, Ireland Li et al

    Hierarchical Text-Conditional Image Generation with CLIP Latents.Arxiv (2022). MM ’25, October 27–31, 2025, Dublin, Ireland Li et al

    2022

  78. [2023]

    Diffusion Models in Vision: A Survey.PAMI(2023)

    2023

  79. [2024]

    Playground v2.5: Three Insights towards Enhancing Aesthetic Quality in Text-to-Image Generation.Arxiv(2024)

    2024

  80. [2025]

    BusterX++: Towards Unified Cross-Modal AI-Generated Content Detection and Explanation with MLLM.Arxiv(2025)

    2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.