Conditioning GAN Without Training Dataset

TL;DR

The paper addresses conditioned image generation without using training data by leveraging a pretrained unconditioned generator (BigGAN) and a pretrained ethnicity classifier. It introduces an Input Generator that, through a reparameterization $z = \mu + \sigma \epsilon$ with $\epsilon \sim \mathcal{N}(0,1)$, supplies latent inputs to the fixed generator, guiding outputs toward a target attribute. Using UTKFace-derived components, the approach achieves competitive Inception Score and FID across full and reduced data regimes while attaining $100\%$ classifier accuracy on validation, demonstrating feasibility of dataset-free conditioning. This method enables practical conditioned GANs in data-scarce settings, though future work should address output diversity via regularization or alternative conditioning signals.

Abstract

Deep learning algorithms have a large number of trainable parameters often with sizes of hundreds of thousands or more. Training this algorithm requires a large amount of training data and generating a sufficiently large dataset for these algorithms is costly\cite{noguchi2019image}. GANs are generative neural networks that use two deep learning networks that are competing with each other. The networks are generator and discriminator networks. The generator tries to generate realistic images which resemble the actual training dataset by approximating the training data distribution and the discriminator is trained to classify images as real or fake(generated)\cite{goodfellow2016nips}. Training these GAN algorithms also requires a large amount of training dataset\cite{noguchi2019image}. In this study, the aim is to address the question, "Given an unconditioned pretrained generator network and a pretrained classifier, is it feasible to develop a conditioned generator without relying on any training dataset?" The paper begins with a general introduction to the problem. The subsequent sections are structured as follows: Section 2 provides background information on the problem. Section 3 reviews relevant literature on the topic. Section 4 outlines the methodology employed in this study. Section 5 presents the experimental results. Section 6 discusses the findings and proposes potential future research directions. Finally, Section 7 offers concluding remarks. The implementation can be accessed \href{https://github.com/kidist-amde/BigGAN-PyTorch}{here}.

Figures (2)

• Figure 1: Conditional GAN
• Figure 2: Images generated by conditioning the GAN. Top-left: Conditioned with African ethnicity. Top-right: Conditioned by East Asian ethnicity. Bottom-left: Conditioned by European ethnicity. Bottom-right: Conditioned by South Asian ethnicity.