📄 Paper | 🌐 Project Page | 💾 OmniFake Dataset | 🤗 Model Weights
VIGIL is a part-centric structured forensic framework for interpretable and generalizable deepfake detection. Current MLLM-based detectors combine evidence generation and manipulation localization in a single step, blurring faithful observations and hallucinated explanations. VIGIL decouples them through a plan-then-examine pipeline: the model first plans which facial parts to inspect based on global visual cues, then examines each part with independently sourced forensic evidence via stage-gated injection. A progressive three-stage training paradigm (SFT → hard-sample self-training → RL with part-aware rewards) ensures genuine evidence reasoning rather than template memorization.
We also introduce OmniFake, a hierarchical 5-Level benchmark (200K+ images) for fine-grained generalizability evaluation, where the model is trained on only three foundational generators and progressively tested up to in-the-wild social-media data.
- Part-Grounded Reasoning — Every forensic claim is anchored to a specific facial part, decoupling claims from evidence to eliminate hallucination.
- Stage-Gated Injection — Part-level forensic signals (frequency-domain + pixel-level) are delivered only during examination, preserving autonomous planning.
- Reasoning Reversion — Accumulated part-level evidence can overturn an initially incorrect global judgment.
- OmniFake Benchmark — 5-Level hierarchical evaluation from in-domain to in-the-wild social-media data.
| Stage | Description |
|---|---|
| Plan | Observe global visual cues and select which facial parts to inspect — without exposure to external forensic signals. |
| Examine | Inject frequency-domain and pixel-level evidence into each selected part via stage-gated mechanism. |
| Synthesize | Aggregate part-level findings into a final verdict, with the ability to overturn initial judgments. |
Training: (1) SFT on signal-semantic annotations, (2) hard-sample self-training via rejection sampling, (3) RL with part-aware & evidence-conclusion consistency rewards (GRPO).
| Level | Name | Description |
|---|---|---|
| L1 | In-Distribution | Same generators as training |
| L2 | Cross-Architecture | Unseen models within related paradigm families |
| L3 | Cross-Model | Entirely unseen generation principles (flow matching, autoregressive, commercial generators, etc.) |
| L4 | Cross-Task | Localized manipulation (inpainting, face restoration) |
| L5 | In-the-Wild | Social media, unknown methods & real-world degradations |
Accuracy (%) on OmniFake. Bold = best, underline = second best.
| Method | L1 | L2 | L3 | L4 | L5 | Avg. |
|---|---|---|---|---|---|---|
| AIDE ICLR'25 | 72.3 | 89.1 | 81.6 | 67.6 | 75.0 | 78.9 |
| Co-SPY CVPR'25 | 83.5 | 89.9 | 89.0 | 71.3 | 82.5 | 84.7 |
| DDA NeurIPS'25 | 97.8 | 94.6 | 91.9 | 81.0 | 80.7 | 88.8 |
| GPT-5.2 | 65.8 | 82.7 | 73.6 | 58.8 | 69.0 | 71.4 |
| Gemini-3-Pro | 72.3 | 85.9 | 78.5 | 61.0 | 69.9 | 75.0 |
| FakeVLM NeurIPS'25 | 83.5 | 81.0 | 76.8 | 74.5 | 74.9 | 77.1 |
| Veritas ICLR'26 | 96.8 | 94.9 | 89.9 | 79.0 | 81.1 | 87.6 |
| VIGIL (Ours) | 98.6 | 96.7 | 95.5 | 89.5 | 86.0 | 93.1 |
We are currently organizing the code and data. Stay tuned!
| Resource | Status |
|---|---|
| Training & Inference Code | Coming Soon |
| Model Weights | Coming Soon |
| OmniFake Dataset | Coming Soon |
| Demo | Coming Soon |
If you find our work useful, please consider citing:
@article{li2026vigil,
title={VIGIL: Part-Grounded Structured Reasoning for Generalizable Deepfake Detection},
author={Li, Xinghan and Xu, Junhao and Chen, Jingjing},
journal={arXiv preprint arXiv:2603.21526},
year={2026}
}This project is released under the Apache 2.0 License.