A framework for unsupervised multi-class visual anomaly detection that fixes the over-generalization of reverse-distillation decoders, where a single decoder trained on many classes starts reconstructing abnormal patches and hides defects. Class-agnostic learnable text prompts run through a frozen CLIP text encoder provide a universal textual anchor of 'normality' that constrains the decoded features away from anomalies, and a gated mixture-of-experts routes distinct patch patterns to specialized experts to cut inter-class interference, yielding strong results on the MVTec AD and VisA industrial benchmarks.
Unsupervised visual anomaly detection trains only on normal images and flags anything that deviates, which matters for industrial defect inspection where anomalies cannot be exhaustively collected. R…