Make Your Own Sprites

Pixel art is a unique style with the appearance of low resolution images. In this paper, we propose data-driven pixelization method that can produce sharp and crisp cell effects controllable sizes. Our approach overcomes limitation existing learning-based methods in size control by introducing reference pixel to explicitly regularize structure. particular, structure features are used as an auxiliary input for process, measuring similarity between generated result art. Furthermore, disentangle process into specific cell-aware aliasing-aware stages, mitigating ambiguities joint learning size, aliasing effect, color assignment. To train our model, construct dedicated dataset augment it different sizes degrees anti-aliasing effects. Extensive experiments demonstrate its superior performance over state-of-the-arts terms sharpness perceptual expressiveness. We also show promising results video game first time. Code available at https://github.com/WuZongWei6/Pixelization.