Daily digest — 2026-04-10

A foundations day. Before re-engaging with the 2023–2024 splatting and generative papers, I went back to the work the field is still standing on.

The anchor was the original NeRF paper by Mildenhall et al. Two ideas still feel under-credited five years later: the sinusoidal positional encoding (which is what made coordinate MLPs fit high frequencies at all), and the coarse-to-fine importance sampling along each ray. Every "fast NeRF" paper since has chipped away at one of these, not invented something orthogonal.

NeRF in the Wild is the one I always forget to cite. Martin-Brualla et al. layer per-image appearance and transient embeddings on top of NeRF so it can swallow unconstrained photo collections — variable lighting, moving tourists, occluders. The trick is small but the framing matters: you can decouple the scene from the capture conditions if you give the model a per-frame latent.

Nerfies by Park et al. is the deformable-NeRF entry point. A per-observation deformation field warps the canonical radiance field, so a casually-captured selfie video becomes a 4D reconstruction. It still per-scene-optimises and still bakes ~hours of training per clip — both costs the corpus's 2023+ papers will revisit.

On the classical-geometry side, I re-read Direct Sparse Odometry (Engel et al., 2016) and DeepV2D (Teed & Deng, 2018) to remember why differentiable SfM mattered. DSO is the last great non-learned visual odometry baseline; DeepV2D is the first serious attempt at a learned, differentiable bundle adjustment, and it seeds the DROID-SLAM line I'll get to later this month.

Five papers, ~45 min. The thread for the coming weeks: every 2023+ paper is a reaction to a constraint one of these set.