Daily digest — 2026-04-15

Today was the "fast NeRF" arc. After yesterday's foundations re-read, the question was: how did the field get from hours-per-scene to minutes-per-scene between 2021 and 2022, and which of those tricks survived the splatting era?

Mip-NeRF (Barron et al., 2021) is the aliasing-aware revision of NeRF: rays become cones, point samples become integrated positional encodings over conical frustums. Useful on multiscale captures. Mip-NeRF 360 extends it to unbounded scenes via a non-linear contraction of space and a proposal-network sampler. The 360 benchmark this paper introduced is still the splatting community's default scene set.

PlenOctrees (Yu et al., 2021) bakes a trained NeRF into a sparse octree of spherical-harmonic coefficients for real-time rendering. It traded training-time flexibility for inference speed — a pattern Plenoxels then pushed further by skipping the MLP entirely. Plenoxels is a remarkable paper to read in 2026 because everything it argues — "you don't need an MLP, you need positional features and a good optimisation prior" — became conventional wisdom. The svox2 reference implementation is still the cleanest place to read that argument as code.

The cleanest synthesis of these ideas is Instant-NGP (Müller et al., 2022): a multiresolution hash grid plus a tiny MLP, trained from scratch in seconds on a single GPU. The NVlabs/instant-ngp reference codebase has been ported into more downstream systems than any other in this corpus. Instant-NGP is what 3D Gaussian Splatting had to beat on quality-at-FPS — and what it then did beat, on training-time fidelity at 1080p.

Five papers, ~40 min. Tomorrow: surfaces.