MTUSI: An algorithm for generating images by neural networks

Main article: Neural networks (neural networks)

2024: Creating an algorithm to generate images by neural networks

MTUSI has developed an algorithm for generating images by neural networks. The university announced this on January 29, 2024.

In almost all spheres of human life, machine learning methods are increasingly being introduced. At the same time, some of the most effective and often used in practice are methods based on artificial neural networks (NS). They allow you to process images, video and audio data.

Another direction of the HC development is the generation of text and images according to a user-generated request (prompt). The National Assembly learned to create images that are difficult to distinguish from real ones created by man, so they became a powerful working tool in compiling advertising slogans, creating unusual advertising banners. As a result, the task of developing a query generation algorithm that allows generating a high-quality image in reasonable time intervals becomes particularly urgent. Employees of the Department of Mathematical Cybernetics and Information Technologies of MTUSI were able to solve it.

When developing an algorithm for compiling industrial pots that allow obtaining a high-quality image by neural network online services, factors have been identified that lead to incorrect image generation results through natural language, informal requests. A methodology for describing an object, phenomenon or process has been created, neural online services "Masterpiece" and Kandinsky have been selected to test the developed algorithm and determine its effectiveness, - explained Yury Leokhin, Doctor of Technical Sciences, Professor of the Department of MKiIT MTUSI.

The choice of the National Assembly is determined by the fact that they are domestic developments, which are primarily aimed at the Russian-speaking audience.

The results of the experiments showed that the developed algorithm for generating a request to neural network services for generating an image allows you to obtain a more correct result compared to a regular natural language request. This is especially noticeable on queries containing negative particles and prepositions, relative adjectives that assume the representation of an abstract concept, and when using numerals, - said Timur Fathulin, Ph.D., associate professor of the Department of MKiIT MTUSI.

In the future, scientists plan to modify and modernize the developed algorithm taking into account the increasing functionality of neural network services and consider the possibility of using it to generate text content.