Research map / Vision-language and multimodal / Multimodal LLMs
AIGI-Holmes: Towards Explainable and Generalizable AI-Generated Image Detection via Multimodal Large Language Models
Ziyin Zhou, Yunpeng Luo, Yuanchen Wu, Ke Sun, Jiayi Ji, Ke Yan, Shouhong Ding, Xiaoshuai Sun, Yunsheng Wu, Rongrong Ji
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Visual Expert Pre-training Stage (VEP-S) (ablation) | LLaVA-based AIGI-Holmes · P1 / P3 mean accuracy | accuracy: 83.3 -> 86.8 (P1); 90.1 -> 97.2 (P3) | Improved |
| add DPO to baseline LLaVA (ablation) | AIGI-Holmes · P1 / P3 | accuracy: 83.3 -> 84.8; 90.1 -> 92.3 | Improved |
| add DPO on top of VEP-S (ablation) | AIGI-Holmes · P1 / P3 | accuracy: 86.8 -> 87.4; 97.2 -> 97.6 | Improved |
| add Collaborative Decoding on top of VEP-S (ablation) | AIGI-Holmes · P1 / P3 | accuracy: 86.8 -> 90.8; 97.2 -> 98.9 | Improved |
| add DPO and CD on top of VEP-S (ablation) | AIGI-Holmes · P1 / P3 | accuracy: 87.4 -> 93.2; 97.6 -> 99.2 | Improved |
| add DPO stage after SFT (ablation) | AIGI-Holmes explanations · Explanation quality | human ELO rating: 10.670 -> 11.420 | Improved |
| add DPO stage after SFT (ablation) | AIGI-Holmes explanations · Explanation quality | BLEU-1: 0.445 -> 0.622 | Improved |
| add NPR expert alongside CLIP visual expert (ablation) | AIGI-Holmes · P1 / P3 | accuracy: 92.0 -> 93.2; 98.9 -> 99.2 | Improved |
| add CLIP visual expert over LLaVA-only encoder (ablation) | AIGI-Holmes · P1 / P3 | accuracy: 86.8 -> 92.0; 98.2 -> 98.9 | Improved |
| use Holmes pipeline instead of LoRA on LLM only (ablation) | AIGI-Holmes · P1 / P3 | accuracy: 83.3 -> 86.8; 90.1 -> 98.2 | Improved |
| use Holmes pipeline instead of LoRA on all modules (ablation) | AIGI-Holmes · P1 / P3 | accuracy: 84.8 -> 86.8; 91.2 -> 98.2 | Improved |
| use Mistral-7B instead of Llama3-8B as LLM (ablation) | AIGI-Holmes · P1 / P3 | accuracy: 91.5 -> 93.2; 97.8 -> 99.2 | Improved |
Later papers that compared with this method
2 later accepted papers report AIGI-Holmes as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- REVEAL: Reasoning-Enhanced Forensic Evidence Analysis for Explainable AI-Generated Image Detection (ECCV 2026 · better on 3 of 3 shared results)
- TranX-Adapter: Bridging Artifacts and Semantics within MLLMs for Robust AI-generated Image Detection (ICML 2026 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Is Artificial Intelligence Generated Image Detection a Solved Problem? (NeurIPS 2025 · similarity 0.87)
- A Sanity Check for AI-generated Image Detection (ICLR 2025 · similarity 0.86)
- Are High-Quality AI-Generated Images More Difficult for Models to Detect? (ICML 2025 · similarity 0.85)
- I-CEE: Tailoring Explanations of Image Classification Models to User Expertise (AAAI 2024 · similarity 0.85)
- SNIFFER: Multimodal Large Language Model for Explainable Out-of-Context Misinformation Detection (CVPR 2024 · similarity 0.85)
- AI-generated Image Quality Assessment in Visual Communication (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
