FUSE: Unifying Spectral and Semantic Cues for Robust AI-Generated Image Detection

TL;DR

This work tackles robust detection of AI-generated images in the face of evolving generation methods. It introduces FUSE, a hybrid framework that unifies FFT-derived spectral features with CLIP Vision encoder semantic features, forming a joint representation for classification. The model is trained in two progressive stages, with Stage 1 achieving strong performance on benchmarks like Chameleon and GenImage, and Stage 2 further improving generalization to unseen generators. Across GenImage, WildFake, DiTFake, GPT-ImgEval, and Chameleon, FUSE demonstrates robust cross-generator detection, illustrating the value of combining spectral and semantic cues for resilient AIGI detection.

Abstract

The fast evolution of generative models has heightened the demand for reliable detection of AI-generated images. To tackle this challenge, we introduce FUSE, a hybrid system that combines spectral features extracted through Fast Fourier Transform with semantic features obtained from the CLIP's Vision encoder. The features are fused into a joint representation and trained progressively in two stages. Evaluations on GenImage, WildFake, DiTFake, GPT-ImgEval and Chameleon datasets demonstrate strong generalization across multiple generators. Our FUSE (Stage 1) model demonstrates state-of-the-art results on the Chameleon benchmark. It also attains 91.36% mean accuracy on the GenImage dataset, 88.71% accuracy across all tested generators, and a mean Average Precision of 94.96%. Stage 2 training further improves performance for most generators. Unlike existing methods, which often perform poorly on high-fidelity images in Chameleon, our approach maintains robustness across diverse generators. These findings highlight the benefits of integrating spectral and semantic features for generalized detection of images generated by AI.

Figures (7)

• Figure 1: Overview of the FUSE framework.
• Figure 2: Number of training samples from different generators. Most samples are from GenImage 4genimage, while DALL-E 2 and other GANs (StarGAN stargan, StyleGAN stylegan) come from WildFake 27.
• Figure 3: Number of test samples from different generators and real images. Chameleon is from Yan et al. 25, GPT-4o from GPT-ImgEval 34, and Flux, PixArt, and SD 3 from DiTFake 31. Test sets not explicitly mentioned were taken from the corresponding datasets used for first-stage training.
• Figure 4: Confusion matrices for various generators/dataset for Stage 1.
• Figure 5: Confusion matrices for various generators for Stage 2.