AI Artist with Reinforcement Learning and Graph Neural Network

Abstract

Recently, researchers have had much progress in the field of AI painting, inventing advanced machine learning models like Stable Diffusion. However, past methods mostly neglect the clients’ customized requests and need for humanized service. Facing this weakness, this paper constructs an AI to receive text input and to literally paint the image in a humanized way with a customized style. It analyzes the clients’ style requests and breaks down the need into two parts, the painting style, and the painting technique. After generating the initial image with Stable Diffusion models, this AI pre-process it with graph neural networks to satisfy the need for customized painting styles. Then, Reinforcement Learning is implemented to train the AI to paint the image like human painters with distinct strokes. In this step, the strokes can be modified to mimic various painting techniques like oil painting, pencil, etc. By two aspects of the style, this AI artist provides clients with a personalized painting with and a video of the full painting process.