Zip-NeRF: Anti-Aliased Grid-Based Neural Radiance Fields
ICCV 2023 (Oral Presentation, Best Paper Finalist)


Neural Radiance Field training can be accelerated through the use of grid-based representations in NeRF's learned mapping from spatial coordinates to colors and volumetric density. However, these grid-based approaches lack an explicit understanding of scale and therefore often introduce aliasing, usually in the form of jaggies or missing scene content. Anti-aliasing has previously been addressed by mip-NeRF 360, which reasons about sub-volumes along a cone rather than points along a ray, but this approach is not natively compatible with current grid-based techniques. We show how ideas from rendering and signal processing can be used to construct a technique that combines mip-NeRF 360 and grid-based models such as Instant NGP to yield error rates that are 8%-77% lower than either prior technique, and that trains 24x faster than mip-NeRF 360.


360° Video Flythroughs


When Training When Rendering

We use multisampling to approximate the average NGP feature over a conical frustum, by constructing a 6-sample pattern that exactly matches the frustum's first and second moments. When training, we randomly rotate and flip (along the ray axis) each pattern, and when rendering we deterministically flip and rotate each adjacent pattern by 30 degrees.

XY aliasing

A naive baseline (left) combining mip-NeRF 360 and Instant NGP results in aliasing as the camera moves laterally. Our full method (right) produces prefiltered renderings that do not flicker or shimmer.

Z aliasing

The proposal network used for resampling points along rays in mip-NeRF 360 results in an artifact we refer to as z-aliasing, where foreground content alternately appears and disappears as the camera moves toward or away from scene content. Z-aliasing occurs when the initial set of samples from the proposal network is not dense enough and misses thin structures, such as the chair above. Missed content can not be recovered by later rounds of sampling, since no future samples will be placed at that location along the ray. Our improvements to proposal network supervision result in a prefiltered proposal output that preserves the foreground object for all frames in this sequence. The plots above depict samples along a ray for three rounds of resampling (blue, orange, and green lines), with the y axis showing rendering weight (how much each interval contributes to the final rendered color), as a normalized probability density.



Thanks to Janne Kontkanen, Rick Szeliski, and David Salesin for their comments on the text, and to Ricardo Martin-Brualla, Keunhong Park, Ben Poole, Aleksander Hołyński, Etienne Pot, Kostas Rematas, Daniel Duckworth, Marcos Seefelder, Cardin Moffett, and Peter Zhizhin for their advice and help.

The website template was borrowed from Michaël Gharbi and Ref-NeRF.