Огляд латентних дифузійних моделей

Abstract

In recent years, algorithms for solving a wide variety of generative modeling tasks, such as image generation, super-resolution, inpainting, editing, have evolved rapidly. Diffusion models raising the bar to a new level in these fields. Latent diffusion model is an extension of the classical diffusion model and takes into account the underlying structure and dependencies in an image to generate more realistic and efficient image samples with less training and less time to inference. In this work, the concept and architecture of these models were reviewed and described.