Detecting Advanced AI Generated Images

Summary & Methodology Analysis

The paper addresses a critical forensic gap where current AI generated image detection benchmarks rely on early generative models like traditional generative adversarial networks and diffusion models, failing against high-quality, multi-form outputs from multimodal large language models. Existing detectors show severe performance degradation on these newer images, particularly struggling with structural artifacts in non-natural content like documents and tables. To evaluate this, the authors construct a benchmark dataset using GPT Image2 and Nano Banana2 across three generation protocols: text-to-image, image-to-text-to-image, and image plus prompt-to-image.

To solve the detection problem, the authors propose the SAP-DSP framework, which stands for a structural-artifact-prior-guided dual-stream prompt learning framework. The system operates via a Structural Artifact Prior module that extracts structural responses such as edges, high-frequency signals, lines, and variance to build a structural prior vector. This vector is then fed into a dual-stream prompt learning architecture that separates texture and structure streams, dynamically modulating the structure stream with the structural prior vector. A structure-aware routing fusion module subsequently uses structural statistics and the structural prior vector to adaptively fuse features from both the texture and structure streams.

Evaluated against traditional benchmarks like UFD and GenImage, and existing baselines including LTD, PGC, GAPL, and IAPL, the SAP-DSP model demonstrates superior detection performance across texture, structure, and hybrid content types. However, the study reveals notable limitations, highlighting that current detection methods still struggle specifically with structural artifacts found in non-natural images like documents and tables, and that previous benchmarks fail to represent current multimodal large language model challenges.