Multi-class anomaly detection has been a promising research area. However, most methods focus on increasing backbone parameters or the depth of the network. This study uses multi-texture anomaly detection as an example to validate a lightweight flow-based pipeline called Multi-Scale Siamese Flow (MSSF) with a Multi-level Feature Fusion (MLFF) to fully use extracted shallow and deep features. Besides, a Mixed anomalies synthesis (MAS) method is incorporated into the MSSF and trains our pipeline in a self-supervised manner by designing a novel training loss combining negative log-likelihood with a changeable self-supervised hindering loss. Extensive experiments on real-world texture subsets or texture datasets, including MVTec-AD, KSDD2, MT, and AITEX, indicate the effectiveness of our MSSF. The inference speed surpasses the second fastest method, UniAD, about 2 times. Compared with other cutting-edge methods, the MSSF achieves an effective balance between performance and speed.