A Additional experiments

In this section, we present extended evaluation of all models introduced in the main work. Following [17], we show the assessment of generations quality in terms of additional metrics namely Inception Score [20] and spatial Fréchet Inception Distance [16] - a version of standard FID score but based on spatial image features. Additionally to standard models, we also include evaluation for DAED setup where DAE model is trained only on ImageNet dataset. Figure 9: Generations from different models trained on FashionMNIST dataset. All models were trained with Simple loss function.