Verus-Insight: Towards Explainable Artifact Detection in AI-Generated Videos
As AI-generated videos approach photorealism, detectors must do more than classify real versus fake; they need to surface the specific visual artifacts that ground the decision. Verus-Insight decomposes artifact detection into three explicit capabilities—deciding what to inspect, verifying whether those inspections reveal genuine generation artifacts, and aggregating the verified evidence into a final verdict—and trains an MLLM with human-anchored, full-process supervision to strengthen each stage. The result is a detection system that improves not only video-level accuracy but also the fidelity of fine-grained artifact recovery, supported by a new benchmark with dense human annotations of spatiotemporal artifacts.
The paper makes a compelling case that reliable AI-generated video detection requires explicit, intermediate artifact verification rather than end-to-end classification alone. Verus-Insight delivers strong empirical gains on both video-level balanced accuracy and fine-grained Plan/Finding F1, and the three-stage decomposition is well motivated. However, the evaluation depends heavily on an automated Gemini 3.1 Pro judge for semantic matching, and Verus-Bench's small size (192 videos) limits the breadth of fine-grained claims; baseline comparisons are also complicated by divergent training data and backbone sizes.
The decomposition into Inspection Planning, Evidence Verification, and Evidence Aggregation is sharply motivated by observed failure modes where detectors miss relevant content, misinterpret evidence, or classify correctly for the wrong reasons. Ablations show that rubric warm-up combined with explicit rubric generation yields the highest Plan and Finding F1, and that region tracking with explicit spatiotemporal grounding improves verification. The finding that stronger video-level accuracy does not guarantee better artifact recovery is an important insight that justifies process-level evaluation.
First, the fine-grained metrics rely on automated extraction and semantic matching via Gemini 3.1 Pro, which introduces opacity and potential judge bias; the authors acknowledge that stored automated matches do not constitute independent human agreement measurements. Verus-Bench is also small, with only 192 videos and 326 annotated artifacts, so broad generalization claims should be treated cautiously.
Second, the RL reward remains gated by source-verdict correctness, meaning a model can receive partial credit even when it misses most artifacts, which may undercut the goal of faithful recovery. Baseline comparisons are further complicated by divergent backbone sizes and training sets, making it difficult to attribute gains solely to the three-stage design.
The experimental evidence supports the core claim that explicit rubric-based planning and fine-grained perceptual warm-up improve artifact recovery: ablations show consistent gains from each component, and the conversion-rate analysis demonstrates that Verus-Insight turns planned inspections into findings more reliably than its backbone. Comparisons to prior specialized detectors are generally fair in highlighting that artifact coverage and source accuracy are not synonymous, though the magnitude of the gap is partly a function of supervision density. Notably, VideoVeritas achieves high video-level balanced accuracy yet almost no fine-grained artifact recovery, underscoring the paper's central argument that verdict correctness is an insufficient proxy for diagnostic quality.
The authors release code, model weights, prompts, and detailed hyperparameters, which is exemplary. However, independent reproduction faces practical barriers: the training data requires extensive manual artifact annotation and verified-normal auditing, the RL reward and evaluation metrics depend on calls to a proprietary model (Gemini 3.1 Pro), and rubric-generation and normal-verification pipelines may fluctuate across LLM versions. The prompts are fully documented, meaning the protocol is transparent, but the cumulative cost and non-determinism of the annotation and judging pipeline make exact replication difficult.
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.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.