Anysize GAN: A solution to the image-warping problem

TL;DR

Anysize GAN introduces a flexible architecture that enables GANs to generate images of arbitrary sizes without resizing, addressing dataset diversity and preserving image quality.

Contribution

The paper presents a novel GAN architecture with dynamic resizing and multi-resolution capabilities, allowing training on datasets with varied image sizes without preprocessing.

Findings

Successfully generates images at multiple sizes

Preserves spatial and feature relationships across resolutions

Eliminates need for image resizing in training datasets

Abstract

We propose a new type of General Adversarial Network (GAN) to resolve a common issue with Deep Learning. We develop a novel architecture that can be applied to existing latent vector based GAN structures that allows them to generate on-the-fly images of any size. Existing GAN for image generation requires uniform images of matching dimensions. However, publicly available datasets, such as ImageNet contain thousands of different sizes. Resizing image causes deformations and changing the image data, whereas as our network does not require this preprocessing step. We make significant changes to the standard data loading techniques to enable any size image to be loaded for training. We also modify the network in two ways, by adding multiple inputs and a novel dynamic resizing layer. Finally we make adjustments to the discriminator to work on multiple resolutions. These changes can allow multiple resolution datasets to be trained on without any resizing, if memory allows. We validate our results on the ISIC 2019 skin lesion dataset. We demonstrate our method can successfully generate realistic images at different sizes without issue, preserving and understanding spatial relationships, while maintaining feature relationships. We will release the source codes upon paper acceptance.