Verus-Insight Teaches an AI Video Detector to Find Artifacts and Prove Its Verdicts

Detectors of generated video still explain their verdicts poorly. The team behind Verus-Insight split the task into three separate skills and trained a model that does not simply name a clip's source, but shows the artifacts its decision is based on. Alongside the model, the team released Verus-Bench, a benchmark in which artifacts visible to humans are annotated manually and in fine detail.

Anyone who has tried to catch synthetic media will recognize the problem. Video generators are getting more realistic, and crude defects such as melting fingers or flickering textures are becoming rarer. Existing artifact-based detectors look for such traces and use them as evidence when determining the source. The approach has two weak spots. First, the detector misses subtle artifacts. Second, and more troublesome, it can reach the correct answer — "this was generated by such-and-such model" — without ever pointing to real evidence. The verdict is right, but the justification is fictitious. For moderation and investigations, an answer like that is almost useless.

Three skills instead of a single verdict

The authors describe detection as a structured process built on three complementary abilities:

• Inspection Planning: decide what exactly is worth examining in the video — which regions, objects and segments of motion.
• Evidence Verification: check that the chosen segment really contains a generation artifact, rather than a quirk of filming, compression or editing.
• Evidence Aggregation: combine the confirmed evidence into a final decision about the video's source.

This split is useful because each step can be evaluated separately. When a detector makes a mistake, you can see where: it did not look in the right place, it looked but failed to recognize the artifact, or it recognized it but put the picture together incorrectly. Further analysis showed that the three skills complement each other but are not reducible to one another, meaning that improving one does not automatically bring gains in the others.

Training with a human in the loop

Verus-Insight was trained with a scheme called human-anchored full-process supervision: human annotation ties the entire process to real observations, from choosing where to inspect to the final decision, rather than just the final label. The authors also used progressive capability training, which strengthens the skills in stages. The goal is twofold: to improve the detection of subtle artifacts and the accuracy of classifying the video as a whole.

Verus-Bench is a contribution in its own right. Most detection datasets provide only a "real or generated" label. This one includes detailed annotation of perceivable artifacts, so it is possible to check whether the model finds what a human sees, rather than merely guessing the answer.

What the experiments showed

By the results across several benchmarks, Verus-Insight demonstrates strong performance both in fine-grained artifact analysis and in detecting generated video. The gain is especially noticeable in recovering artifacts that people notice. The authors' main argument is that the model more reliably turns relevant inspection targets into confirmed evidence, while the number of false artifact claims does not grow. This matters: a detector that accuses everything easily achieves high recall and becomes useless in practice.

Where this is useful

For platforms and newsrooms, the value lies in explainability. A moderator or fact-checker does not need a confidence percentage; they need a pointer: here is the frame, here is the distortion, here is why it indicates generation. The plan–verify–aggregate structure maps well onto how a human analyst works and makes it possible to hand intermediate findings over to that analyst.

For detector developers, there is now a better yardstick than plain accuracy: Verus-Bench makes it possible to check whether a model relies on real evidence or on statistical cues, such as codec traces. For generator developers, the benchmark is also useful as a map of subtle defects that humans can still spot and that are worth eliminating.

The limitation is obvious for the whole field: the set of artifacts is tied to the current generation of generators. With new models, the evidence will change, and the annotation will have to be updated. But the approach itself, in which a verdict must be backed by verified evidence rather than just a correct label, looks like a workable standard for the next generation of detectors.