Verus-Insight: Towards Explainable Artifact Detection in AI-Generated Videos
AuthorsZhuoran Zhang, Yang Shi, Qixun Wang, Tengfei Liu, Yuqi Tang, Xuanyu Zhu, Yizhuo Jia, Jingyun Hua, Shixuan Liu, Xianghua Ying, Haotian Wang, Yuanxing Zhang, Pengfei Wan, Wenjing Yang
AffiliationsPeking University · Kling Team · HKUST (GZ) · Fudan University · Tsinghua University
Verus-Insight aims to detect AI-generated videos more reliably by finding and verifying the visual artifacts people can actually see.
Key results
Full-process training data combined Verus-Annotation and RewardData.
Benchmark includes real and AI-generated videos with human-annotated artifacts.
Verus-Insight video-level detection score.
Fine-grained artifact recovery score; the Qwen3-VL-8B backbone scores 3.38.
Rate at which covered inspection targets become correctly recovered artifact findings.
What the paper found
Verus-Insight treats AI-generated video detection as an evidence audit, not just a real-or-fake classification. Built on Qwen3-VL-8B-Instruct, it first plans video-specific inspections, then verifies each candidate artifact with spatial and temporal evidence, and finally aggregates confirmed findings into a verdict. Training combines region-tracking and rubric-planning warm-ups, full-process supervised fine-tuning, and GRPO reinforcement learning, using 8,411 training instances. The researchers also introduce Verus-Bench, a 192-video benchmark with human-annotated artifact descriptions and locations; its generated-video sources include Kling and Google Veo 3. On RewardData, Verus-Insight reaches 94.79% balanced accuracy and 35.42 Finding F1, compared with 48.81% and 3.38 for its Qwen3-VL-8B backbone. It converts covered inspection targets into findings at 75.58%, showing that the gains come from better verification, not simply making more artifact claims. The study compares its approach with systems including Google’s Gemini 3.6 Flash and OpenAI’s GPT-5.5, and finds that a correct final verdict alone does not guarantee accurate artifact explanations.
Original abstract
Recent advances in video generation have made synthetic videos increasingly realistic, placing greater demands on reliable AI-generated video detection. Existing artifact-based detectors typically identify generation artifacts as evidence for source prediction, yet they may either fail to recognize subtle artifacts or reach the correct source decision without faithfully identifying the evidence underlying it. In this work, we systematically investigate how to design and train effective artifact detectors and formulate artifact detection as a structured process comprising three complementary capabilities: Inspection Planning, which determines what should be examined; Evidence Verification, which verifies whether the inspected content exhibits genuine generation artifacts; and Evidence Aggregation, which integrates the verified evidence into the final source decision. Building on this formulation, we introduce Verus-Insight, trained with human-anchored full-process supervision and progressive capability training to strengthen both fine-grained artifact detection and video-level detection. We further introduce Verus-Bench, a benchmark with fine-grained human annotations of perceptible artifacts for evaluating artifact detection. Extensive experiments across diverse benchmarks demonstrate that Verus-Insight achieves strong performance in both fine-grained artifact analysis and AI-generated video detection, with substantial gains in recovering human-perceived artifacts. Further analyses reveal that the three capabilities are complementary yet distinct, and that Verus-Insight more reliably converts relevant inspection targets into supported artifact evidence without increasing spurious artifact claims.
Read the original paperMore in Computer Vision
Browse all 73 papers →Foundation4D: Towards a unified perception and reconstruction model for autonomous vehicles
NVIDIA, :, Sérgio Agostinho, Maximilian Apodaca, Emmanuel Attia, Guillem Braso, George Dong, Ilya Fedorov, Michele Fenzi, Gang Huang, Jiahui Huang, Shengyu Huang, Bingxin Ke, Laura Leal-Taixé, Kuan-Ting Lee, Mingxuan Liu, Tian Liu, Tim Meinhardt, Aljoša Ošep, Naveen Kumar Rai, Kefei Ren, Cristiano Saltori, Keqiang Wu, Ning Xu, Fei Xue, Jin-Bey Yu, Lei Zhang, Qunjie Zhou
Foundation4D combines multi-camera driving video to estimate scene geometry, detect objects, map occupancy, and reconstruct scenes for simulation.
PhysInsert: Physically Plausible Object Insertion with Geometric Priors and Bidirectional Object-Scene Interactions
Yue Gong, Le Zhang, Rui Chen, Song Chang, Xiaochao Qu, Luoqi Liu, Ting Liu, Yebin Liu
PhysInsert aims to make inserted objects fit their scenes by adapting their pose, perspective, and lighting using geometric and object-scene cues.
WorldCast: World-Consistent Character Video Generation with Large-Angle Camera Control
Lei Cao, Yu Liu, Wenhua Liao, Peng Zhang, Fei Wang, Hao Zhang, Jiajun Le, Daiguo Zhou, Jian Luan, Jiayi Ma
WorldCast aims to keep generated characters visually consistent as the camera makes large, complex moves around them.