This is WORK IN PROGRESS, please feel free to contribute via pull request.

We are trying to make NeRF train super fast in pytorch by using pytorch bindings for Instant-NGP.

Current Progress:

  • Code is implemented and runs, but cannot achieve super good results.
  • Per iteration it is ~3.5x faster than the nerf-pytorch code it is built upon.
  • VERY quickly (1 min) gets up to ~20 PSNR (this is MUCH faster, even in iteration count than the normal NeRF).
  • But doesn’t really get above ~25 PSNR even when training for a long time.
  • There is a bug where running the ‘fine’ network doesn’t work, results above are only for coarse network (e.g. N_importance = 0), and speed comparisons also to only course network, this might be the reason the PSNR doesn’t get super high.

How to get running:

  1. Install tiny-cuda-nn (
  2. Download, install dependencies and run this code.

Links to sources:


  • Jonathon Luiten
  • Kangle Deng


Specify --backbone ngp to enable Instant-NGP (already done in configs/fern_ngp.txt).


View Github