USING NEURAL NETWORK METHODS FOR GENERATION AND RECONSTRUCTION OF 3D IMAGES

Abstract

This article examines modern neural network methods for generating and reconstructing three-dimensional images, applied in computer graphics, virtual and augmented reality, medicine, and digitalization of objects. Particular attention is paid to implicit neural network representations (such as distance and occupancy functions), neural radial density fields (Neural Radiance Fields , NeRF ), as well as approaches using pretrained diffusion models for text- driven 3D content generation. A comparative analysis of architectures, training data requirements, computational complexity, and the quality of the resulting 3D models is conducted. The key advantages and limitations of existing solutions are identified, including issues of scalability, the lack of labeled 3D data, and the accuracy of geometric reconstruction.