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

Authors: Chenfan Qu, Yachao Liang, Mingqi Fang, Jiangwei Xie, Zhenming Wang, Baoyue Zhang, Changjiang Jiang, Zheng Chang, Fanli Jin, Xuekang Zhu, Kaiwen Feng, Junchi Li, Weihong Zhang, Haoyue Wang, Chenfeng Zhang, Yinan Xia, Fangming Cui, Dong Fan, Jingjing Liu, Tianyi Zhang, Luobin Chen, Ji-Zhe Zhou, Sheng Li, Wei Wang, Bin Li, Ran He, Xinpeng Zhang, Linna Zhou, Jian Liu, Weiqiang Wang

Published: 2026-10-27 20:01:00+00:00

Comment: Technical Report

AI Summary

This paper introduces Venus-DefakerOne2, a vision-forensics foundation model that significantly expands the capability boundary for detecting heterogeneous real-world forgery across digital and physical forms, and image and video media types. The model achieves state-of-the-art performance on a comprehensive suite of benchmarks by incorporating novel training strategies like Tide-Style-Tuning and Entropy-based Test Adaptation, establishing itself as a powerful and versatile solution for binary-decision forensic tasks.

Abstract

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. Despite substantial progress, a unified anti-forgery foundation model covering such a broad threat spectrum remains absent. In this paper, we present Venus-DefakerOne2, which substantially extends the forensic capability boundary along both the forgery-form and media-type dimensions. Trained on tens of millions of open-source and in-house samples, it incorporates two key designs. Tide-Style-Tuning alternates LoRA and full-parameter fine-tuning to balance computational efficiency, generalization, and hard-sample fitting while mitigating catastrophic forgetting. Entropy-based Test Adaptation determines the optimal prediction modality by analyzing prediction entropy. Venus-DefakerOne2 achieves new state-of-the-art performance across a comprehensive suite of benchmarks, including DeepFake faces (96.7%), AI-generated images (93.3%), tampered documents (87.6%), manipulated natural images (89.6%), physical spoofs (93.1%), and AI-generated videos (96.0%). It significantly outperforms both specialized expert models and its predecessor, DefakerOne, establishing itself as a powerful and versatile foundation model for binary-decision forensic tasks.


Key findings
Venus-DefakerOne2 achieves new state-of-the-art performance across various benchmarks, including DeepFake faces (96.7%), AI-generated images (93.3%), tampered documents (87.6%), manipulated natural images (89.6%), physical spoofs (93.1%), and AI-generated videos (96.0%). It significantly outperforms both specialized expert models and its predecessor, DefakerOne, demonstrating robust generalization to unseen, realistic synthetic manipulations and effectively addressing the fragmented nature of anti-forgery research.
Approach
The model, Venus-DefakerOne2, is built upon a multimodal large language model (MLLM), specifically Qwen3.5-2B, augmented with an Anomaly Magnifier module for fine-grained forensic cue perception. It employs 'Tide-Style Tuning' which alternates LoRA and full-parameter fine-tuning to balance efficiency, generalization, and hard-sample fitting, and 'Entropy-based Test Adaptation' for video forensics to determine the optimal prediction modality (native clip or frame sequence) by analyzing prediction entropy.
Datasets
Tens of millions of open-source and in-house samples, including: DiffusionForensics, CommunityForensics, GenImage, LAION, ForenSynths, FaceForensics++, CelebDF-v2, DFD, DFDC, ScaleDF, DF40, WDF, MFFI, DanceText, DocTamper, OSTF, T-SROIE, RTM, SACP, RIFLC, MIML, CASIA-v2, OpenSDID, SIDSet, MS-COCO, So-Fake, So-Fake-OOD, tampCOCO, CelebA-Spoof, MSU-MFSD, OULU-NPU, DLC2021, SCVD, GVF, GVD, AIGVDBench, MintVid, TalkingHeadBench, ComGenVid, GenBuster, RABench, ManipulationBench, and a newly constructed Realistic Rumor Video Benchmark (RRVBench).
Model(s)
Qwen3.5-2B (as MLLM backbone), DINOMAC (DINOv3-Large architecture fine-tuned with LoRA adapter as Anomaly Magnifier).
Author countries
UNKNOWN