NeRFs are getting faster. Much faster. With Thomas Müller’s latest merge into Instant NGP, the rendering rates of NeRFs are increasing by 5X, though I have seen higher speed ups on my GPU. This is all thanks to a deep integration of just-in-time (JIT) compilation in the engine that powers it all: tiny-cuda-nn and its simultaneous 2.0 release.
The performance improvements in 2.0 don’t come from changing how NeRFs work, but rather from how well the system can execute what it already knows. Instead of separately handling neural networks, ray marching, encodings, and back propagation, the new compiler turns it all into a single, streamlined CUDA operation.
The same one click executable files on Windows will still work to let you install Instant NGP. Linux users can pull the latest source code, update the submodules, and recompile following the standard instructions. What’s even more exciting is that any of your saved checkpoints or .ingp files will work without any modification. I have several hundred ingp files if anyone would like to try!
Once you’re up and running, look for the “instant-ngp v2.0dev” title in the GUI, along with a new checkbox labeled “JIT fusion.” This toggle lets you compare the new implementation against the older version in real time.
Beyond performance, Instant NGP 2.0 introduces several quality of life improvements. There are new interpolation modes and live keyframe recording for camera paths, better support for different lens types (including orthographic), and newly added Python bindings for easier programmatic editing. Most camera lenses, except FTheta, now support DLSS as well. And you’re no longer limited to NeRF mode when choosing a lens type; SDF, Image, and Volume modes have all been upgraded to match.
Longtime users should know that the Glow mode feature on rendering has been removed, but for most people this won’t be noticeable.
With this update, NeRFs have been pushed well into real time rendering rates and the training time has only continued to decrease. This is a greatly exciting day that I believe will reopen a lot of avenues for NeRFs. To get started with Instant NGP, head to the NV Labs GitHub page. Instant NGP comes with the standard NVIDIA license.
New to Gaussian Splatting? Start here
