VerusVIDEO FORENSICS · HUMAN-ANCHORED EVIDENCE
Verus-Insight
Towards Explainable Artifact Detection
in AI-Generated Videos
Learn what to inspect. Verify what is visible.
Ground the final decision in artifact evidence.
From a source label to visible evidence.
A correct real or AI-generated prediction does not guarantee that a detector has recovered the visible artifacts. It may overlook a relevant event, misinterpret what it sees, or reach the right verdict with incomplete evidence.
Verus-Insight separates this process into three complementary capabilities, and learns them through human-anchored full-process supervision.
- 01
Inspection Planning
Determine what to inspect with content-specific criteria.
- 02
Evidence Verification
Verify each criterion against the video and recover concrete artifact evidence.
- 03
Evidence Aggregation
Integrate the verified observations into the final source decision.
Human-anchored, full-process supervision.
Human artifact annotations become outcome-neutral inspection criteria paired with their verification targets. Independently verified normal checks complement the artifact examples. The complete supervision follows Context → Rubric → Audit → Aggregation.
Built on Qwen3-VL-8B-Instruct, Verus-Insight learns through tracking and rubric warm-up, full-process supervised fine-tuning, and GRPO. At inference, the model generates its own inspection criteria and verifies them against the video.
Verus-Bench
Verus-Bench evaluates artifact detection on real and high-quality AI-generated videos. Each artifact annotation records an observation, its temporal extent, and frame-wise bounding boxes.
Generated videos are collected independently of the training and development sets. All artifact annotations undergo manual quality review.
96 real · 96 AI-generated
artifact findings
categories
Ground truth covers all 192 samples. Distributed media cover 166; acquisition references are provided for the remaining 26. Explore the dataset ↗
Explore the 17 artifact categories
Each released finding is counted once by its artifact category. View the source counts ↗
Experimental results.
Paper-reported results for fine-grained artifact analysis, video-level detection, and component ablations.
Video-level detection across four benchmarks
All values are percentages (↑ higher is better). Artifact-Bench T1: balanced accuracy; T2: accuracy; T3: exact-match accuracy. RewardData, ViF-Bench, and Verus-Bench: balanced accuracy. Bold green and underlined cells reproduce the manuscript's best and second-best markings among open-source models.
Verus-Insight leads the evaluated open-source models on RewardData, Verus-Bench, and Artifact-Bench T1 and T3. Performance varies across benchmarks; Skyra leads on ViF-Bench.
Inspection planning and artifact finding
P / R / F1 are precision, recall, and F1 (%; ↑ higher is better). Plan is evaluated on artifact-positive videos; Finding on the full set. Both use one-to-one semantic matching. Ranking marks follow the manuscript's open-source comparison.
On RewardData, Verus-Insight increases Finding F1 from 3.38 for its backbone to 35.42, with gains in both precision and recall. An unmatched reference comparison does not, by itself, establish that a visual observation is unsupported.
What each training component contributes
BA: balanced accuracy. All values are percentages (↑ higher is better). Best and second-best markings follow the manuscript across the listed variants.
The full model has the highest Plan F1 and Finding F1 on both benchmarks. Removing the Stage-II rubric slightly improves BA, while reducing both fine-grained metrics. Stronger source classification does not necessarily imply better artifact understanding.
How to read these results
Balanced accuracy averages recall over the real and AI-generated classes. Plan measures coverage of inspection targets associated with reference artifacts. Finding measures recovery of the artifacts themselves. Predictions are matched one-to-one against the reference annotations; unmatched predicted findings count as false positives in this scoring protocol.
Results are transcribed from the manuscript and have not been recomputed for this release. Reference matching and visual validity are different questions. ViF-Bench source classification and the ViF-CoT-4K finding subset are separate evaluations.
Take a closer look.
Explore the inference code, the 8B model, and the benchmark.
Human annotations & model output.
Fine-grained human reference annotations and Verus-Insight's artifact analysis.
Loading authors, figures, and evaluation tables…