Latest Breakthrough in Diffusion Model and its Applications

No Thumbnail Available

Files

URL

Journal Title

Journal ISSN

Volume Title

Sähkötekniikan korkeakoulu | Bachelor's thesis
Electronic archive copy is available locally at the Harald Herlin Learning Centre. The staff of Aalto University has access to the electronic bachelor's theses by logging into Aaltodoc with their personal Aalto user ID. Read more about the availability of the bachelor's theses.

Date

2024-04-26

Department

Major/Subject

Digital System and Design

Mcode

ELEC3056

Degree programme

Aalto Bachelor's Programme in Science and Technology

Language

en

Pages

26+4

Series

Abstract

The Diffusion Probabilistic Model (DPM; Sohl-Dickstein et al., 2015), first pro- pose the concept of the diffusion model, which restores data from data that is distorted gradually until it becomes pure noise. After various research works, diffusion models not only are able to restore data but also generate data, since then diffusion models started to stand out from other generative models especially in the realm of image generation. However, it still faces challenges such as the extensive computational demands in sample production and uneven, ineffective training. And with the challenges, diffusion based generative model is still the dominant model in image generation applications. Although diffusion based image generation models are capable of generating high quality images, the generated data still have shortcomings such as implausibility, and low aesthetic quality. This thesis aims to address these challenges by exploring state-of-the-art techniques for improving generative diffusion models. Specifically, it will focus on enhancements in sampling efficiency, framework innovations, refined training strategies, and the application of the model in conditional image generation.

Description

Supervisor

Ylirsiku, Salu

Thesis advisor

Haitsiukevich, Katsiaryna

Keywords

diffusion model, machine learning, reinforcement learning

Other note

Citation