--- license: apache-2.0 base_model: Qwen/Qwen3.5-4B library_name: peft pipeline_tag: video-text-to-text language: - en tags: - ai-generated-video-detection - video-forensics - omni-modal - audio-visual - agent - multi-agent - tool-use - reinforcement-learning - lora - qwen3.5 --- # WeVisForensics: Self-Evolving Agent for Omni AI-Generated Video Detection **Paper:** [WeVisForensics: Self-Evolving Agent for Omni AI-Generated Video Detection](https://arxiv.org/abs/2611.21373) | **Code:** [Tencent/WeVisForensics](https://github.com/Tencent/WeVisForensics) | **Project Page:** [WeVisForensics](https://tencent.github.io/WeVisForensics) **WeVisForensics** is a self-evolving agentic detector of AI-generated video. A main investigator agent coordinates forensic inspector sub-agents for visual appearance, physical consistency, common sense, on-screen text, speech content, and voice profile. This repository releases the five LoRA adapters of WeVisForensics-4B, trained with DTPO on top of Qwen3.5-4B. ## Released Adapters | Adapter | Purpose | | --- | --- | | `main` | Main investigator agent | | `visual_appearance` | Visual-appearance inspector | | `physical_consistency` | Physical-consistency inspector | | `common_sense` | Common-sense inspector | | `ocr_consistency` | On-screen text inspector | All five adapters are rank 64 and alpha 128 on [Qwen/Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B), in BF16. `skill_card.json` holds the inspector skill cards used at inference. ## Quick Start Clone the [code repository](https://github.com/Tencent/WeVisForensics), then deploy in one of three ways. **Multi-GPU.** Merge each adapter into the base model, then point the inference config at the merged model and at the bundled `skill_card.json`: ```bash python scripts/merge_lora.py --base Qwen/Qwen3.5-4B --adapters --output ``` **Single GPU.** Serve all five adapters from one vLLM process without merging: ```bash BASE=Qwen/Qwen3.5-4B ADAPTERS= \ GPU_MEMORY_UTILIZATION=0.66 bash scripts/serve_lora.sh ``` **Transformers.** Run the whole agent in-process on one GPU with at least 24 GB, without vLLM. This backs the Gradio demo and the Hugging Face Space: ```bash pip install -r demo/requirements.txt "gradio>=6.0" WVF_ADAPTERS=tencent/WeVisForensics python demo/app.py ``` See the repository README for the full commands. ## Intended Use The release is intended for research in AI-generated video forensics. It is not intended for malicious use or for making autonomous high-impact decisions about individuals. ## License Unless otherwise noted, the adapters and Tencent-authored code are released under the Apache License 2.0. Speech, speaker, and OCR third-party models keep their own licenses. ## Citation If WeVisForensics is helpful to your work, please consider giving us a ⭐ and citing: ```bibtex @misc{du2026wevisforensicsselfevolvingagentomni, title={WeVisForensics: Self-Evolving Agent for Omni AI-Generated Video Detection}, author={Yipeng Du and Yuxi Li and Ying Zhang and Chong Sun and Chen Li and Jing Lyu and Ying Tai}, year={2026}, eprint={2611.21373}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2611.21373}, } ```