This visual detection update focuses on content that does not contain faces. A lot of deepfake detection leans on faces, and this release improves how the image and video models handle scenes without them.

The work responded to a rise in war-related image and video deepfakes circulating on social media, where the subject is often a place or event rather than a person. Training added several datasets toward that goal, including DeepAction and SDFVD, generator sets such as Veo3, Kling2, and LTX-2, and landmark and hard-to-detect image sets like nano-banana-2 and nano-banana-pro, among others.

The result is a model that holds up better on the kind of faceless content that had been harder to judge.