Transformational models underlying large language models cope well not only with the task of content generation in natural language but also with multimodal tasks, to which the task of image description can be attributed. The development of systems for solving the Image to Caption task has progressed from convolutional neural networks and encoder-decoder models to basic transformer-based systems using vision language pre-training (VLP). Experimental studies show that pre-trained models significantly improve the performance of transformer models. However, the performance of such models remains an open and important issue. Using pre-trained transformer models, fine-tuning a large language model, and transfer learning will improve the quality of audio-descriptors created for visually impaired people to perceive visual information of visual arts subjects. Nevertheless, including important factual information necessary for people with visual impairments to fully perceive visual information is challenging. Models may miss information about the author of the work, its size, style of execution, and other key characteristics that are essential for the correct description of works of art. These shortcomings indicate the need for further improvement of teaching methods and model tuning to ensure high-quality audio descriptions that meet the strict standards of professional peer review.