UnstableDiffusion

Abstract

Due to the widespread distribution of AI-synthesized images on the Internet, it is increasingly difficult to collect pristine image data sets these days. Therefore, the main purpose of our project is to evaluate the impact of using synthetic images during the training of text-to-image generative models. We trained multiple generations of generative models, each generation using synthetic images produced by the previous generation of models as training data. We also used multiple data sets mixing real images with synthetic images at different scales for training state-of-the-art models and collecting experimental results. The original data sets (real images data sets) we mainly used are ImageNet & CIFAR10, and the image model used in the experiment is a diffusion model - an open source implementation of the Imagen model developed by Google. Through the methods described above, our project finally provides an overview of the potential impact of synthetic images on generative models.