MetaFake extends LaDeDa (Cavia et al., 2024), a real-time detector that works well on images from the wild but is weaker on diffusion-transformer and autoregressive generators such as FLUX and SD 3.5. On unseen 2025 generators it raises zero-shot AUC from 78.3 to 88.7, and on Synthbuster against real photos from 72.0 to 81.0. Only a small part of the model is trained, so experiments run in seconds on an Apple M1 Max.
Details will follow after the conference.