Conceitos essenciais
AI-generated image detection relies on texture patches to identify fake images across various generative models.
Resumo
PatchCraft approach focuses on identifying AI-generated images using texture patches.
Texture patches reveal traces left by generative models more effectively than global semantic information.
Smash&Reconstruction preprocessing enhances texture patches and erases global semantic information.
Inter-pixel correlation contrast between rich and poor texture regions boosts detection performance.
Comprehensive benchmark evaluates the effectiveness of the approach.
Outperforms state-of-the-art detectors by a significant margin.
Proposed method extracts a universal fingerprint for various generative models.
Ablation studies show the importance of Smash&Reconstruction and inter-pixel correlation contrast.
Performance remains robust even with image distortions.
Estatísticas
Figure 1. We conduct a comprehensive AI-generated image detection benchmark, including 16 kinds of prevalent generative models [3, 6, 9, 15, 19–21, 31, 33, 39, 52] and commercial APIs [37, 50] like Midjourney [30].
Our approach outperforms the state-of-the-art detector about 4% over average detection accuracy.
Citações
"Texture patches of images reveal more traces left by generative models compared to global semantic information."
"Synthesizing realistic rich texture regions proves to be more challenging for existing generative models."