taichi-nerfs

Why we included this project

Neural radiance field work usually means wrestling with hand-written CUDA kernels, and this project is a useful alternative for anyone who would rather not. It reimplements the instant-ngp training pipeline with the Taichi language doing the heavy compute and PyTorch handling the training loop, so the code is much easier to read and modify than the original CUDA-heavy reference. Ready-to-run scripts cover the synthetic NeRF and 360_v2 datasets, and there is a video-to-NeRF path that runs COLMAP on your own footage and trains a scene from it. The interactive GUI and the published training times (the Lego scene hits 35.0 PSNR in about 208 seconds on an RTX 3090) make it handy for quick experiments in 3D reconstruction and novel-view synthesis. It also works as a backend for text-to-3D work through stable-dreamfusion, so it is useful beyond pure reconstruction.