A new approach to generative art using deep reinforcement learning algorithms

Generative art encompasses the creation of artworks through algorithms and computational processes, allowing for the exploration of novel visual aesthetics. However, one significant drawback in this field is the reliance on simplistic techniques that limit the agent’s ability to produce complex, high-quality art. This study addresses this limitation by introducing an innovative computerized art architecture that trains a media painting model using deep reinforcement learning techniques. The model's goal is to generate a desired result using fundamental drawing behaviors, which may be generalized to diverse painting genres. In this paper, a fine-tuned trust region policy optimization approach based digital paint (art) model is developed. The suggested approach involves teaching a painting agent to produce pictures by selecting a sequence of continuous-valued activities rather than employing simplistic painting techniques. Over the boundaries of the activity areaand the agent's acquired policy, these strokes are collected on a computerized artwork that is governed by designated reference picture. Employing a learning technique, patches that are ever more complex are fed to the agent throughout training, taken from a collection of reference pictures. For successful training of the agent, loss functions like pixel-based and visual loss are generated. According to the study's findings, painting agents can pick up effective painting techniques. In results, the overall performance of our proposed framework is (converge in 98,000 episodes) achieved better outcomes. Furthermore, the agent can create complicated pictures in the specified style at random using a coarse-to-fine improvement method. This technique represents a promising direction for generative art, offering flexibility in developing excellent digital artworks.