Several published solutions exist for the automatization of seismic facies labeling. We suggest an approach that applies tools from deep learning and semantic image segmentation, such as specific UNet-based neural network structure, total variation (TV) loss, pseudolabels, as well as domain-specific attributes and a novel image-warping augmentation technique. We use a patch-based training and prediction approach, and at the prediction stage, the facies labels for the test cube are collected with the patch overlap and with the averaging of the predictions. When applied to two open-source labeled seismic cubes, the algorithm demonstrates superior performance compared with the published results with regard to several metrics computed, such as accuracy, intersection-over-union, and F1 score. We compare the model predictions with and without the domain-specific augmentation and the pseudolabel approach with the metrics suggesting that the augmentation and the pseudolabels provide an increase in the model’s performance. Our method provides smoother labels due to the use of TV loss and pseudolabels, which is proved by the visual observation of the predictions of the final model in comparison with the baseline raw UNet model results.