Online Detection of AI-Generated Images
David C. Epstein
Ishan Jain
Oliver Wang
Richard Zhang


With advancements in AI-generated images coming on a continuous basis, it is increasingly difficult to distinguish traditionally-sourced images (e.g., photos, artwork) from AI-generated ones. Previous detection methods study the generalization from a single generator to another in isolation. However, in reality, new generators are released on a streaming basis. We study generalization in this setting, training on N models and testing on the next (N+k), following the historical release dates of well-known generation methods. Furthermore, images increasingly consist of both real and generated components, for example through image inpainting. Thus, we extend this approach to pixel prediction, demonstrating strong performance using automatically-generated inpainted data. In addition, for settings where commercial models are not publicly available for automatic data generation, we evaluate if pixel detectors can be trained solely on whole synthetic images.

Online Detection Results

Accuracy measures if a classifier is able to identify fakes. AuC measures if it has the right features to do so. Generators are added into the classifier one at a time (x-axis), while accuracy/auc are measured across test sets (y-axis). Values below the green boxes are already trained on (and achieve near 100%). Values above the green boxes indicate generalization.

While the specifics of many methods are unknown, this study indicates that certain methods may have underlying similarities. For example, LDM generalizes well to Midjourney (v2-v3), Stable Diffusion (v1).

Detecting Inpainting

Additionally, we show that GenAI methods contain enough local cues to train a per-pixel detector for inpainting applications.


D. C. Epstein, I. Jain, O. Wang, R. Zhang.
Online Detection of AI-Generated Images.
In ICCV DFAD Workshop, 2023.
(hosted on ArXiv)



We thank Sheng-Yu Wang and Alexei A. Efros for helpful discussion, Charlie Scheinost and Josh Arteaga for their support and discussion, and Deepti Clark for help gathering inpainting data.