More about HKUST
Image Editing via Generative Models: A Survey
PhD Qualifying Examination Title: "Image Editing via Generative Models: A Survey" by Mr. Jiapeng ZHU Abstract: Image editing has witnessed significant advancements with the advent of deep learning techniques, particularly generative models. In this paper, we present a comprehensive study of three popular generative models - Generative Adversarial Networks (GANs), diffusion models, and autoregressive models, and their applications in image editing tasks. We introduce the basics of these models and discuss their pros and cons. GANs excel in generating high-quality single-object images while maintaining a desirable continuous latent space. Diffusion models use step-by-step denoising to generate images, which have garnered significant interest, particularly in the domain of large-scale text-to-image models. Autoregressive models synthesize images of arbitrary sizes and resolutions pixel-by-pixel by modeling the conditional distribution of each pixel given its predecessors. We compare these models and discuss their strengths and weaknesses in different scenarios. We also talk about current challenges and future research in image editing with generative models. Our study aims to help researchers and practitioners understand and apply these models in image editing. Date: Friday, 24 January 2025 Time: 2:00pm - 4:00pm Venue: Room 3494 Lifts 25/26 Committee Members: Dr. Qifeng Chen (Supervisor) Prof. Dit-Yan Yeung (Chairperson) Dr. Dan Xu Dr. Wenhan Luo (AMC)