Verus-Insight: Towards Explainable Artifact Detection in AI-Generated Videos

Verus-Insight. Towards Explainable Artifact Detection in AI-Generated Videos. By Zhuoran Zhang and thirteen others, from Peking University, the Kling Team, and three other institutions. Published on the twenty-second of October. Its code repository is public and real: twelve source files.

Video generation keeps getting better, and synthetic videos are increasingly realistic. That raises the demand on detectors that can reliably tell AI-generated video apart. One family of detectors works from artifacts: it identifies generation artifacts and uses them as evidence for deciding where a video came from. The abstract names two ways this goes wrong. A detector may fail to recognise subtle artifacts at all. Or it may reach the correct source decision without faithfully identifying the evidence underneath it, getting the right answer for reasons it cannot actually point to.

The authors systematically investigate how to design and train effective artifact detectors, and reframe artifact detection as a structured process built from three complementary capabilities. Inspection Planning decides what should be examined. Evidence Verification checks whether the inspected content really shows genuine generation artifacts. Evidence Aggregation integrates the verified evidence into the final decision about the video's source. Looking, confirming and concluding become separate jobs.

Verus-Insight is built on that formulation. It is trained with human-anchored full-process supervision, meaning the supervision covers the whole process rather than only the final verdict, and with progressive capability training. The aim is to strengthen two things together: fine-grained detection of individual artifacts, and detection at the level of the whole video. Alongside the model, the authors introduce Verus-Bench, a benchmark carrying fine-grained human annotations of perceptible artifacts, built for evaluating artifact detection itself.

Extensive experiments across diverse benchmarks show Verus-Insight performing strongly on both fine-grained artifact analysis and AI-generated video detection, with substantial gains in recovering the artifacts that humans perceive. Further analyses find that the three capabilities are complementary yet distinct, so each one contributes something the others do not. And Verus-Insight more reliably turns relevant inspection targets into supported artifact evidence, without increasing spurious artifact claims.

That last result is the one aimed squarely at the original complaint. A detector that is right for unfaithful reasons is hard to trust, and one that invents artifacts to justify itself is no better. Splitting detection into planning, verification and aggregation, and supervising every stage against human judgement, is how Verus-Insight tries to make its verdicts rest on evidence a person could also see.