The article is devoted to the application of modern methods of generative image compression using variational autoencoders and neural network architectures. Special attention is paid to the analysis of existing approaches to image generation and restoration, as well as a comparative assessment of compression quality in terms of visual perception and metric indicators. The aim of the study is to systematize deep image compression methods and identify the most effective solutions based on the variational Bayesian approach. The paper considers various architectures, including conditional autoencoders and hypernetwork models, as well as methods for evaluating the quality of the data obtained. The main research methods used were the analysis of scientific literature, a comparative experiment on the architectures of generative models and a computational estimation of compression based on metrics. The results of the study showed that the use of variational autoencoders in combination with recurrent and convolutional layers makes it possible to achieve high-quality image recovery with a significant reduction in data volume. The conclusion is made about the prospects of using conditional variational autoencoders in image compression tasks, especially in the presence of additional information (for example, metadata). The presented approaches can be useful for developing efficient systems for storing and transmitting visual data.

Keywords: variational autoencoders, generative models, image compression, deep learning, neural network architectures, data recovery, conditional models