GANzilla: User-Driven Direction Discovery in Generative Adversarial Networks

Abstract

Figure 1: GANzilla is a tool that allows users to discover editing directions in Generative Adversarial Networks (GAN) via iterative scatter/gather interactions-a user-driven approach that complements many existing algorithm-driven methods. (a) A user starts by highlighting a region of interest (an optional step). (b) Based on the highlight (if there is), directions are sampled and clustered, each shown as an image edited by that direction. The user can gather clusters by selecting thumbnail images (indicated by a red border). (c) The user can see all the directions of the gathered clusters and (d) scatter them into new clusters. (e) The user can go back-and-forth across iterations to explore alternate choices of scatter/gather. (f) The user can test a selected direction (red border in c) on other images with individual sliders controlling the strength to apply the direction. (g) The user can bookmark directions that meet their editing goals.

GANzilla: User-Driven Direction Discovery in Generative Adversarial Networks

“…StyleCLIP accepts a textual description of the direction and finds it using the CLIP model [31,33]. GANzilla allows users to discover directions via iterative scatter/gather interactions and complements other direction discovery methods [10]. GANzilla is a 'complementary' solution to the algorithm-driven discovery methods.…”

“…Baseline Methods. We quantitatively compared GANravel with four state of the art methods for editing human face user study: InterFaceGAN [37], GANSpace [15], StyleFlow [1] and GANzilla [10]. GANSpace and StyleFlow are unsupervised direction discovery methods and InterFaceGAN is a supervised direction discovery method that leverages classifiers.…”

“…10 https://github.com/harskish/ganspace11 https://github.com/RameenAbdal/StyleFlow Coarse-Grained Glasses Smile Age interfaceGAN .74 ± .11 .78 ± .13 .60 ± .21 GANspace .65 ± .21 .76 ± .15 .42 ± .34 StyleFlow .55 ± .24 .69 ± .15 .48 ± .40 GANzilla .58 ± .22 .71 ± .22 .45 ± .39 GANravel .84 ± .19 .86 ± .18 .67 ± .23 Facial identity metrics of GANravel with baselines IFG, GS, and SF for wearing glasses, smiling, and increasing age tasks. Higher values indicate higher disentanglement.…”