NeRF & Radiance Fields

NeRF (Neural Radiance Fields, Mildenhall et al., ECCV 2020) was the radiance-field workflow that mattered from 2020 to 2023; in 2024 Gaussian Splatting largely displaced it for production captures because splats are 10–100x faster to render. NeRFs still win for view-dependent transparency, atmospheric scattering, and very-high-quality offline renders, and the tooling lives on inside the same frameworks. Pairs with Photogrammetry Software (capture / pose solving), Mesh Processing for mesh extraction, and 3D / WebGL for display.

Training frameworks (OSS)

• ★ Nerfstudio — free, OSS (Apache-2). The de-facto NeRF (and now splat) research framework; pluggable methods, real-time web viewer, COLMAP-friendly. The starting point in 2026.
• ★ Instant-NGP (NVIDIA) — free, OSS (NVIDIA license, non-commercial-friendly). The breakthrough "NeRF in seconds" paper; still the fastest pure-NeRF trainer for many cases. CUDA-only.
• Nerfacto — Nerfstudio's first-party balanced NeRF method; reasonable defaults.
• Mip-NeRF / Mip-NeRF 360 — Google research; anti-aliasing / unbounded scenes; reference impl in Jax.
• Plenoxels — Berkeley research; voxel-grid alternative, no neural net; very fast.
• TensoRF / KiloNeRF / Nerfacc — research variants; available as Nerfstudio plugins or standalone.
• Zip-NeRF / MERF / SMERF — Google research; extreme-quality + browser-streaming NeRFs.

Mesh extraction from NeRFs

• Nerfstudio ns-export poisson / tsdf — extract a mesh from a trained NeRF.
• Marching cubes on the density field — built into Instant-NGP, Nerfacto.
• NeuS / VolSDF — research methods that train an SDF directly for cleaner mesh extraction.

Hosted / paid

• Luma AI — paid; pioneered consumer NeRF capture from any phone; now also splats.
• Polycam NeRF — paid Pro tier.
• KIRI Engine — paid Pro tier.

Viewers / display

• Nerfstudio web viewer (viser) — free; live during training and after.
• MERF / SMERF browser viewers — Google research; demo-grade splat-like NeRF rendering in browser.
• WebGL NeRF viewers (community) — exist but largely superseded by Gaussian Splatting viewers.

When NeRF still beats splats in 2026

• Transparency, smoke, fog, glass, fluids — view-dependent volumetrics where splats struggle.
• Reflective / refractive scenes — Mip-NeRF 360 + ZipNeRF still set the bar.
• Offline cinematic renders — when you can wait minutes per frame.
• Research / publishing — NeRF math is more thoroughly studied.

When to skip NeRF and use splats

• Real-time rendering, browser display, mobile playback, editing — see Gaussian Splatting.

Pick this if…

• Default OSS NeRF training: Nerfstudio + Nerfacto.
• Fastest training, OK with NVIDIA-only: Instant-NGP.
• Highest quality unbounded scene: Zip-NeRF / Mip-NeRF 360 (research; expect to compile).
• No-code phone capture, paid: Luma AI or Polycam NeRF.
• Real-time / interactive / web: switch to Gaussian Splatting.