In Russia, Image Generation for AI Training Is Being Put on an Assembly Line
Russian researchers are tackling the data shortage for neural network training by building a factory for synthetic image generation.
At Tomsk State University of Control Systems and Radioelectronics, researchers are developing a pipeline for generating synthetic images to train neural networks. The project, supported by the Russian Science Foundation, is intended to help computer vision systems avoid “going blind” when operating in real-world conditions.
A Shortage of Images
According to the university, the project aims to address the lack of high-quality data for training neural networks. In many applied fields, collecting real images and labeling them requires significant time and resources and can sometimes be impossible — for example, when analyzing emergency situations, rare events, or hazardous scenarios.
As explained by Roman Kulshin, a senior lecturer at the department, the proposed approach is based on creating a digital model of an object or process, from which images are generated automatically. These datasets are then refined step by step to make them as close as possible to real-world shooting conditions.
A Factory of Images
The image “factory” under development is designed to be universal. The technology could be used in industrial quality control, transport and autonomous systems, security systems, medicine, and robotics — essentially anywhere computer vision is in demand but access to large volumes of photographs is limited.
The researchers have already developed individual methods for generating synthetic images using 3D models and game engines. In 2026, the team plans to combine them into a full-fledged pipeline and run a series of experiments. The launch of the pipeline is scheduled for 2027.
