Rex Tsou

MC/WASN Conference Paper Abstract

This study explores the improvements and possibilities of using Diffusion Model for Image Inpainting. We consider the missing parts of the image as an initial normal distribution (such as Gaussian Noise), and progressively generate images that match the original image through the Diffusion Model, thereby achieving Image Inpainting. First, we propose a new training strategy that enables Diffusion Model to effectively learn and simulate the diffusion process of images through textual descriptions, and explore how to restore detailed parts while maintaining the overall consistency of the image. Therefore, this study proposes a method that combines the whole image, text, and local features to obtain high-quality inpainting results. In addition, we also conducted a series of experiments to evaluate the performance of this experiment's methods in various image inpainting tasks. According to the experimental results, compared with previous research, this study's method has a significant advantage in restoring highly complex image details.