pith. sign in

arxiv: 2511.00181 · v2 · submitted 2025-10-31 · 💻 cs.CV · cs.CR

From Evidence to Verdict: An Agent-Based Forensic Framework for AI-Generated Image Detection

Mengfei Liang , Yiting Qu , Yukun Jiang , Michael Backes , Yang Zhang This is my paper

Pith reviewed 2026-05-18 02:01 UTC · model grok-4.3

classification 💻 cs.CV cs.CR
keywords AI image detectionmulti-agent collaborationforensic frameworktraining-freevision language modelsmedia forensicsgenerative AI
0
0 comments X

The pith

A multi-agent forensic framework detects AI-generated images by debating evidence from multiple tools without any model training.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper presents AIFo, a framework that turns AI-generated image detection into a collaborative forensic investigation. Multiple agents examine different types of evidence such as reverse image searches, metadata, pre-trained classifiers, and vision-language model outputs. A debate mechanism helps resolve cases where the evidence is incomplete or contradictory. The system also includes an optional memory component to draw on past detections. Evaluations on thousands of images from lab and real-world sources show it reaching over 97 percent accuracy while beating standard detection methods.

Core claim

AIFo formulates AI-generated image detection as a multi-stage forensic analysis process through multi-agent collaboration that integrates reverse image search, metadata extraction, pre-trained classifiers, and vision-language model analysis, resolving insufficient or conflicting evidence through a structured multi-agent debate mechanism, achieving 97.05% accuracy on a 6,000-image benchmark spanning controlled and real-world scenarios.

What carries the argument

The multi-agent debate mechanism, which structures collaboration among agents to reconcile conflicting forensic evidence from diverse tools.

Load-bearing premise

That the structured debate among agents can consistently produce accurate verdicts even when the individual forensic tools provide insufficient or conflicting information.

What would settle it

Observing a set of AI-generated images where the agents reach an incorrect verdict due to unresolved conflicts in the evidence from the tools.

Figures

Figures reproduced from arXiv: 2511.00181 by Mengfei Liang, Michael Backes, Yang Zhang, Yiting Qu, Yukun Jiang.

Figure 1
Figure 1. Figure 1: High-level overview of our proposed AIFo. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Examples of our agent framework’s decision-making process, demonstrating diverse evidence integration across different [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Analysis of individual tool contributions to the agent [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Performance degradation when each tool is disabled from the framework. [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Overview of the memory-augmented reasoning mod [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Example of memory-augmented reasoning showing [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
read the original abstract

The rapid evolution of AI-generated images poses growing challenges to information integrity and media authenticity. Existing detection approaches face limitations in robustness, interpretability, and generalization across diverse generative models, particularly when relying on a single source of visual evidence. We introduce AIFo (Agent-based Image Forensics), a training-free framework that formulates AI-generated image detection as a multi-stage forensic analysis process through multi-agent collaboration. The framework integrates a set of forensic tools, including reverse image search, metadata extraction, pre-trained classifiers, and vision-language model analysis, and resolves insufficient or conflicting evidence through a structured multi-agent debate mechanism. An optional memory-augmented module further enables the framework to incorporate information from historical cases. We evaluate AIFo on a benchmark of 6,000 images spanning controlled laboratory settings and challenging real-world scenarios, where it achieves 97.05% accuracy and consistently outperforms traditional classifiers and strong vision-language model baselines. These findings demonstrate the effectiveness of agent-based procedural reasoning for AI-generated image detection.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces AIFo, a training-free agent-based framework for detecting AI-generated images. It integrates forensic tools including reverse image search, metadata extraction, pre-trained classifiers, and vision-language models, then employs a structured multi-agent debate to resolve insufficient or conflicting evidence, with an optional memory-augmented module for historical cases. The framework is evaluated on a 6,000-image benchmark spanning laboratory and real-world scenarios, reporting 97.05% accuracy and consistent outperformance over traditional classifiers and strong VLM baselines.

Significance. If the multi-agent debate reliably converts tool outputs into correct verdicts, the work offers a promising direction for interpretable, generalizable, and training-free detection methods that combine multiple evidence sources. The explicit avoidance of fitted parameters and reliance on external pre-trained tools is a clear strength that supports broader applicability across generative models.

major comments (2)
  1. [§5] §5 (Evaluation): The central claim of 97.05% accuracy and outperformance is reported without error bars, confidence intervals, or statistical significance tests against baselines. This omission makes it impossible to determine whether the gains are robust or could be explained by variance in the 6,000-image benchmark.
  2. [§3.2] §3.2 (Multi-agent Debate Mechanism): The robustness argument rests on the debate resolving conflicts among reverse search, metadata, classifiers, and VLMs, yet no ablation studies (e.g., replacing debate with majority vote or single-VLM decision) or quantitative metrics on conflict frequency and resolution success are provided. Without these, the contribution of the agent component remains unproven and is the weakest link in the generalization claim.
minor comments (2)
  1. [Abstract] Abstract and §4: The benchmark is described as covering 'controlled laboratory settings and challenging real-world scenarios' but lacks explicit enumeration of the generative models, image sources, or construction protocol, which would aid reproducibility.
  2. [Figure 1] Figure 1 (Framework Overview): The diagram would benefit from clearer labeling of the debate stages and information flow between agents to improve readability for readers unfamiliar with multi-agent systems.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive feedback on our manuscript. The comments highlight important aspects regarding the statistical rigor of our evaluation and the validation of the multi-agent component. We provide detailed responses to each major comment and indicate the revisions planned for the next version of the paper.

read point-by-point responses

  1. Referee: [§5] §5 (Evaluation): The central claim of 97.05% accuracy and outperformance is reported without error bars, confidence intervals, or statistical significance tests against baselines. This omission makes it impossible to determine whether the gains are robust or could be explained by variance in the 6,000-image benchmark.

    Authors: We agree that the current presentation lacks error bars, confidence intervals, and statistical significance tests, which weakens the ability to assess robustness. In the revised manuscript we will add 95% bootstrap confidence intervals for all reported accuracies on the 6,000-image benchmark and will include paired statistical tests (McNemar’s test) comparing AIFo to each baseline. These results will appear in Section 5 together with the existing accuracy figures. revision: yes

  2. Referee: [§3.2] §3.2 (Multi-agent Debate Mechanism): The robustness argument rests on the debate resolving conflicts among reverse search, metadata, classifiers, and VLMs, yet no ablation studies (e.g., replacing debate with majority vote or single-VLM decision) or quantitative metrics on conflict frequency and resolution success are provided. Without these, the contribution of the agent component remains unproven and is the weakest link in the generalization claim.

    Authors: We acknowledge the value of explicit ablations to isolate the contribution of the multi-agent debate. We will add experiments that replace the debate with majority voting and with single-VLM decisions, and we will report quantitative metrics on conflict frequency across evidence sources and on the fraction of conflicts successfully resolved by the debate. These new results and analysis will be placed in Section 3.2 and in an expanded evaluation subsection. revision: yes

Circularity Check

0 steps flagged

No significant circularity; framework relies on external tools and rules

full rationale

The paper presents AIFo as a training-free framework that integrates pre-existing forensic tools (reverse search, metadata, classifiers, VLMs) and applies a structured multi-agent debate to resolve conflicts. No equations, fitted parameters, or self-referential definitions appear in the description. Performance claims rest on evaluation against an external 6000-image benchmark rather than any internal reduction of outputs to inputs by construction. No load-bearing self-citations, uniqueness theorems, or ansatzes are invoked that would make the central claims tautological. The method is self-contained against external components and benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the effectiveness of agent collaboration and the assumption that complementary forensic signals exist and can be reconciled through debate; no free parameters or new physical entities are introduced.

axioms (1)
  • domain assumption Forensic tools (reverse search, metadata, classifiers, VLMs) produce complementary or reconcilable evidence that multi-agent debate can integrate into reliable verdicts.
    Invoked to justify the multi-stage process and claimed robustness across generative models.

pith-pipeline@v0.9.0 · 5713 in / 1296 out tokens · 34870 ms · 2026-05-18T02:01:44.593597+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 2 Pith papers

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

  1. Authenticated Contradictions from Desynchronized Provenance and Watermarking

    cs.CR 2026-03 unverdicted novelty 6.0

    C2PA manifests and AI watermarks can independently validate contradictory claims on the same asset, and a cross-layer audit protocol resolves this with 100% accuracy on 3500 images.

  2. HEDGE: Heterogeneous Ensemble for Detection of AI-GEnerated Images in the Wild

    cs.CV 2026-04 unverdicted novelty 4.0

    HEDGE is a heterogeneous ensemble using progressive DINOv3 training, multi-scale features, and MetaCLIP2 diversity with dual-gating fusion to achieve robust AI-generated image detection and 4th place in the NTIRE 2026...

Reference graph

Works this paper leans on

72 extracted references · 72 canonical work pages · cited by 2 Pith papers · 4 internal anchors

  1. [1]

    AI Image Detector.https://huggingface.co/ haywoodsloan/ai-image-detector-deploy. 5

    work page

  2. [2]

    Anime Image Detector.https://huggingface.co/ legekka/AI-Anime-Image-Detector-ViT. 6

    work page

  3. [3]

    6, 7, 20

    Civitai.https://civitai.com. 6, 7, 20

    work page

  4. [4]

    Flickr.https://flickr.com. 6, 7

    work page

  5. [5]

    Google CLoud Vision.https://cloud.google.com/ vision?hl=en. 5

    work page

  6. [6]

    2, 6, 8, 20

    GPT-4.1.https://openai.com/index/gpt-4-1/. 2, 6, 8, 20

    work page

  7. [7]

    6, 7, 20

    Lexica.https://lexica.art/. 6, 7, 20

    work page

  8. [8]

    6, 7, 20

    NightCafe.https://creator.nightcafe.studio/. 6, 7, 20

    work page

  9. [9]

    co/NYUAD-ComNets/NYUAD_AI-generated_images_ detector

    NYUAD AI Image Detector.https://huggingface. co/NYUAD-ComNets/NYUAD_AI-generated_images_ detector. 6

    work page

  10. [10]

    SDXL-Detector.https://huggingface.co/Organika/ sdxl-detector. 6

    work page

  11. [11]

    SMOGY AI Image Detector.https://huggingface.co/ Smogy/SMOGY-Ai-images-detector. 6

    work page

  12. [12]

    Wikimedia Commons.https://commons.wikimedia.org. 6, 7

    work page

  13. [13]

    As Social Media Guardrails Fade and AI Deepfakes Go Mainstream, Experts Warn of Impact on Elec- tions.https://apnews.com/article/election- 2024-misinformation-ai-social-media-trump- 6119ee6f498db10603b3664e9ad3e87e, 2023. 1

    work page 2024

  14. [14]

    NTIRE 2017 Challenge on Single Image Super-Resolution: Dataset and Study

    Eirikur Agustsson and Radu Timofte. NTIRE 2017 Challenge on Single Image Super-Resolution: Dataset and Study. In IEEE Conference on Computer Vision and Pattern Recogni- tion Workshops (CVPRW), pages 1122–1131. IEEE, 2017. 6, 7

    work page 2017

  15. [15]

    Xiuli Bi, Bo Liu, Fan Yang, Bin Xiao, Weisheng Li, Gao Huang, and Pamela C. Cosman. Detecting Generated Images by Real Images Only.CoRR abs/2311.00962, 2023. 1

    work page arXiv 2023

  16. [16]

    Antifakeprompt: Prompt-tuned vision-language models are fake image detectors.arXiv preprint arXiv:2310.17419,

    You-Ming Chang, Chen Yeh, Wei-Chen Chiu, and Ning Yu. AntifakePrompt: Prompt-Tuned Vision-Language Models are Fake Image Detectors.CoRR abs/2310.17419, 2023. 13

    work page arXiv 2023

  17. [17]

    Zero-Shot Detection of AI-Generated Im- ages

    Davide Cozzolino, Giovanni Poggi, Matthias Nießner, and Luisa Verdoliva. Zero-Shot Detection of AI-Generated Im- ages. InEuropean Conference on Computer Vision (ECCV), pages 54–72. Springer, 2024. 1

    work page 2024

  18. [18]

    ImageNet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. InIEEE Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 248–255. IEEE, 2009. 6, 7

    work page 2009

  19. [19]

    Tenen- baum, and Igor Mordatch

    Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenen- baum, and Igor Mordatch. Improving Factuality and Reason- ing in Language Models through Multi-Agent Debate. InIn- ternational Conference on Machine Learning (ICML). JMLR,

    work page

  20. [20]

    Theory is All You Need: AI, Human Cognition, and Causal Reasoning.Strategy Sci- ence, 2024

    Teppo Felin and Matthias Holweg. Theory is All You Need: AI, Human Cognition, and Causal Reasoning.Strategy Sci- ence, 2024. 3 14

    work page 2024

  21. [21]

    Denoising Diffu- sion Probabilistic Models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffu- sion Probabilistic Models. InAnnual Conference on Neural Information Processing Systems (NeurIPS). NeurIPS, 2020. 1

    work page 2020

  22. [22]

    MetaGPT: Meta Programming for A Multi-Agent Collabo- rative Framework

    Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, Chenyu Ran, Lingfeng Xiao, Chenglin Wu, and Jürgen Schmidhuber. MetaGPT: Meta Programming for A Multi-Agent Collabo- rative Framework. InInternational Conference on Learning Representations (ICLR). JMLR, 2024. 13

    work page 2024

  23. [23]

    Tanzib Hosain, Salman Rahman, Md

    Md. Tanzib Hosain, Salman Rahman, Md. Kishor Morol, and Md. Rizwan Parvez. Xolver: Multi-Agent Reasoning with Holistic Experience Learning Just Like an Olympiad Team. CoRR abs/2506.14234, 2025. 11

    work page arXiv 2025

  24. [24]

    Holopix50k: A Large-Scale In-the-wild Stereo Image Dataset

    Yiwen Hua, Puneet Kohli, Pritish Uplavikar, Anand Ravi, Saravana Gunaseelan, Jason Orozco, and Edward Li. Holopix50k: A Large-Scale In-the-wild Stereo Image Dataset. CoRR abs/2003.11172, 2020. 6, 7

    work page arXiv 2003

  25. [25]

    Unmasking Digital Falsehoods: A Comparative Analysis of LLM-Based Misinformation Detection Strategies.CoRR abs/2503.00724, 2025

    Tianyi Huang, Jingyuan Yi, Peiyang Yu, and Xiaochuan Xu. Unmasking Digital Falsehoods: A Comparative Analysis of LLM-Based Misinformation Detection Strategies.CoRR abs/2503.00724, 2025. 13

    work page arXiv 2025

  26. [26]

    Hengrui Kang, Siwei Wen, Zichen Wen, Junyan Ye, Weijia Li, Peilin Feng, Baichuan Zhou, Bin Wang, Dahua Lin, Linfeng Zhang, et al

    Yikun Ji, Yan Hong, Jiahui Zhan, Haoxing Chen, Jun Lan, Huijia Zhu, Weiqiang Wang, Liqing Zhang, and Jianfu Zhang. Towards Explainable Fake Image Detection with Multi-Modal Large Language Models.CoRR abs/2504.14245, 2025. 1, 13

    work page arXiv 2025

  27. [27]

    Can ChatGPT Detect DeepFakes? A Study of Using Multimodal Large Language Models for Media Forensics

    Shan Jia, Reilin Lyu, Kangran Zhao, Yize Chen, Zhiyuan Yan, Yan Ju, Chuanbo Hu, Xin Li, Baoyuan Wu, and Siwei Lyu. Can ChatGPT Detect DeepFakes? A Study of Using Multimodal Large Language Models for Media Forensics. In IEEE Conference on Computer Vision and Pattern Recogni- tion Workshops (CVPRW), pages 4324–4333. IEEE, 2024. 13

    work page 2024

  28. [28]

    Analyzing and Improving the Image Quality of StyleGAN

    Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and Improving the Image Quality of StyleGAN. InIEEE Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 8107–

    work page

  29. [29]

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. BLIP: Bootstrapping Language-Image Pre-training for Uni- fied Vision-Language Understanding and Generation.CoRR abs/2201.12086, 2022. 6

    work page arXiv 2022

  30. [30]

    Malthouse

    Xinyi Li, Yongfeng Zhang, and Edward C. Malthouse. Large Language Model Agent for Fake News Detection.CoRR abs/2405.01593, 2024. 13

    work page arXiv 2024

  31. [31]

    Fakebench: Probing ex- plainable fake image detection via large multimodal models.arXiv preprint arXiv:2404.13306, 2024

    Yixuan Li, Xuelin Liu, Xiaoyang Wang, Bu Sung Lee, Shiqi Wang, Anderson Rocha, and Weisi Lin. FakeBench: Prob- ing Explainable Fake Image Detection via Large Multimodal Models.CoRR abs/2404.13306, 2024. 1, 2, 6, 7, 13, 20

    work page arXiv 2024

  32. [32]

    Improving Multi- Agent Debate with Sparse Communication Topology

    Yunxuan Li, Yibing Du, Jiageng Zhang, Le Hou, Peter Grabowski, Yeqing Li, and Eugene Ie. Improving Multi- Agent Debate with Sparse Communication Topology. InCon- ference on Empirical Methods in Natural Language Process- ing (EMNLP), pages 7281–7294. ACL, 2024. 2

    work page 2024

  33. [33]

    Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate

    Tian Liang, Zhiwei He, Wenxiang Jiao, Xing Wang, Yan Wang, Rui Wang, Yujiu Yang, Shuming Shi, and Zhaopeng Tu. Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate. InConference on Em- pirical Methods in Natural Language Processing (EMNLP), pages 17889–17904. ACL, 2024. 2

    work page 2024

  34. [34]

    MIND: A Multi-agent Framework for Zero-shot Harmful Meme Detection

    Ziyan Liu, Chunxiao Fan, Haoran Lou, Yuexin Wu, and Kai- wei Deng. MIND: A Multi-agent Framework for Zero-shot Harmful Meme Detection. InAnnual Meeting of the Asso- ciation for Computational Linguistics (ACL), pages 923–947. ACL, 2025. 2, 13

    work page 2025

  35. [35]

    GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models

    Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards Photorealistic Image Genera- tion and Editing with Text-Guided Diffusion Models.CoRR abs/2112.10741, 2021. 1

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  36. [36]

    Large Language Models and Cog- nitive Science: A Comprehensive Review of Similarities, Dif- ferences, and Challenges.CoRR abs/2409.02387, 2024

    Qian Niu, Junyu Liu, Ziqian Bi, Pohsun Feng, Benji Peng, Keyu Chen, and Ming Li. Large Language Models and Cog- nitive Science: A Comprehensive Review of Similarities, Dif- ferences, and Challenges.CoRR abs/2409.02387, 2024. 3

    work page arXiv 2024

  37. [37]

    GPT-4o.https://openai.com/index/hello- gpt-4o/

    OpenAI. GPT-4o.https://openai.com/index/hello- gpt-4o/. 2, 8, 20

    work page

  38. [38]

    Plummer, Liwei Wang, Chris M

    Bryan A. Plummer, Liwei Wang, Chris M. Cervantes, Juan C. Caicedo, Julia Hockenmaier, and Svetlana Lazeb- nik. Flickr30k Entities: Collecting Region-to-Phrase Corre- spondences for Richer Image-to-Sentence Models. InIEEE International Conference on Computer Vision (ICCV), pages 2641–2649. IEEE, 2015. 2, 6, 7

    work page 2015

  39. [39]

    SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rom- bach. SDXL: Improving Latent Diffusion Models for High- Resolution Image Synthesis.CoRR abs/2307.01952, 2023. 6

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  40. [40]

    Learning Transferable Visual Models From Natural Language Supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. InInternational Conference on Machine Learning (ICML), pages 8748–8763. PMLR, 2021. 11, 13

    work page 2021

  41. [41]

    Hierarchical Text-Conditional Image Generation with CLIP Latents

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical Text-Conditional Image Gen- eration with CLIP Latents.CoRR abs/2204.06125, 2022. 1, 6

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  42. [42]

    High-Resolution Image Syn- thesis with Latent Diffusion Models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Syn- thesis with Latent Diffusion Models. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684–10695. IEEE, 2022. 1, 6, 7

    work page 2022

  43. [43]

    Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding.CoRR abs/2205.11487, 2022. 1

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  44. [44]

    DE-FAKE: Detection and Attribution of Fake Images Generated by Text- to-Image Diffusion Models.CoRR abs/2210.06998, 2022

    Zeyang Sha, Zheng Li, Ning Yu, and Yang Zhang. DE-FAKE: Detection and Attribution of Fake Images Generated by Text- to-Image Diffusion Models.CoRR abs/2210.06998, 2022. 1, 2, 7, 8, 13, 20

    work page arXiv 2022

  45. [45]

    ZeroFake: Zero-Shot Detection of Fake Images Generated and Edited by Text-to-Image Generation Models

    Zeyang Sha, Yicong Tan, Mingjie Li, Michael Backes, and Yang Zhang. ZeroFake: Zero-Shot Detection of Fake Images Generated and Edited by Text-to-Image Generation Models. InACM SIGSAC Conference on Computer and Communica- tions Security (CCS), pages 4852–4866. ACM, 2024. 13

    work page 2024

  46. [46]

    Shoaib, Zefan Wang, Milad Taleby Ahvanooey, and Jun Zhao

    Mohamed R. Shoaib, Zefan Wang, Milad Taleby Ahvanooey, and Jun Zhao. Deepfakes, Misinformation, and Disinforma- tion in the Era of Frontier AI, Generative AI, and Large AI Models. InInternational Conference on Computer and Appli- cations (ICCA), pages 1–7. IEEE, 2023. 1 15

    work page 2023

  47. [47]

    Smit, Nathan Grinsztajn, Paul Duckworth, Thomas D

    Andries P. Smit, Nathan Grinsztajn, Paul Duckworth, Thomas D. Barrett, and Arnu Pretorius. Should we be going MAD? A Look at Multi-Agent Debate Strategies for LLMs. InInternational Conference on Machine Learning (ICML). JMLR, 2024. 2

    work page 2024

  48. [48]

    Media Forensics and DeepFakes: An Overview.Journal of Selected Topics in Signal Processing,

    Luisa Verdoliva. Media Forensics and DeepFakes: An Overview.Journal of Selected Topics in Signal Processing,

    work page

  49. [49]

    DFBench: Benchmarking Deepfake Image Detection Capability of Large Multimodal Models.CoRR abs/2506.03007, 2025

    Jiarui Wang, Huiyu Duan, Juntong Wang, Ziheng Jia, Woo Yi Yang, Xiaorong Zhu, Yu Zhao, Jiaying Qian, Yuke Xing, Guangtao Zhai, and Xiongkuo Min. DFBench: Benchmarking Deepfake Image Detection Capability of Large Multimodal Models.CoRR abs/2506.03007, 2025. 1, 13

    work page arXiv 2025

  50. [50]

    Meng, Zibin Zheng, Liang Chen, and Bingzhe Wu

    Qichao Wang, Tian Bian, Yian Yin, Tingyang Xu, Hong Cheng, Helen M. Meng, Zibin Zheng, Liang Chen, and Bingzhe Wu. Language Agents for Detecting Implicit Stereotypes in Text-to-image Models at Scale.CoRR abs/2310.11778, 2023. 13

    work page arXiv 2023

  51. [51]

    Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A. Efros. CNN-Generated Images Are Surprisingly Easy to Spot... for Now. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8692–8701. IEEE, 2020. 1, 2, 7, 8, 13, 20

    work page 2020

  52. [52]

    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. InIEEE International Conference on Computer Vision (ICCV), pages 22388–22398. IEEE, 2023. 1, 13

    work page 2023

  53. [53]

    Reverse Im- age Retrieval Cues Parametric Memory in Multimodal LLMs

    Jialiang Xu, Michael Moor, and Jure Leskovec. Reverse Im- age Retrieval Cues Parametric Memory in Multimodal LLMs. CoRR abs/2405.18740, 2024. 5

    work page arXiv 2024

  54. [54]

    Narasimhan, and Yuan Cao

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R. Narasimhan, and Yuan Cao. ReAct: Synergizing Reasoning and Acting in Language Models. InInternational Conference on Learning Representations (ICLR). ICLR, 2023. 13

    work page 2023

  55. [55]

    Unlocking the Capabilities of Vision- Language Models for Generalizable and Explainable Deep- fake Detection.CoRR abs/2503.14853, 2025

    Peipeng Yu, Jianwei Fei, Hui Gao, Xuan Feng, Zhihua Xia, and Chip-Hong Chang. Unlocking the Capabilities of Vision- Language Models for Generalizable and Explainable Deep- fake Detection.CoRR abs/2503.14853, 2025. 1, 13

    work page arXiv 2025

  56. [56]

    Autodefense: Multi-agent LLM defense against jailbreak attacks,

    Yifan Zeng, Yiran Wu, Xiao Zhang, Huazheng Wang, and Qingyun Wu. AutoDefense: Multi-Agent LLM Defense against Jailbreak Attacks.CoRR abs/2403.04783, 2024. 13

    work page arXiv 2024

  57. [57]

    Patchcraft: Exploring texture patch for efficient ai-generated image detection.arXiv preprint arXiv:2311.12397, 2023

    Nan Zhong, Yiran Xu, Sheng Li, Zhenxing Qian, and Xinpeng Zhang. PatchCraft: Exploring Texture Patch for Efficient AI- generated Image Detection.CoRR abs/2311.12397, 2024. 1, 7, 8, 13, 20

    work page arXiv 2024

  58. [58]

    Aigi-holmes: Towards explainable and gener- alizable ai-generated image detection via multimodal large language models.arXiv preprint arXiv:2507.02664, 2025

    Ziyin Zhou, Yunpeng Luo, Yuanchen Wu, Ke Sun, Jiayi Ji, Ke Yan, Shouhong Ding, Xiaoshuai Sun, Yunsheng Wu, and Rongrong Ji. AIGI-Holmes: Towards Explainable and Gener- alizable AI-Generated Image Detection via Multimodal Large Language Models.CoRR abs/2507.02664, 2025. 1, 13

    work page arXiv 2025

  59. [59]

    GenImage: A Million-Scale Benchmark for Detecting AI-Generated Image

    Mingjian Zhu, Hanting Chen, Qiangyu Yan, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. GenImage: A Million-Scale Benchmark for Detecting AI-Generated Image. InAnnual Conference on Neural Infor- mation Processing Systems (NeurIPS). NeurIPS, 2023. 2, 6, 7, 20 A Prompts Used in Our Framework This appendix documents the exact...

    work page 2023

  60. [61]

    AI-generated

    analysis_details: A detailed analysis explaining your decision Table 10: Second prompt template for the Reasoning Agent. evaluates the sufficiency of each debate round and can decide to terminate the debate early if the arguments are deemed suf- ficient. Table 13 and Table 14 are the prompt templates used to instruct the LLM: A.5 VLM Analysis Tool Prompt ...

    work page

  61. [62]

    If one side’s evidence is strong and the other’s is weak or has been effectively countered, the information is likely sufficient

    work page

  62. [63]

    If both sides have presented compelling but con- flicting evidence that has not yet been reconciled, more analysis is needed

    work page

  63. [64]

    Based on these criteria, decide if you have enough information to make a high-confidence final judgment

    If the discussion become repetitive, further rounds are unlikely to be productive. Based on these criteria, decide if you have enough information to make a high-confidence final judgment. Answer ’True’ if sufficient, ’False’ if more debate and analysis would be helpful. Table 13: First prompt template for the Judge Agent. 18 You are an AI Image Forensics ...

    work page

  64. [65]

    Weigh the Evidence: Identify the most compelling piece of evidence from EACH side

    work page

  65. [66]

    Resolve the Core Conflict: Directly address the central disagreement

    work page

  66. [67]

    Required output format:

    State Your Final Conclusion: Based on your analysis, provide a clear final verdict. Required output format:

    work page

  67. [68]

    is_ai_generated: boolean (True if AI-generated, False if real image)

    work page

  68. [69]

    Table 14: Second prompt template for the Judge Agent

    analysis_details: A detailed analysis explaining your decision Format the response as a structured object. Table 14: Second prompt template for the Judge Agent. As a professional AI image detector, please analyze this image carefully:

    work page

  69. [70]

    Determine if this is an AI-generated image or a real image. - Real images include images that are created by humans, including photographs captured by cameras, photos that have been edited with software such as Pho- toshop, or human artistic creations such as hand-drawn sketches and paintings. - AI-generated images include images that are fully or partial...

    work page

  70. [71]

    If you determine it’s an AI-generated image, please specifically identify and list the visual artifacts or characteristics that indicate AI generation, such as: - Unnatural textures or patterns - Inconsistent lighting or shadows - Anatomical errors in humans or animals - Unusual distortions or blending of elements - Text or writing abnormalities - Symmetr...

    work page

  71. [72]

    If you determine it’s a real image, explain what characteristics support this conclusion

    work page

  72. [73]

    Table 15: Prompt template for the VLM Analysis Tool

    Provide your final classification with confidence level (high, medium, or low). Table 15: Prompt template for the VLM Analysis Tool. 19 Table 16: Metadata fields and prefixes considered in the analysis tool. Category Field / Prefix Description Exact Key Fields XMP:CreatorTool Creator tool Software used to generate or edit the image. EXIF:Software Software...

    work page