Dual-Flow Knowledge Distillation with Category-Adaptive Inference for Unsupervised Industrial Anomaly Detection

Our contributions are:
1. We propose a dual-branch scoring architecture jointly training hierarchical distillation and per-level normalizing flows on a shared frozen teacher pyramid, fused via depth-weighted, standardized combination.
2. We include frequency-Aware Attention (FAA) residual block for the student, re-weighting channel responses via learned pooling over grouped frequency-oriented projections, with a cosine+MSE distillation loss.
3. We also introduce category-adaptive policy (selective augmentation, adaptive top-k scoring, soft foreground masking, optional rotation TTA) from simple, interpretable object priors rather than exhaustive tuning.
4. We perform transparent ablation showing that two of four fixes did not generalize and were reverted; and results on all 15 MVTec AD categories: mean image-AUROC 94.07%, mean pixel-AUROC 96.71% are achieved, which can be considered a competitive score.