A SURVEY ON GAN INVERSION APPROACHES FOR IMAGE ATTRIBUTE MANIPULATION

PhD Qualifying Examination


Title: "A SURVEY ON GAN INVERSION APPROACHES FOR IMAGE ATTRIBUTE MANIPULATION"

by

Mr. Tengfei WANG


Abstract:

With the rapid advance of generative adversarial networks (GANs) in image 
synthesis, there have been numerous attempts to apply pre-trained GAN 
models (e.g., StyleGAN) for high-quality semantic image editing. The key 
idea for the real image attribute manipulation is to combine GAN inversion 
and latent space exploration. GAN inversion aims in embedding a real image 
into the latent space of a pre-trained generator. The latent space of GANs 
often has semantically meaningful directions under the vector arithmetic, 
e.g., viewpoint and objects appearance. Exploring these directions enables 
diverse attribute editing operations.

In this survey, we give a comprehensive review of GAN inversion-based 
image attribute manipulation. We first briefly introduce the 
state-of-the-art GAN architectures. We then introduce various methods of 
GAN inversion, which can be categorized into optimization-based, 
learning-based, and hybrid approaches. Next, supervised and un-supervised 
latent direction exploration approaches are presented. Their limitations 
are also analyzed. In the end, we conclude this survey by summarizing 
several future research directions.


Date:			Friday, 9 July 2021

Time:                  	11:00am - 1:00pm

Zoom meeting:
https://hkust.zoom.us/j/97353348345?pwd=NGRBVXlnNHFBajNZSmxuaGpzbFd0UT09

Committee Members:	Dr. Qifeng Chen (Supervisor)
 			Dr. Dan Xu (Chairperson)
 			Dr. Xiaojuan Ma
 			Prof. Chiew-Lan Tai


**** ALL are Welcome ****