Venus-DefakerOne2 Expands Unified Visual Forensics

Introduction

Real-world visual forgery is broader than face swapping. An attack may involve image editing, generative content, document manipulation, or a physical presentation designed to fool a vision system. The problem also spans both still images and video. Venus-DefakerOne2 approaches this diversity with a single visual-forensics foundation model aimed at binary forensic decisions that would otherwise be handled by separate specialist detectors.

Key points

• Broader threat coverage. The reported evaluation includes deepfake faces, AI-generated images, tampered documents, manipulated natural images, physical spoofs, and AI-generated videos.
• A mixed fine-tuning schedule. Trained on tens of millions of open-source and in-house samples, the model uses Tide-Style-Tuning. The method alternates LoRA updates with full-parameter fine-tuning to balance compute efficiency, generalization, and fitting of difficult samples while reducing catastrophic forgetting.
• Entropy-guided test adaptation. Entropy-based Test Adaptation analyzes prediction entropy to select an appropriate prediction modality. The supplied abstract does not define the modalities in detail, but the stated goal is to make inference more suitable for varying input conditions.
• Reported benchmark results. The paper lists 96.7% for deepfake faces, 93.3% for AI-generated images, 87.6% for tampered documents, 89.6% for manipulated natural images, 93.1% for physical spoofs, and 96.0% for AI-generated videos. The authors describe the model as outperforming both specialized experts and the previous DefakerOne.

Why it matters

The main contribution is a shift in the scope of visual forensics. Many detectors are designed around one manipulation type or one media format, and their performance may not transfer cleanly to new distributions, compression settings, or attack channels. Venus-DefakerOne2 instead treats coverage itself as a foundation-model problem, seeking one system that can move across digital manipulation, physical spoofing, images, and video.

The tuning strategy also exposes a practical tension in foundation-model adaptation. Parameter-efficient tuning lowers resource requirements, but may limit adaptation to hard examples. Full-parameter training can provide more flexibility, yet costs more and may overwrite previously learned capabilities. Alternating the two is a reasonable compromise, but its value should be assessed alongside ablations, dataset composition, and cross-domain testing.

The reported percentages should not be read as direct guarantees for production security. The supplied material does not provide dataset sizes, class balance, thresholds, false-positive rates, false-negative rates, or inference costs. For identity verification, content moderation, or media provenance, the model’s output would therefore be better treated as one piece of evidence alongside human review and other forensic signals.

Overall, Venus-DefakerOne2 represents a move from narrowly focused forensic experts toward broader visual anti-forgery models. Its approach combines expanded training coverage, alternating adaptation methods, and test-time uncertainty analysis. Whether it remains reliable against unseen generators, changing capture conditions, and evolving attacks will require more technical disclosure and independent reproduction.

Hugging Face Daily Papers