A multi-task, contrastive-constraint-based algorithm was proposed to address the challenge of ambiguous boundaries between normal and anomalous features in unsupervised visual anomaly detection. First, highly realistic anomalous images were generated by a controllable anomaly synthesis module and used as multi-task supervisory signals to train the projection layer. Second, in the feature compactness aggregation task, the distances between normal sample feature embeddings were minimized to guide the projection layer toward a compact representation of normal features. In the anomaly sensitivity task, contrastive loss was minimized to increase the projection layer’s focus on anomaly-sensitive regions. In the anomaly feature suppression task, feature similarity loss was minimized to suppress anomalous features during feature propagation. Finally, multi-scale features extracted by the projection layer were fused by an adaptive fusion reconstruction module to enable accurate anomaly detection and localization. Effectiveness validation experiments were performed on multiple public datasets for the proposed algorithm. Results showed that the proposed algorithm achieved an image-level area under the receiver operating characteristic curve (AUROC) of 99.3% on the MVTec dataset.