I note with interest a new Ukraine-led paper indicating that the older generation of deepfake video detectors perform very poorly on current state-of-the-art genAI video, whether those generators are open source, local AI installs such as Alibaba's Wan 2.2 or Hunyuan Video; or API-only, closed-source models such as Grok Imagine or Kling. Click to play (AUDIO CONTENT). Three examples from the 'Direct to camera casual' category in the DF26 dataset. Source Back in 2024, another paper had already…
This article has been indexed from Unite.AI
Read the original article: