A pytorch implementation of instant-ngp, as described in Instant Neural Graphics Primitives with a Multiresolution Hash Encoding.

Note: This repo is far from instant, and is still a work in progress.



  • Fully-fused MLP
    • basic pytorch binding of the original implementation (but only slightly faster than pytorch built-in FP16, which seems to use CUTLASS …)
    • further benchmark
  • HashGrid Encoder
    • basic pytorch CUDA extension
    • fp16 support
    • improve performance
  • Experiments
    • SDF
      • baseline
      • better SDF calculation (especially for non-watertight meshes)
    • NeRF
      • baseline (although much slower)
      • ray marching in CUDA.


  • 1.30:
    • fixed atomicAdd() to use __half2 in HashGrid Encoder’s backward, now the training speed with fp16 is as expected!
  • 1.29:
    • finished an experimental binding of fully-fused MLP.
    • replace SHEncoder with a CUDA implementation.
  • 1.26: add fp16 support for HashGrid Encoder (requires CUDA >= 10 and GPU ARCH >= 70 for now…).


pip install -r requirements.txt

Tested on Ubuntu with torch 1.10 & CUDA 11.3


We use the same data format as instant-ngp, e.g., armadillo and fox.
Please download and put them under ./data.

First time running will take some time to compile the CUDA extensions.

# SDF experiment
bash scripts/

# NeRF experiment
bash scripts/

python data/fox --workspace trial_nerf # fp32 mode
python data/fox --workspace trial_nerf --fp16 # fp16 mode (pytorch amp)
python data/fox --workspace trial_nerf --fp16 --ff # (experimental) fp16 mode + fully-fused MLP


View Github