AIGI-Holmes: Towards Explainable and Generalizable AI-Generated Image Detection via Multimodal Large Language Models

Conference: ICCV 2025 arXiv: N/A (CVF OpenAccess) Code: https://github.com/wyczzy/AIGI-Holmes Area: Multimodal VLM Keywords: AI-generated image detection, multimodal large language models, explainable detection, direct preference optimization, collaborative decoding

TL;DR

This paper proposes AIGI-Holmes, which adapts MLLMs into a "Holmes"-style detector capable of both accurately identifying AI-generated images and providing human-verifiable explanations. This is achieved by constructing the Holmes-Set dataset with explanatory annotations and a carefully designed three-stage training pipeline (visual expert pre-training → SFT → DPO). At inference time, a collaborative decoding strategy further enhances generalization.

Background & Motivation

Problem Definition

The rapid advancement of AI-generated content (AIGC) technology has enabled highly realistic AI-generated images (AIGI) to be misused for spreading misinformation, threatening public information security. Existing detection methods face two core issues:

Lack of Explainability: Current detection models are black boxes, and their outputs are difficult for humans to verify. Without human-verifiable explanations, detection results remain untrustworthy.

Lack of Generalizability: AIGC technology evolves rapidly (e.g., FLUX, SD3.5, VAR), and existing methods struggle to generalize to the latest generative techniques.

Why MLLMs?

MLLMs possess commonsense understanding and natural language generation capabilities, enabling semantic-level analysis of visual content—making them ideal candidates for addressing explainability and generalizability. However, directly applying MLLMs faces two challenges:

• Scarcity of Training Data: Existing AIGI detection datasets (e.g., CNNDetection, GenImage, DRCT) contain only visual modality data and lack instruction-tuning datasets suitable for MLLM SFT. FakeBench and LOKI offer preliminary attempts but rely on GPT-4o annotations and are too small in scale.
• Suboptimal Supervised Fine-Tuning: Training MLLMs solely on SFT datasets yields limited results, as MLLMs are insufficient in image classification and low-level perception tasks, and SFT models may mechanically replicate explanation templates rather than genuinely understanding the root causes of artifacts or semantic errors.

Method

Overall Architecture

AIGI-Holmes consists of two core components: the Holmes-Set dataset and the Holmes Pipeline training framework.

Architecturally, an NPR (Neighboring Pixel Relationships) visual expert is added on top of LLaVA to capture low-level artifact information. The input processing pipeline is as follows: - A CLIP visual encoder \(F\) extracts high-level semantic features \(f_{img}\) - An NPR visual expert \(R\) extracts low-level artifact features \(f_{npr}\) - Both are injected into the LLM via a projector: \(H = \text{LLM}(\text{proj}([f_{img}, f_{npr}]), f_t)\)

Key Design 1: Holmes-Set Dataset

Holmes-SFTSet (65K Images)

Data sources consist of two parts: 1. Existing datasets: 45K images selected from CNNDetection, GenImage, and DRCT 2. Expert-filtered images: 20K images filtered by specialist small models to contain common AI-generated defects (text, human body, face, projective geometry, commonsense, physical laws)

Annotation employs a Multi-Expert Jury approach, with cross-annotation and evaluation by four open-source MLLMs (Qwen2VL-72B, InternVL2-76B, InternVL2.5-78B, Pixtral-124B): - General Positive Prompt: Analyzes high-level semantic dimensions (anatomy, physical laws) and low-level dimensions (texture, sharpness) - General Negative Prompt: Generates adversarial annotations that form DPO data pairs \(D_1\) with positive annotations - Specialist Prompt: Targeted annotations for specific defects in the 20K expert-filtered images

Quality control adopts an MLLM-as-a-judge paradigm, retaining only annotations with the highest consensus.

Holmes-DPOSet (65K + 4K)

To address the "mechanical template copying" problem in SFT models, a human-aligned preference dataset is constructed: - \(D_1\): Natural positive-negative pairs from General Positive/Negative Prompts - \(D_2\): Manually modified — 2K human-annotated samples + 2K samples modified using Specialist Prompts. Human experts provide revision suggestions on SFT model outputs (adding correct information, removing erroneous or irrelevant explanations), which are then executed by DeepSeek-V3

Key Design 2: Holmes Pipeline (Three-Stage Training)

Stage 1: Visual Expert Pre-training

The goal is to equip the visual experts with generalization capability in the AIGI detection domain. Binary classification pre-training is performed separately on two visual encoders: - CLIP-ViT-L/14 is fine-tuned with LoRA (\(r=4, \alpha=8\)), obtaining classification results via an MLP from the CLS feature \(f_{cls}\) - The first two layers of the NPR-based ResNet are fully fine-tuned, with classification also performed via an MLP - Loss functions: \(l_{clip} = l_{bce}(y_{clip}, y), \quad l_{npr} = l_{bce}(y_{npr}, y)\)

Why pre-train? Neither vanilla CLIP nor ResNet was designed for AIGI detection; they are insufficient in low-level artifact perception and classification when applied directly to downstream tasks. Pre-training endows them with domain-specific feature extraction capabilities.

Stage 2: Supervised Fine-Tuning (SFT)

The pre-trained visual experts are integrated into the LLM, and autoregressive text loss training is conducted on Holmes-SFTSet: - Visual experts are frozen - The projector and LLM LoRA components (rank=128, α=256) are trained - Loss: \(l_{txt} = l_{ce}(H, H_{txt})\)

Stage 3: Direct Preference Optimization (DPO)

Human preference alignment is performed on Holmes-DPOSet (\(D = D_1 \cup D_2\)):

\[L_{DPO}(\phi) = -\mathbb{E}_{(x, y_w, y_l) \sim D} \left[ \log \sigma \left( \beta \left( \log \frac{\pi_\phi(y_w|x)}{\pi_{base}(y_w|x)} - \log \frac{\pi_\phi(y_l|x)}{\pi_{base}(y_l|x)} \right) \right) \right]\]

The key role of DPO is to reshape the MLLM's reasoning pattern, aligning explanations with human judgment standards rather than remaining at the level of suboptimal fine-tuning.

Key Design 3: Collaborative Decoding

At inference time, the MLLM and pre-trained visual experts make joint decisions by adjusting the logit values for "real" and "fake" tokens:

\[\text{logit}_{new}(y=k) = \alpha \cdot \text{logit}_{raw}(y=k) + \beta \cdot \text{logit}(y_{clip}=k) + \gamma \cdot \text{logit}(y_{npr}=k)\]

where \(\alpha=1, \beta=1, \gamma=0.2\).

Why does this work? By retaining the MLLM's predictions while incorporating the visual experts' judgments, the approach prevents the MLLM from overfitting to forgery types seen during training, thereby improving generalization to unseen domains.

Loss & Training

Stage	Trainable Parameters	Loss Function	Hyperparameters
Visual Expert Pre-training	CLIP LoRA(r=4) + ResNet full params	Binary CE	batch=32, 5 epochs
SFT	Projector + LLM LoRA(r=128)	Autoregressive CE	lr=5e-5, batch=16, 3 epochs
DPO	Projector + LLM LoRA(r=48)	DPO Loss	lr=5e-7, batch=4, β=0.1, 2 epochs

Key Experimental Results

Main Results

The paper evaluates under three protocols (Protocol-I/II/III), with P3 being the most challenging — training on diffusion model data and testing on entirely new autoregressive generative models and state-of-the-art diffusion models.

Protocol-III Detection Accuracy (Acc. %):

Method	VAR	FLUX	Janus-Pro-7B	SD3.5-Large	Mean Acc.	Mean A.P.
CNNSpot	59.9	63.8	85.0	78.2	72.9	85.6
NPR	85.9	91.6	73.9	93.4	84.0	89.5
UnivFD	64.3	87.8	96.4	75.7	83.6	95.9
RINE	85.0	97.8	97.2	98.9	96.2	99.5
AIDE	93.6	99.4	97.8	98.6	97.0	99.7
AIGI-Holmes	99.6	99.4	98.0	99.9	99.2	99.9

AIGI-Holmes achieves 98%+ accuracy across all generators, with Mean Acc. surpassing AIDE by 2.2% and RINE by 3.0%.

Explanation Quality Comparison (MLLM vs. AIGI-Holmes):

Model	BLEU-1	ROUGE-L	CIDEr	ELO Rating
GPT-4o	0.433	0.308	0.005	10.271
Pixtral-124B	0.428	0.270	0.010	10.472
AIGI-Holmes (w/o DPO)	0.445	0.315	0.023	10.670
AIGI-Holmes (w/ DPO)	0.622	0.375	0.107	11.420

Ablation Study

Core Component Ablation (Acc. %):

VEP-S	DPO	CD	P1	P3
✗	✗	✗	83.3	90.1
✓	✗	✗	84.8	92.3
✓	✓	✗	87.4	97.6
✓	✗	✓	90.8	98.9
✓	✓	✓	93.2	99.2

• Visual Expert Pre-training yields +2.2% on P3
• DPO contributes +0.4% (but improves ELO Rating by 0.75)
• Collaborative Decoding contributes the most: +1.7%
• The combination of all three improves approximately 10% over the baseline

Robustness Evaluation (P3 Mean Acc. %):

Method	JPEG QF=75	Gaussian σ=2	Resize ×0.5
AIDE	92.8	90.7	89.2
RINE	92.4	92.8	92.3
AIGI-Holmes	99.0	97.9	95.9

Key Findings

1. Collaborative Decoding is key to generalization: By incorporating domain knowledge from visual experts, collaborative decoding effectively prevents the MLLM from overfitting to forgery types seen during training.
2. DPO is critical for explanation quality: Although DPO yields limited improvement in detection accuracy (+0.4%), it substantially improves explanation quality (ELO +0.75).
3. MLLMs focus on high-level semantic features: Even under perturbations such as JPEG compression and Gaussian blur, explanation quality metrics do not significantly degrade, indicating that the MLLM relies on high-level semantics rather than low-level artifacts.
4. Multi-expert cross-validation improves data quality: The Multi-Expert Jury approach is more reliable than single-model annotation.

Highlights & Insights

1. Data-driven methodology: Holmes-Set is the first AIGI detection dataset containing explanatory annotations and human preference data, filling a critical data gap.
2. Elegant three-stage training design: The progressive Visual Expert → SFT → DPO design addresses a specific problem at each stage (feature extraction → explanation generation → human alignment).
3. Inference-stage innovation: Collaborative Decoding introduces no additional training cost, incorporating visual expert judgments only at inference time — a low-cost yet effective generalization enhancement strategy.
4. Multi-Expert Jury annotation method: Using multiple MLLMs for cross-annotation replaces costly GPT-4o or human annotation, reducing cost while maintaining quality.

Limitations & Future Work

1. Inference overhead: Collaborative decoding requires running both the MLLM and visual experts simultaneously, increasing inference time.
2. Backbone dependency: The method is built on LLaVA-1.6-mistral-7B; adopting a stronger MLLM backbone may yield further improvements.
3. DPO data scale: The manually modified DPO data contains only 4K samples; scaling up may further improve explanation quality.
4. Real-time detection scenarios: As an MLLM-based solution, the approach is difficult to deploy for real-time detection requirements.
5. Video generation detection: The current work focuses on images; detection of video AIGC (e.g., Sora) remains an unexplored direction.

Related Work & Insights

• Inspiration from NPR visual expert: Low-level neighboring pixel relationships are valuable for AIGI detection but must be combined with high-level semantic features to achieve maximum effectiveness.
• DPO in specialized domains: This work demonstrates that DPO is effective not only for general-purpose dialogue but also for improving output quality and human preference alignment in vertical detection tasks.
• A new paradigm for explainable AI: Shifting from "post-hoc explanation" to "generative explanation," allowing models to naturally output their reasoning process alongside detection decisions.

Rating

• Novelty: ⭐⭐⭐⭐ (The three-stage training framework is distinctive in design, though individual components are not entirely novel)
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ (Three protocols, robustness testing, explanation quality evaluation, and comprehensive ablation)
• Writing Quality: ⭐⭐⭐⭐ (Clear structure, though some details require consulting the appendix)
• Value: ⭐⭐⭐⭐⭐ (Explainable and generalizable AIGI detection is an important and practically significant direction)

Related Papers

• [ICCV 2025] CompCap: Improving Multimodal Large Language Models with Composite Captions
• [ICCV 2025] Bidirectional Likelihood Estimation with Multi-Modal Large Language Models for Text-Video Retrieval
• [ICCV 2025] ShortV: Efficient Multimodal Large Language Models by Freezing Visual Tokens in Ineffective Layers
• [ICCV 2025] DocThinker: Explainable Multimodal Large Language Models with Rule-based Reinforcement Learning for Document Understanding
• [ICCV 2025] On Large Multimodal Models as Open-World Image Classifiers