Neural Radiance Fields
Reconstructing a 3D scene from photos by training a network to be the scene — a beautiful idea, and it was replaced in three years.
When not to use it
- NeRF, for anything new. Gaussian Splatting is faster to train, real-time to render, and comparable quality.
- On reflective or transparent surfaces. The representation assumes a point has a colour. A mirror doesn't.
- On dynamic scenes. People, leaves, water — anything that moves breaks it.
- With bad camera poses. This is where projects actually fail, and it looks like the method failed.
Reach for something else instead
- 3D Gaussian Splatting — the successor. Use this.
- Photogrammetry — classical mesh reconstruction. Boring, robust, editable.
- LiDAR scanning — measure it; no reconstruction ambiguity.
- Structure from motion alone — if you want a point cloud, not a renderable scene.
Sources & further reading
- Mildenhall et al. (2020), NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis — the landmark, and positional encoding is the trick.
- Kerbl et al. (2023), 3D Gaussian Splatting for Real-Time Radiance Field Rendering — the replacement, in three years, on rendering speed.
- Tancik et al. (2020), Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains — why coordinate networks need frequency encoding at all.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Starting a new project on NeRF. It was superseded in three years.
- Blaming the method for bad camera poses. COLMAP failing is where most reconstructions die.
- Expecting mirrors to work. The representation can't express "this colour depends on where you stand."
- Missing why Gaussian Splatting won. Rendering speed — an engineering property. The hardware likes rasterising.
At a glance
FieldComputer Vision
The ideatrain a network to be the scene; query a 3D point, get colour and density
The trickpositional encoding, because networks have a spectral bias toward smooth
Replaced by3D Gaussian Splatting, in three years, on rendering speed
The lessona representation the hardware can execute has a structural advantage
DifficultyAdvanced
Flashcards for this concept
Question
—
Answer
—
1 / 4
Often compared with
NeRF vs. Gaussian Splatting — one is an elegant implicit network that takes seconds per frame; the other is a million explicit blobs the GPU rasterises in real time. Three years, and the practical one won.