Generative models for the synthesis of three-dimensional objects

Abstract

The chapter examines modern generative models used for the synthesis of three-dimensional objects and analyzes their potential for automated creation of complex 3D content. Particular attention is paid to neural network approaches capable of generating three-dimensional geometry textual descriptions, images, point clouds, voxel representations, and other forms of input data. The principles of operation of generative adversarial networks, variational autoencoders, diffusion models, and transformer-based architectures are considered in the context of 3D object synthesis. The main stages of the generative pipeline are described, including data preprocessing, feature extraction, latent-space representation, geometry generation, texture formation, and post-processing of the resulting models. The advantages and limitations of different approaches are analyzed with respect to geometric accuracy, visual realism, computational complexity, training requirements, and the possibility of controlling the generated result. Special attention is given to current methods for improving the quality of generated polygonal models, meshes, implicit surfaces, and neural representations. The study demonstrates that generative artificial intelligence significantly reduces the time required for creating three-dimensional content and enables the automation of several stages of traditional 3D modeling. The considered technologies have considerable potential for computer graphics, virtual and augmented reality, computer-aided design, simulation, digital entertainment, education, and engineering applications.