Venus-DefakerOne2: A Vision-Forensics Foundation Model with Broader Coverage

Venus-DefakerOne2. A Vision-Forensics Foundation Model with Broader Coverage. By Chenfan Qu and twenty-nine others, from Ant Group's GuangJian Team and the China Society of Image and Graphics. Published on the twenty-seventh of October.

Forgery in the real world does not come in one form. It appears as digital manipulation, and also as physical spoofing. It is carried by images, and also by videos. Despite substantial progress in detecting it, a unified anti-forgery foundation model that covers a threat spectrum that broad has been missing. Defences have tended to cover one corner of it at a time.

Venus-DefakerOne2 sets out to be that unified model. It pushes the boundary of forensic capability along two dimensions at once: the form a forgery takes, and the type of media that carries it. The model is trained on tens of millions of samples, drawn from both open-source and in-house data.

Two key designs carry it. Tide-Style-Tuning alternates between LoRA and full-parameter fine-tuning. The alternation is there to balance computational efficiency, generalization, and the fitting of hard samples, while mitigating catastrophic forgetting. The second design acts when the model is used. Entropy-based Test Adaptation analyses prediction entropy to determine which prediction modality is the optimal one.

The results are reported as new state-of-the-art performance across a comprehensive suite of benchmarks. On DeepFake faces the model reaches ninety-six point seven percent. On AI-generated images it reaches ninety-three point three. On tampered documents, eighty-seven point six, and on manipulated natural images, eighty-nine point six.

Physical spoofs come in at ninety-three point one percent, and AI-generated videos at ninety-six point zero. Across these, the model significantly outperforms both specialised expert models and its own predecessor, DefakerOne.

Breadth is the point of the second version. Where specialised experts each handle one slice of the problem, Venus-DefakerOne2 is offered as a single powerful and versatile foundation model for binary-decision forensic tasks, one question asked of everything from a face to a document to a video: is this real, or forged.