Conducting Computer Graphics with Generative Reinforcement Networks
Title: Conducting Computer Graphics with Generative Reinforcement Networks
SNIC Project: Berzelius-2022-83
Project Type: LiU Berzelius
Principal Investigator: Lida Huang <lida.huang@dsv.su.se>
Affiliation: Stockholms universitet
Duration: 2022-07-01 – 2023-01-01
Classification: 10207
Keywords:

Abstract

Human world has many materials to teach people how to conduct a digital drawing on computer. Human can learn after they watch the tutorials on drawing processes, including conceiving ideas, drafting, coloring. If the computer can learn the way how human complete drawing by itself, it will be able to learn from many drawing tutorials in an unsupervised manner. Current technologies have been able to generate images via generative adversarial networks (GAN) in one step. However, it cannot understand the drawing processes of human. In this paper, we propose a novel method called generative reinforcement exploration (GREX) to encourage exploration in reinforcement learning via introducing an intrinsic reward output from a generative adversarial network, where the generator provides fake samples of states that help discriminator identify those less frequently visited states. We will show how to teach computers to paint like human painters. We will implement experiments to demonstrate that excellent visual effects can be achieved using the processes of human drawing including conceiving ideas, drafting, and coloring. We especially hope to continue the current resource, as there are many deep learning frameworks in the original database.