NTH

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

AuthorsChenfan 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

November 8, 2026 2 min read
Watch on YouTube
The one-line take

Venus-DefakerOne2 aims to detect many kinds of image and video forgery with one broadly trained vision-forensics model.

Key results

50M
Training corpus

Heterogeneous anti-forgery training samples.

2B
Model size

Qwen3.5-based foundation model.

90.9%
Image-domain average

Average performance across five image domains.

95.7%
Video-domain average

Average performance across three video domains.

95.3%
RRVBench balanced accuracy

Zero-shot detection on realistic rumor videos.

96.7%
DeepFake average AUC

Average across DeepFake detection benchmarks.

What the paper found

Venus-DefakerOne2 targets a gap in visual forensics: most detectors specialize in one kind of forgery or one media type, while real threats span digital and physical manipulation in both images and videos. Built on a 2B-parameter Qwen3.5 model and trained on a 50M-sample corpus, it uses Tide-Style Tuning: LoRA adaptation, full-parameter fine-tuning, then a final LoRA phase focused on difficult and new examples. For videos, Modality Synergism trains on both native clips and sampled frames, then uses prediction entropy to select the more confident modality for each input. A DINOMAC-based Anomaly Magnifier also supports pixel-level localization, with a global-confidence calibration step to adjust the mask threshold. The team introduces RRVBench, a benchmark of realistic rumor videos, and reports 95.3% balanced accuracy on it. Across benchmark suites, the model averages 90.9% on image domains and 95.7% on video domains; its DeepFake average AUC is 96.7%. The data pipeline draws on models including GPT-5.5 and GPT-Image-2, and the video tests include ByteDance’s SeeDance 2.5. The results point to a single model covering tasks that have typically required separate specialist detectors.

Original 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.

Read the original paper

More in Foundation Models

Browse all 53 papers →
02Foundation Model

Towards Foundation Decision Models: Drex

Nischay Dhankhar, Alisher Amantay, Zhanibek Datbayev, Darkhan Akhmetov, Yash Shah, Nurmukhamed Ubaidullayev, Raman Saparkhan, Dos Baha, Abulhair Saparov

Drex aims to make AI decisions faster and more consistent by predicting directly among options instead of generating long answers.

Read analysis
03Foundation Model

Nemotron 3.5 Super and Lightning: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Models for Agentic AI

NVIDIA, :, Aastha Jhunjhunwala, Aayushi Ashokkumar Srivastava, Abhibha Gupta, Abhinav Garg, Abhinav Khattar, Achsah Wildish, Adam Rajfer, Adi Renduchintala, Adil Asif, Adriana Flores Miranda, Aileen Zaman, Aishwarya Chhabra, Akanksha Shukla, Akhiad Bercovich, Alejandra Rico, Aleksa Gordic, Aleksander Ficek, Alex Gronskiy, Alex Steiner, Alexander Bukharin, Alexandre Milesi, Ali Roshan Ghias, Ali Shehper, Ali Taghibakhshi, Alice Gatti, Alisa Liu, Alok Kumar, Alok Tripathy, Amala Sanjay Deshmukh, Aman Miryala, Aman Shanbhag, Ameya Sunil Mahabaleshwarkar, Ameya Velingker, Amir Klein, Amir Samani, Amit Zuker, Amnon Geifman, Anahita Bhiwandiwalla, Ananth Subramaniam, Andre Manoel, Andrea Santilli, Andrew Tao, Andrii Skliar, Anil Thomas, Anish Mahishi, Anna Shors, Anna Warno, Anuj Doshi, Anusha Kamath, Anwith Kiran, Ao Tang, Arham Mehta, Arnab Dhara, Arnav Komaragiri, Arnav Kundu, Arti Jain, Arun Rao, Arun Venkatesan, Arushi Goel, Asha Anoosheh, Ashish Vaswani, Ashwath Aithal, Asif Ahame...

Nemotron 3.5 pairs hybrid Mamba-Transformer models with agent training and inference optimizations to deliver capable, efficient open AI systems.

Read analysis