DF26: We Cannot Tell Fake From Real Anymore
On DF26, both humans and current deepfake detectors perform close to random chance.
The benchmark contains 271 real and 2,420 synthetic single-person speaking videos. The synthetic clips were made with seven recent text-to-video and image-to-video models. The paper argues that existing evaluation protocols are not keeping up with modern generative model shifts. Source: HF Daily Papers' note
The benchmark contains 271 real and 2,420 synthetic single-person speaking videos. The synthetic clips were made with seven recent text-to-video and image-to-video models. The paper argues that existing evaluation protocols are not keeping up with modern generative model shifts. Source: HF Daily Papers' note
score 5