Abstract
We introduce k-planes, a white-box model for radiance fields in arbitrarydimensions. Our model uses d choose 2 planes to represent a d-dimensionalscene, providing a seamless way to go from static (d=3) to dynamic (d=4)scenes. This planar factorization makes adding dimension-specific priors easy,e.g. temporal smoothness and multi-resolution spatial structure, and induces anatural decomposition of static and dynamic components of a scene. We use alinear feature decoder with a learned color basis that yields similarperformance as a nonlinear black-box MLP decoder. Across a range of syntheticand real, static and dynamic, fixed and varying appearance scenes, k-planesyields competitive and often state-of-the-art reconstruction fidelity with lowmemory usage, achieving 1000x compression over a full 4D grid, and fastoptimization with a pure PyTorch implementation. For video results and code,please see sarafridov.github.io/K-Planes.