Unsupervised Discovery of Disentangled Manifolds in GANs

TL;DR

This paper introduces a framework for discovering interpretable directions in the latent space of pre-trained GANs, enabling attribute editing and better understanding of the generation process.

Contribution

It proposes a method to learn attribute directions in GAN latent space using a centroid loss, applicable to various datasets and models.

Findings

Discovered interpretable attribute directions in GAN latent spaces.

Enabled attribute editing through identified directions.

Improved consistency and smoothness in attribute traversal.

Abstract

As recent generative models can generate photo-realistic images, people seek to understand the mechanism behind the generation process. Interpretable generation process is beneficial to various image editing applications. In this work, we propose a framework to discover interpretable directions in the latent space given arbitrary pre-trained generative adversarial networks. We propose to learn the transformation from prior one-hot vectors representing different attributes to the latent space used by pre-trained models. Furthermore, we apply a centroid loss function to improve consistency and smoothness while traversing through different directions. We demonstrate the efficacy of the proposed framework on a wide range of datasets. The discovered direction vectors are shown to be visually corresponding to various distinct attributes and thus enable attribute editing.