Veritas++: Value-aware On-Policy Distillation for Perception-Enhanced AIGI Detection

[ 📄Paper | 🚀Model ]

Veritas++ is an interpretable multimodal large language model (MLLM) for generalizable AI-generated image (AIGI) detection. It treats reliable visual perception as the foundation of authenticity reasoning: the model first learns to perceive fine-grained details, semantic anomalies, and pixel-level discrepancies, and then internalizes these abilities through Value-aware On-Policy Distillation (VaOPD).

Highlights

• Perception-oriented Learning (PoRL) strengthens fine-grained visual, semantic-anomaly, and pixel-level perception with verifiable rewards.
• Value-aware On-Policy Distillation (VaOPD) internalizes perception-aware reasoning through a privileged EMA self-teacher.
• Veritas++ achieves strong generalization on standard OOD, in-the-wild, and emerging AIGI detection benchmarks.

Method

The pipeline contains a high-quality cold start, PoRL-based perception learning, and VaOPD-based perception-enhanced reasoning. The model is built on Qwen3-VL-8B.

Value-aware On-Policy Distillation

VaOPD scores student-generated trajectories with a privileged EMA self-teacher and adapts the distillation signal at three levels: it prioritizes incorrect trajectories, emphasizes informative tokens, and adjusts the distillation direction for corrective signals.

Installation

conda create -n veritaspp python=3.10 -y
conda activate veritaspp

pip install -e .

🔎 Inference on a single image

Download Veritas++ 🔥🔥🔥. We recommend using vLLM for model deployment:

sh self_scripts/deploy/deploy_model.sh /path/to/Veritas++

Set the checkpoint path in vllm_server_config inside self_scripts/infer/infer_vllm_single.py, then run:

python self_scripts/infer/infer_vllm_single.py \
  --model model \
  --image_path /path/to/image.jpg

Citation

If you find this work useful, please cite:

@article{tan2026veritaspp,
  title   = {Veritas++: Value-aware On-Policy Distillation for
             Perception-Enhanced AIGI Detection},
  author  = {Tan, Hao and Lan, Jun and Tan, Zichang and Liu, Ajian and
             Yu, Zijian and Song, Chuanbiao and Zhu, Huijia and
             Wang, Weiqiang and Wan, Jun and Lei, Zhen},
  journal={arXiv preprint arXiv:2607.27113},
  year={2026}
}

License

This repository is released under the Apache License 2.0.

Acknowledgements

This project is built upon ms-swift and Qwen3-VL. We thank their authors and the open-source community for their contributions.