LIIF-GAN: Learning Representation With Local Implicit Image Function and GAN for Realistic Images on a Continuous Scale
Jun Seok Kang (University of Science and Technology), Sang Chul Ahn (KIST)*
The 33rd British Machine Vision Conference
Abstract
Recently, the Local Implicit Image Function (LIIF) has been proposed, which can generate continuous 2D image representation for pixel-based images. The continuous image representation can be presented at any resolution. However, the LIIF representation has limited fidelity when presented at higher resolution, resulting in unrealistic images. To solve this problem, simply adding a GAN can make the image realistic, but it degrades the local structure of the image. In this paper, we propose the LIIF-GAN, a novel architecture-based deep model, to generate realistic images at continuous scales while maintaining local image structures. It utilizes a generative adversarial network (GAN) and multiple decoders for encoder features at different levels. We show that the LIIF-GAN can generate a more realistic continuous image representation than previous methods. Furthermore, we show that our new architecture retains the local image structure better than simply using a GAN with the existing architecture. We demonstrate the performance of the proposed method qualitatively and quantitatively through various experiments.Video
Citation
@inproceedings{Kang_2022_BMVC,
author = {Jun Seok Kang and Sang Chul Ahn},
title = {LIIF-GAN: Learning Representation With Local Implicit Image Function and GAN for Realistic Images on a Continuous Scale},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year = {2022},
url = {https://bmvc2022.mpi-inf.mpg.de/0703.pdf}
}
Copyright © 2022
The British Machine Vision Conference is organised by