Verus-Insight Moves AI-Video Detection from Prediction to Evidence

Introduction

AI-generated videos are becoming increasingly realistic, making it harder to rely on obvious visual defects for forensic analysis. Many existing detectors treat artifact recognition mainly as a means to predict whether a video is real or synthetic. This approach can produce a correct source label for the wrong reason: a model may exploit incidental correlations, inspect irrelevant content, or miss subtle artifacts that a human viewer can notice.

The paper “Verus-Insight: Towards Explainable Artifact Detection in AI-Generated Videos” addresses this gap by treating artifact detection as a structured reasoning process. The goal is not only to decide where a video came from, but also to connect that decision to observable and verifiable evidence.

Key ideas

• Inspection Planning: The system first determines what should be examined. The target may involve particular objects, regions, frames, or temporal segments. This stage is intended to prevent the detector from treating the entire video as an undifferentiated signal.
• Evidence Verification: After selecting an inspection target, the model checks whether it actually exhibits a generation artifact. This distinction matters because unusual motion, blur, texture, or appearance changes are not automatically evidence of synthetic generation.
• Evidence Aggregation: Verified clues are then combined to support the final source decision. The separation makes the relationship between local observations and video-level classification more explicit.
• Human-anchored full-process supervision: Verus-Insight is trained with supervision grounded in human judgments about perceptible artifacts. Progressive capability training is used to strengthen both fine-grained artifact analysis and overall video detection.
• A dedicated benchmark: Verus-Bench provides fine-grained human annotations of perceptible artifacts. Instead of evaluating only the final real-or-generated label, it tests whether a detector recovers the kinds of evidence people can actually observe.

Why it matters

The central contribution is a change in evaluation philosophy. A detector that returns a probability or a binary label may be useful for large-scale filtering, but it is difficult to audit. In practical settings such as content moderation, media forensics, and authenticity review, analysts may also need to know what the model inspected, which irregularity it identified, and whether that irregularity genuinely supports the conclusion.

The three-part formulation offers a useful way to diagnose failures. If a model overlooks an artifact, the problem may lie in its inspection plan. If it looks at the right place but mistakes a normal visual variation for a synthetic defect, evidence verification is the weak link. If several local observations are valid but the final decision is unstable, evidence aggregation may need improvement. This decomposition is more informative than judging a detector only by its final accuracy.

According to the supplied material, experiments across diverse benchmarks show strong performance for Verus-Insight in both fine-grained artifact analysis and video-level detection. The model also improves recovery of human-perceived artifacts and more reliably turns relevant inspection targets into supported evidence without increasing unsupported artifact claims.

The available material does not include full tables, dataset sizes, or numerical gains, so broad superiority claims should be avoided. Still, the direction is significant: as synthetic videos become harder to distinguish by appearance alone, detection systems may need to move from simply guessing provenance toward presenting evidence that can be inspected and challenged.

Source: Hugging Face Daily Papers