Patch-Based Generalized Neural Texture Upscaling

Abstract

We propose a lightweight neural texture upscaling method. A single generalized MLP is trained on a diverse collection of texture patches and integrated into the texture sampling stage, upscaling low-resolution textures on-the-fly at rendering time. Because the model is generalized rather than overfit to specific content, no per-asset training is required - the same model applies to any existing scene without modifying texture data or the asset pipeline. Storing lower-resolution base textures reduces disk footprint, upscaling at sampling time eliminates the need to store high-resolution textures in VRAM, and sharing a single set of model weights across all textures ensures non-divergent GPU execution. We evaluate our method across multiple texture classes, demonstrating consistent improvement over analytic upscaling methods in perceptual quality.