Boosting explainability of general visual inspection with a multihead vision-language model

We present a multihead vision-language model for general visual inspection that jointly generates natural-language rationales for defects and their bounding-box localizations in a one-shot setting using a single in-context exemplar. Building on ViP-LLaVA, our architecture incorporates a box head trained with L1+(1−GIoU) loss and a count head trained with L1 loss while retaining the lm head for textual decision criteria. Evaluation on MVTec AD shows that the approach maintains high inspection F1 scores (0.950→0.932) but exhibits reduced pixel-level AUROC (0.730→0.646). To quantify localization traceability, we additionally report oracle best-box AUROC (and best IoU) computed from the predicted bounding boxes, indicating that at least one plausible localization is often produced even when the union-based pixel AUROC declines. We analyze common failure modes including formatting errors, coordinate drift, and count-box mismatches. Our training insights reveal when multihead learning enhances inspection traceability through explicit explanations and localizations, even when pixel-level performance metrics decline.