Paper Image

Detecting AI Image Sources

Published on:

28 March 2024

Primary Category:

Computer Vision and Pattern Recognition

Paper Authors:

Katherine Xu,

Lingzhi Zhang,

Jianbo Shi

Bullets

Key Details

Achieved over 90% accuracy attributing images across 12 text-to-image models

Initialization seeds are highly detectable along with subtle inference variations

Detectable traces exist in style representations and mid-level image structure

Perturbing high-frequencies causes only minor attribution accuracy declines

Training attributor on style features outperforms training on RGB images

AI generated summary

Detecting AI Image Sources

This paper explores detecting and attributing images to 12 state-of-the-art text-to-image AI models. Experiments show initialization seeds are highly detectable, along with subtle inference variations. Analyses also reveal fake images have detectable traces in style representations and mid-level image structure, beyond just high frequencies.

Answers from this paper

Comments

No comments yet, be the first to start the conversation...

Sign up to comment on this paper

Sign Up