MMVIAD - Unseen Categories: leaderboard
MMVIAD (Zhao et al., 2026) is rendered 2-second inspection videos that sweep about 120 degrees around Anomaly-ShapeNet objects (48 categories, six structural defect types), asking whether the object is anomalous, which defect it has, what the object is and when the defect is visible. The score is the mean of the three accuracies and the visible-time IoU, in percent, on the Unseen test split (1,062 clips of 12 categories held out from the training split); higher is better.
Metric: Mean of three accuracies and visible-time mIoU (%; MMVIAD-Unseen test split, 1,062 clips of 12 object categories held out from training, 4,248 questions). Source: arxiv.org. Saturation forecast: Around February 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 63.9 | #91 |
| 2 | GPT-5.4 | 53.5 | #121 |
| 3 | GPT-5.4 Mini | 53.5 | #389 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mmviad-unseen-categories · How It Works · Data refreshed daily, snapshot 2026-11-12.