Using Hidden Feature Space of Diffusion Neural Networks for Image Blending Problem

Abstract

In this paper, a new augmentation algorithm based on the idea of blending two images is proposed. The method is developed using state-of-the-art generative diffusion neural networks and can be used to solve the problem of data scarcity, improve the training quality and robustness of neural networks. Flexible customization of the algorithm allows adding snow, rain and other weather conditions to the selected.