Unveiling Perceptual Artifacts: A Fine-Grained Benchmark for Interpretable AI-Generated Image Detection

TL;DR

X-AIGD introduces a fine-grained, pixel-level benchmark for interpretable AI-generated image detection, cataloging perceptual artifacts into low-level, high-level, and cognitive-level categories and providing paired real/fake images across 13 generators. The study reveals that existing detectors largely ignore artifact cues and rely on opaque features, even when tasked with artifact detection; to address this, the authors develop an attention-alignment approach that regularizes model attention to align with annotated artifact regions, yielding improved interpretability and cross-dataset generalization. The dataset comprises 4,000 real and 4,000 fake images per generator, totaling 52,000 fakes, annotated with 3,035 high-quality artifact masks across 7 categories, enabling robust evaluation of grounded interpretations. Overall, X-AIGD sets a new standard for evaluating interpretable AIGI detection and motivates future work on artifact-aware methods and scalable, human-in-the-loop data expansion, including potential integration with Multimodal Large Language Models.

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 investigation using X-AIGD provides several key insights: (1) Existing AIGI detectors demonstrate negligible reliance on perceptual artifacts, even at the most basic distortion level. (2) While AIGI detectors can be trained to identify specific artifacts, they still substantially base their judgment on uninterpretable features. (3) Explicitly aligning model attention with artifact regions can increase the interpretability and generalization of detectors. The data and code are available at: https://github.com/Coxy7/X-AIGD.

Figures (14)

• Figure 1: Taxonomy of Perceptual Artifacts. Our X-AIGD categorizes perceptual artifacts into three hierarchical levels: 1) low-level distortions, 2) high-level semantics, and 3) cognitive-level counterfactuals. Low-level distortions capture fundamental visual irregularities, such as edge misalignments, unnatural textures, and color inconsistencies. High-level semantics address structural and compositional errors that disrupt object integrity and logical arrangement. Cognitive-level counterfactuals encompass commonsense and physical violations that defy real-world knowledge, such as implausible object relationships or violations of physical laws. Red polygons illustrate the precise localization of these artifacts, supporting fine-grained evaluation of AIGI detection models.
• Figure 1: Comparison with existing datasets. X-AIGD provides detailed annotations (pixel-level masks and categorized artifact labels), real images paired with the fake ones, supporting fine-grained evaluation. PAL4VST, SynArtifact, and RichHF-18K are for perceptual artifact localization.
• Figure 2: Human assessment of annotation quality. The distribution of average confidence scores assigned by human annotators indicates the overall high quality of artifact annotations.
• Figure 3: Accuracy of existing AIGI detectors on AIGIs with different fidelity. The image fidelity is measured by (a) NIQE NIQE; (b) MDFS MDFS; (c) Perceptual Artifact Ratio PAL4VST computed based on our annotations. For these three metrics, higher values indicate lower fidelity.
• Figure 4: Qualitative comparison of the interpretations of different models. Columns 2-4 show the interpretation heatmap for existing AIGI detectors (FatFormer FatFormer and DRCT-CLIP DRCT) and the classification head of the multi-task model. Column 5 is the artifact segmentation results of the multi-task model. The last column shows the ground truth annotations.