This project reimplemented NVIDIA’s Instant-NGP method for training NeRFs in pure PyTorch, replacing the original C++/CUDA implementation. It achieves significantly faster convergence by using multi-resolution hash encoding to represent neural graphics primitives, enabling training of neural radiance fields up to 100x faster than vanilla NeRF on a single GPU. The implementation builds on the NeRF-pytorch codebase and provides additional loss options including total variation loss and sparsity-inducing loss.