Mip-NeRF Anti-Aliased 3D Rendering

Mip-NeRF renders cones instead of rays so pixels covering different areas stay sharp at every zoom level.

Why Does This Exist?

NeRF samples infinitesimal points along a ray, but a real pixel covers a footprint that grows with distance. Close-ups then shimmer because training saw one scale. Mip-NeRF (2021) casts a cone per pixel and encodes each interval's mean and covariance, so one model serves all zooms. This page covers integrated encoding; speed comes from Instant NGP instead.

Think of It Like This

A flashlight beam versus a laser pointer

A laser samples one dust mote. A flashlight beam lights a widening disc and you judge the whole disc's average color. Judging a wall by laser dots flickers as you step back, while the beam stays stable.

Cones are beams and rays are lasers. The analogy stops at encoding: beams need Gaussian math, not just averaging.

How It Actually Works

Pixel footprint radius at distance $t$ is $r = t \cdot \text{px} / f$ with pixel width $\text{px}$ and focal $f$ . Interval $[t_0, t_1]$ becomes Gaussian with mean $\mu$ and covariance $\Sigma$ along the cone axis. Integrated positional encoding feeds $\mathbb{E}[\sin(2^k x)]$ and $\mathbb{E}[\cos(2^k x)]$ using $\mu$ and $\Sigma$ instead of raw $x$ . Coarse and fine MLPs both consume intervals.

Worked example

Take $f = 800$ , pixel width $1$ , $t = 10$ m. Radius is $r = 10/800 = 0.0125$ m. Interval $[9.9, 10.1]$ m has length $0.2$ m and mean $\mu = 10.0$ m. Encoding frequency $2^5 = 32$ on raw points swings fully across the interval, while the integrated version damps by $\exp(-32^2 \sigma^2 / 2)$ with $\sigma \approx 0.06$ , cutting the term to about $0.16$ of full strength. That damping is the anti-aliasing.

Watch Out For

Feeding points to a Mip-NeRF checkpoint

Sampling points instead of intervals at test time restores shimmer even with trained weights. The symptom is crisp trains but buzzing gravel. Always render with the cone interval code path.

The Quick Version

Mip-NeRF casts cones matching each pixel's true footprint.
Intervals become Gaussians whose encoding damps high frequencies correctly.
One model handles close-ups and wide shots without retraining.
Do not sample points at test time; use intervals or aliasing returns.
Cone radius $t/f$ sets how fast detail must fade with distance.