Venus-DefakerOne2: A Vision-Forensics Foundation Model with Broader Coverage
The authors report a single model that detects digital forgeries and physical spoofs across images and video.
Venus-DefakerOne2 was trained on tens of millions of open-source and in-house samples. Its training alternates LoRA and full-parameter fine-tuning to balance efficiency and generalization while limiting catastrophic forgetting. At test time, it uses prediction entropy to select the prediction modality. The authors report state-of-the-art results across six benchmark categories, including 96.7% for DeepFake faces, 87.6% for tampered documents and 96.0% for AI-generated videos.
HF Daily Papers' note
Venus-DefakerOne2 was trained on tens of millions of open-source and in-house samples. Its training alternates LoRA and full-parameter fine-tuning to balance efficiency and generalization while limiting catastrophic forgetting. At test time, it uses prediction entropy to select the prediction modality. The authors report state-of-the-art results across six benchmark categories, including 96.7% for DeepFake faces, 87.6% for tampered documents and 96.0% for AI-generated videos.
HF Daily Papers' note
score 4