K-Planes: Explicit Radiance Fields in Space, Time, and Appearance

Fridovich-Keil et al. introduce K-Planes, a factorized explicit representation for radiance fields that generalizes seamlessly from static 3D scenes to dynamic 4D scenes and to scenes with varying appearance. The construction generalizes the planar factorization of TensoRF and its predecessors: a d-dimensional scene is represented by "d choose 2" feature planes, one per pair of axes. For a static 3D scene that is three planes (xy, xz, yz); for a dynamic 4D scene it is six planes (xy, xz, yz, xt, yt, zt).

At a query point, features are bilinearly sampled from each plane and combined by element-wise multiplication into a single feature vector, which is then decoded into density and color. The decoder can be a small MLP, as in prior factorized methods, but the authors show that a linear decoder with a learned global color basis achieves comparable quality, making the entire pipeline a white-box that is easy to inspect and regularize.

The factorization makes adding dimension-specific priors mechanically simple. Temporal smoothness becomes a TV penalty on the time-axis planes; multi-scale spatial detail becomes a multi-resolution stack of spatial planes; static-dynamic decomposition is induced by initializing the time-involving planes to the identity, so the model represents only genuinely time-varying content in those planes. The result is a roughly 1000x memory compression versus a dense 4D grid.

K-Planes outperforms or matches state-of-the-art static and dynamic radiance-field methods across synthetic and real benchmarks while training quickly in pure PyTorch — no custom CUDA. Within this corpus it pairs with Plenoxels (Yu et al., 2021) and Instant-NGP (Müller et al., 2022) as part of the explicit-representation lineage, and with Nerfies/HyperNeRF as part of the dynamic-radiance-field lineage.