Paper presented at CVPR Findings 2026!

Our paper „PRADA: Probability-Ratio-Based Attribution and Detection of Autoregressive-Generated Images“ got accepted at the Findings of the Conference on Computer Vision and Pattern Recognition (CVPR) 2026!

The work by Simon Damm, Jonas Ricker, Henning Petzka, and Asja Fischer tackles the problem of detecting images generated by autoregressive (AR) image generators, a new class of generative models capable of creating highly realistic images at relatively low computational cost. The proposed detection method, PRADA, is based on a special property of AR models: In contrast to other generative models (e.g., GANs or diffusion models), they provide explicit likelihoods for generated images. By learning a simple score function to account for different model architectures, we can reliably estimate how likely an image was generated by a particular model. PRADA is evaluated on a diverse range of class-to-image and text-to-image models, achieving competitive results at drastically reduced training efforts compared to existing methods. The paper will be presented at the Computer Vision and Pattern Recognition Conference (CVPR) taking place from June 3 to 7 in Denver, Colorado. A preprint is available at https://arxiv.org/abs/2511.20068.