Efficient Physically-based Retexturing in Image-space

Abstract

Texture plays a key role in the visual appearance of 3D objects in rendered images. In production, artists often spend a significant amount of time iterating on textures and re-rendering scenes to achieve the desired look. This workflow is generally time-consuming, and the cost increases with detailed geometry, challenging lighting, and material properties. While image-space editing techniques offer an alternative, editing on complex global illumination and fine-grained appearance details during retexturing are highly challenging, and an accurate and efficient solution is yet to be found. We propose a method that can efficiently render many versions of the same image with different textures applied to an object. We avoid expensive re-tracing of light paths by aggregating appropriate auxiliary information (e.g., radiance, UV coordinates) during a single initial rendering pass. Retexturing can then be done as an image-based post-process for arbitrarily many new textures. Our method generates photorealistic results with complex indirect illumination across a wide range of scenes within a few seconds, even at 4K resolution, and naturally extends to video sequences.