Генерация текстур для трёхмерных моделей

This thesis presents a novel approach to the automatic generation of 3D textures using text-to-image diffusion models. The exponential growth of industries such as gaming, animation, and industrial design has highlighted the necessity for high-quality 3D content, particularly textures that enhance the realism of 3D models. The traditional methods for texture generation are both labour-intensive and time-consuming, which has motivated the development of automated solutions. The proposed method employs Score Distillation Sampling (SDS) to transfer the capabilities of pre-trained 2D diffusion models to 3D texture generation. By employing physically-based rendering (PBR) principles, diverse and realistic textures are generated that align with user-defined text prompts. The method employs a differentiable renderer to iteratively optimise material parameters. This ensures that the generated textures are coherent with the 3D mesh and respond accurately to lighting conditions. The efficacy of our approach is demonstrated through a series of experiments. This work contributes to the field of generative models by providing an automated solution that opens new possibilities for realistic and diverse 3D texture generation across various domains.