Content-Aware Pixel Art Rendering on Pixels of Multiple Shapes

Abstract

Pixel art is a well studied art form that arose from technical limitations on computing hardware in the early 1980s. Although the discipline itself is often associated with video games, standalone character and landscape portraits in the pixel art style are also popular. Characterized by a deliberately limited resolution and colour palette, pixel art is as an artistic exercise in the conveyance of visual information with a limited number of samples, while avoiding certain unpleasant visual artifacts. In this thesis, we present a first solution to a novel problem in computer graphics: how do we render images in the pixel art style on other tilings of the plane besides the usual squares, all while respecting image features?

We formulate the non-square (or "any-shape") pixel art rendering task as an energy minimization problem over tile-shaped filter supports, given a conventional raster image and geometric tiling data as input. We compute tile energy gradients via rasterization of the tiling geometry; using this information, we evolve an optimal filter support shape while imposing geometric constraints to balance between distortion and feature clarity. We then demonstrate that our method produces images with superior qualitative and quantitative properties in comparison with naive methods.

Our program can compute finished images in seconds, and allows the user to watch the pixel art evolve in real time. We also provide some basic stylization and interaction features for artists, such as k-means colour quantization, colour palette generation in a perceptually uniform colour space, and brush-based vertex manipulation to adjust the shapes of the filter supports. This method has the potential to be useful in several artistic contexts, such as the creation of highly stylized portraiture and landscapes, and authoring of image and video for real hardware displays that use non-square pixels.