GANSlider: How Users Control Generative Models for Images using Multiple Sliders with and without Feedforward Information

Abstract

Solution Target b a Example: User exploration path (Not part of UI) StartFigure 1: This paper compares different user interfaces with sliders for interactive control of a generative model for face images (StyleGAN2). We study different numbers of sliders and two designs: (a) Regular sliders and (b) Filmstrip sliders that provide feedforward information via preview images. Participants used these UIs in image reconstruction tasks to generate a given target images. The right part of the figure (c) shows one example of a resulting "exploration path" towards a target (for more examples see Appendix A).

GANzilla: User-Driven Direction Discovery in Generative Adversarial Networks

“…We decided to use sliders in the live-testing area following prior work [1,37]. Dang et al compared the regular sliders and sliders that provide feedforward information ('filmstrips') in a comparative study [9]. Zhang et al enabled users to explore the latent space with a grid-like view of sampled images [46].…”

GANzilla: User-Driven Direction Discovery in Generative Adversarial Networks

“…Even prior to the popularity of GAN, researchers have explored interactions with other generative processes, most of which are centered around generative design (e.g., using topology optimization) via sketching [3,14] or visualization [16] to explore a large design space. Given that most conventional ways of controlling GAN is via the use of sliders, Dang et al conducted a comprehensive comparative study to understand the effects of regular sliders vs. sliders that provide a 'filmstrip' of feedforward information (preview images) [6]. Other researchers go beyond sliders to consider alternative techniques to interact with GAN.…”

Generative 3D Animation Pipelines: Automating Facial Retargeting Workflows

“…An additional use case could be to generate unseen, realistic 3D characters based on an interface aimed to offer customization capabilities to artists, as shown by Dang et al [3]. The variables set by the artist could influence the facial hair length, skin tone and other facial features.…”

“…Another form of customization could be artistic modifications to the model parameters. Dang et al [3] researched the implications of different slider setups, where the styling of an image of a human face could be adjusted. This could open up artistic freedom, as in our proposed tool, a 3D animator would be able to adjust the type of facial expression using a slider.…”

“…In recent years, while developments in artificial intelligence (AI) have led to notorious abuses of deepfakes [6], with increasing introspection, more and more generative machine learning models have been widely applied to the design domain to help users explore the possibilities of unseen design space, such as GANSlider [3].…”