A popular method for anomaly detection is to use the generator of an\nadversarial network to formulate anomaly scores over reconstruction loss of\ninput. Due to the rare occurrence of anomalies, optimizing such networks can be\na cumbersome task. Another possible approach is to use both generator and\ndiscriminator for anomaly detection. However, attributed to the involvement of\nadversarial training, this model is often unstable in a way that the\nperformance fluctuates drastically with each training step. In this study, we\npropose a framework that effectively generates stable results across a wide\nrange of training steps and allows us to use both the generator and the\ndiscriminator of an adversarial model for efficient and robust anomaly\ndetection. Our approach transforms the fundamental role of a discriminator from\nidentifying real and fake data to distinguishing between good and bad quality\nreconstructions. To this end, we prepare training examples for the good quality\nreconstruction by employing the current generator, whereas poor quality\nexamples are obtained by utilizing an old state of the same generator. This\nway, the discriminator learns to detect subtle distortions that often appear in\nreconstructions of the anomaly inputs. Extensive experiments performed on\nCaltech-256 and MNIST image datasets for novelty detection show superior\nresults. Furthermore, on UCSD Ped2 video dataset for anomaly detection, our\nmodel achieves a frame-level AUC of 98.1%, surpassing recent state-of-the-art\nmethods.\n