Ant Group releases DefakerOne2, a 2B vision-forensics foundation model covering digital and physical image and video forgery
Ant Group's GuangJian Team released DefakerOne2, a 2B-parameter vision-forensics foundation model unifying AI-generated content detection, DeepFake face detection, document tampering detection, natural-image manipulation detection, face/document anti-spoofing, and AI-generated video detection. Model weights and the RRVBench benchmark were released on ModelScope on 2026-10-29; the technical report and project page followed on 2026-10-31. DefakerOne2 achieves state of the art on 40 of 53 detection benchmarks and 11 of 14 localization benchmarks.
China context
- Original name
- 蚂蚁集团
- Outside China
- Open weights · github.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can access the model weights on ModelScope and integrate DefakerOne2 into their own moderation or forensics pipelines, subject to the model's license.
- For investors
- Ant Group's release of a state-of-the-art open-weight forensics model may pressure commercial deepfake detection startups, as a capable alternative becomes freely available.
DefakerOne2 is a unified vision-forensics foundation model from Ant Group's GuangJian Team. It expands defense coverage along both the forgery-form (digital to digital + physical) and media-type (image to image + video) dimensions. The model is built on the Qwen3.5-2B multimodal backbone with an Anomaly Magnifier auxiliary encoder based on DINOMAC. It was trained on about 50M samples from public datasets, self-synthesized data, and manually annotated in-the-wild attack samples. The model provides both image/video-level verdicts and pixel-level localization masks, and emits a confidence score with every verdict. It outperforms its predecessor DefakerOne and existing state-of-the-art detectors across all image and video forensic domains.
DefakerOne2 uses a three-phase 'Tide-Style Tuning' schedule (LoRA → full-parameter → LoRA) to balance efficiency, generalization, and hard-sample fitting. It employs 'Modality Synergism' with joint video/frame training and entropy-based test adaptation to route samples to the stronger modality. An 'Anomaly Magnifier + Calibration Loop' uses a DINOMAC-based auxiliary encoder and a global-judgment-conditioned pixel threshold for joint verdicts and localization. The model is evaluated on 53 detection and 14 localization benchmarks, achieving SOTA on 40/53 and 11/14 respectively, and ranks first in all eight forensic domains.
Organizations needing to detect AI-generated or manipulated media can now use a single open-weight model instead of multiple domain-specific detectors, reducing integration and maintenance costs. Ant Group's release of model weights on ModelScope makes the technology accessible to developers and researchers outside China, potentially shifting the competitive landscape for commercial deepfake detection vendors.
DefakerOne2 offers a unified solution for media forensics, potentially reducing the need for multiple specialized models. Its open-weight release on ModelScope allows integration into commercial products, subject to the model's license terms. The model's confidence scores and pixel-level localization support evidence presentation and threshold-based operation in moderation workflows.
The next verifiable signal is whether DefakerOne2 is adopted in production moderation pipelines by platforms or enterprises, and whether independent evaluations confirm its benchmark performance. The release of the RRVBench benchmark may lead to further research on robust video forgery detection.