A Two-Level Simplification Framework for Efficient Low-Poly Image Abstraction from AI-Generated Images

Low-polygon image representation is commonly employed into simplify an image to reduced geometrical data. This paper proposes a method to generate a low-poly image from an AI-generated image. The proposed framework is based on a two-level simplification technique to achieve a considerable reduction in resolution while maintaining good visual quality. First, significant feature points are extracted from the segmented portions of the AI-generated image to obtain essential structural information. Subsequently, hexagonal grid-based sampling, which allows for the identification of crucial seed points while respecting significant visual elements, is employed. The Low-poly style is achieved through Delaunay triangulation, and the colors are taken directly from the original AI image to ensure visual consistency. Furthermore, we have compared the performance of entropy and saliency maps when used in the selection process for the hexagonal grid. We present experimental results demonstrating a pixel-to-point reduction of over 99.1%, as well as the ability to compress high-resolution images into a few simple points in a way that powerfully preserves perceptual integrity. Both qualitative and quantitative analyses were conducted on the AI-generated images at multiple resolutions, observing the sensitivity and scalability of the method for generating low-poly images.