Venus-DefakerOne2 claims SOTA on six forgery benchmarks

A new preprint claims one foundation model that detects six distinct categories of forged media at state-of-the-art accuracy. In a paper posted to arXiv, the authors introduce Venus-DefakerOne2 and report 96.7% on DeepFake faces, 93.3% on AI-generated images, 87.6% on tampered documents, 89.6% on manipulated natural images, 93.1% on physical spoofs, and 96.0% on AI-generated videos.

The pitch is consolidation. "Real-world forgery is inherently heterogeneous: it manifests not only as digital manipulation but also as physical spoofing, and is carried not only by images but also by videos," the authors write, framing their system as the missing unified detector where today's content-integrity stacks run separate models per attack type.

Two training ideas carry the claim. "Tide-Style-Tuning alternates LoRA and full-parameter fine-tuning" to balance generalization, hard-sample fitting, and what the authors call catastrophic forgetting; an Entropy-based Test Adaptation step then picks the prediction modality by analyzing entropy at inference. The model, the paper says, is trained on "tens of millions of open-source and in-house samples" and "significantly outperforms both specialized expert models and its predecessor, DefakerOne."

The abstract names no parameter count, no latency numbers, and no breakdown of which generator families sit inside each benchmark. One preprint, one lab's reported numbers.