The malicious use of generative artificial intelligence to create highly realistic deepfake videos raises serious ethical concerns and poses substantial challenges to AI safety. However, existing deepfake video benchmarks provide limited coverage of recent synthesis methods and generally lack reliable fine-grained textual annotations. Meanwhile, conventional detectors and multimodal large language models (MLLMs), whether operating as a single model or relying on a single analytical perspective, often fail to capture subtle forgery artifacts, limiting their generalization to emerging AI-generated methods. To address these limitations, we introduce FaceVid-Forensics-100K, a large-scale deepfake video dataset comprising 100,000 videos and spanning 33 synthesis methods across face swapping, face reenactment, and entire-face synthesis, including recent generators such as Seedance 2.0. The dataset provides fine-grained textual annotations of visual observations and verdict-consistent forensic explanations, automatically synthesized through a multi-model aggregation and conflict-resolution pipeline powered by advanced MLLMs. Building on this benchmark, we propose a multi-agent forensic reasoning framework that employs four specialized domain-expert agents to independently analyze forgery cues from four perspectives: texture, lighting, motion, and physics. A judge agent then reconciles their reports to produce a final prediction together with an explanation. Extensive evaluations on out-of-domain test sets show that, despite being composed entirely of small open-source MLLMs, our framework outperforms all methods including closed-source GPT and Gemini models and ranks first across all reported metrics on this benchmark.
100K
21,075 real and 78,925 fake videos, with training, in-domain, and out-of-domain protocols.
33
Broad coverage of face swapping, reenactment, and recent entire-face generators such as Seedance 2.0.
4 + 1
Four independent observers collect evidence; a separate judge resolves agreement and conflict.
ARGUS separates evidence collection from decision making. Each observer is optimized for one forensic dimension and does not issue a final verdict.
Specialization encourages systematic cue discovery and prevents one early interpretation from anchoring every perspective.
The judge weighs corroborating and conflicting reports rather than treating a single salient artifact as decisive.
The output pairs a binary prediction with a concise forensic rationale grounded in the observations.
The benchmark targets generalization to emerging generators and provides more than binary supervision.
The full system ranks first across all reported metrics on the OOD test set.
| Method | Accuracy | Recall | F1 |
|---|---|---|---|
| Small vision models | |||
| DFGaze (TIFS’24) | 56.67 | 17.92 | 27.24 |
| DFD-FCG (CVPR’25) | 61.25 | 29.22 | 39.16 |
| Effort (ICML’25) | 63.92 | 34.79 | 44.76 |
| TALL++ (IJCV’24) | 63.47 | 42.97 | 45.07 |
| TFCU (CVPR’25) | 64.28 | 33.44 | 45.20 |
| Open-source MLLMs | |||
| Qwen2.5-VL-7B | 35.94 | 16.51 | 13.24 |
| InternVL3.5-8B | 37.87 | 17.60 | 14.53 |
| MiMo-V2.5 (310B-A15B) | 49.85 | 15.00 | 18.69 |
| Qwen3.6-35B-A3B | 53.17 | 50.05 | 35.72 |
| Closed-source MLLMs | |||
| GPT-4o (2024) | 57.63 | 40.36 | 37.55 |
| GPT-5-mini (2025) | 59.31 | 38.70 | 39.00 |
| Gemini-2.5-Pro (2025) | 63.78 | 75.29 | 47.45 |
| Gemini-3.5-Flash (2026) | 63.34 | 58.75 | 46.22 |
| Forensics-tuned MLLMs | |||
| Skyra (CVPR’26) | 60.50 | 30.05 | 38.30 |
| VideoVeritas (ICML’26) | 57.87 | 78.96 | 43.22 |
| Multi-agent system | |||
| Ours (w/o Video) | 67.41 | 65.00 | 51.01 |
| Ours (w/ Video) | 69.87 | 81.82 | 53.28 |
Values are percentages. Results are reported on 7,636 OOD videos from 20 generators excluded from training.
Hover over the filmstrip to inspect the sampled frames, then switch among the four independent Observation Agents. The Judge Agent reconciles their reports into the final ARGUS verdict.
Frame 01 / 09
Facial skin appears overly smooth and airbrushed. The hair looks waxy and lacks strand detail. Soft hairline boundaries and painterly foliage suggest synthetic texture.
Frame 01 / 09
The skin has an unnaturally smooth, waxy quality with a lack of fine pores and micro-textures. Severe blending artifacts, unnatural warping, and overly bright regions lack natural shading.
Browse the substantive Appendix F prompt instructions without expanding the page. Runtime data fields are omitted for clarity.
@misc{argus,
title={Multi-Agent Forensic Reasoning for Generalizable Deepfake Video Detection},
author={Xuechao Zou and Shun Zhang and Kai Li and Yi Zhou and Xinyu Sun and Yuhui Chen and Zhe Wu and Congyan Lang and Junliang Xing},
year={2026},
eprint={2608.06865},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2608.06865},
}