How to spot if your art was used to train image generators

Diffusion models like Stable Diffusion train on billions of images, raising questions about whether copyrighted artwork was included without permission. Researchers developed SD-MIA, a black-box attack that detects whether specific images were in a model's training data by analyzing how the model denoises both the image and perturbed text instructions together. Testing on public benchmarks and a new balanced dataset, SD-MIA beats existing methods—even those with unfair access to internal model features—making copyright auditing practical for closed-source platforms.