Unveiling Perceptual Artifacts: A Fine-Grained Benchmark for Interpretable AI-Generated Image Detection
Yao Xiao, Weiyan Chen, Jiahao Chen, Zijie Cao, Weijian Deng, Binbin Yang, Ziyi Dong, Xiangyang Ji, Wei Ke, Pengxu Wei, Liang Lin
TL;DR
This paper introduces X-AIGD, a detailed benchmark with pixel-level annotations of perceptual artifacts, to evaluate and improve interpretability in AI-generated image detection models.
Contribution
It presents a new fine-grained benchmark with localized artifact annotations, enabling interpretability assessment and insights into model decision-making processes.
Findings
Existing detectors show minimal reliance on perceptual artifacts.
Detectors trained on specific artifacts still rely on uninterpretable features.
Aligning attention with artifact regions improves interpretability and generalization.
Abstract
Current AI-Generated Image (AIGI) detection approaches predominantly rely on binary classification to distinguish real from synthetic images, often lacking interpretable or convincing evidence to substantiate their decisions. This limitation stems from existing AIGI detection benchmarks, which, despite featuring a broad collection of synthetic images, remain restricted in their coverage of artifact diversity and lack detailed, localized annotations. To bridge this gap, we introduce a fine-grained benchmark towards eXplainable AI-Generated image Detection, named X-AIGD, which provides pixel-level, categorized annotations of perceptual artifacts, spanning low-level distortions, high-level semantics, and cognitive-level counterfactuals. These comprehensive annotations facilitate fine-grained interpretability evaluation and deeper insight into model decision-making processes. Our extensive…
Peer Reviews
Decision·ICLR 2026 Poster
- The dataset contribution is concrete and timely, providing pixel-level artifact masks with semantic labels and paired real/fake images. This directly supports studying why a detector made its decision instead of treating the task as only real or fake. - The analysis of existing detectors is systematic. The authors compare saliency and attention maps with ground-truth artifact masks and show that they often do not overlap, which raises doubts about how explainable current high-performing detect
- There is a growing line of work that uses multimodal large language models (MLLMs) to "look at an image and generate textual explanations," and claims to identify "where it is fake and why it is fake." The authors briefly mention related efforts (e.g., using MLLMs to generate explanations), but a more explicit discussion of works such as [1,2,3] would make the connection to this direction clearer for readers and better contextualize the contribution. - The authors report 12 human annotators, 3
- X-AIGD introduces the first benchmark with paired real-fake images and pixel-level annotations across 7 fine-grained artifact categories, spanning low-level distortions, high-level semantics, and cognitive-level counterfactuals, enabling grounded interpretability evaluation. - Empirical analysis surprisingly reveals that SOTA detectors achieve good performance without relying on perceptual artifacts—contrary to common intuition that visible flaws should be primary cues, as evidenced by uncorr
- No comparison or integration with MLLM-based explanation methods (e.g., GPT-4o, LLaVA) is provided, despite critiquing their lack of grounding—missing an opportunity to show whether X-AIGD enables better spatial reasoning in language models. - Inter-annotator agreement (e.g., IoU-based Fleiss’ κ) is not reported, especially critical for subjective high-level and cognitive-level artifacts, potentially affecting reproducibility.
1. The proposed interpretable benchmark with pixel-level artifact masks is of substantial value to the research community, promoting further progress in explainable image forgery detection and attribution. 2. The paper reveals that current detectors make limited use of perceptible artifacts. By introducing an attention alignment loss, the authors encourage the model to make judgments from a human perceptual perspective, which adds credibility to the interpretability of the approach.
1. The paper lacks detailed information about whether and when the proposed benchmark will be publicly released. Clear disclosure of dataset availability is essential for reproducibility and community adoption. 2. The claim that existing detectors rely on imperceptible artifacts is indeed intuitive. To further validate the effectiveness of the proposed attention alignment loss, more ablation studies are needed. For example: - (1) What happens if the ground-truth mask is randomly assigned?
Code & Models
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsExplainable Artificial Intelligence (XAI) · Adversarial Robustness in Machine Learning · Generative Adversarial Networks and Image Synthesis