About

This repository is a fork of Jan Spindler's Neural Radiance Caching (NRC) implementation with two new modifications and an experimental implementation of two transmittance estimation methods from the paper An unbiased ray-marching transmittance estimator

First modification of NRC is adaptive train batch filtering. It splits input image into batches with a multi level uniform grid. Train batches that contained no scattering events are useless for the training and are excluded from the training. Multi level structure allows to join multiple batches into a single bigger batch to reduce training time.

Second modification of NRC adds cloud density to the input of NRC neural network. This modification increases training speed of the neural network by the factor of 6.

Detailed description of modifications present in this fork are described in the text of my master's thesis stored in the doc folder of this repository.

Results

To the left base NRC from master branch, to the right my modified version from combined branch with both modifications and biased transmittance estimator enabled.

Requirements

• CUDA v12.1 and NVIDIA GPU with compute capability >= 8.6
• Visual Studio 22. I used Visual Studio 2022 and it is not guaranteed that everything will work on other versions.
• (optional, for plot recreation): Python and Jupyter Notebook

Installation

IMPORTANT: some links that vcpkg references are unavailable in Russian Federation. You probably know how to deal with this. Travel to another country, do the installation steps, then go home.

Program was tested only on Windows 11 and instruction is written for x64-windows vcpkg triplet. If you have a different OS, you need to replace x64-windows in installation steps paths with your respective platform triplet.

Installation steps:

1. Install requirements
2. Open Developer Command Prompt for VS 22.
3. git clone --recurse-submodules -j8 https://github.com/Fenrir7Asteron/NRC-HPM-Renderer.git
4. cd NRC-HPM-Renderer
5. Open NRC-HPM-Renderer folder in Visual Studio 22
6. Configure root CMakeLists file with right click on Solution Explorer.
   In the process of configuring vcpkg will download all dependencies from vcpkg.json.
7. Return to Developer Command Prompt and execute
   .\install-openvdb-<Target>.bat
8. Go back to VS, set root CMakeLists as startup item and build the project

Run

Project can be run with default arguments with Visual Studio.
If you want to run it with different arguments through command line you can look up
default arguments at the bottom of src/main.cu file and meaning of arguments in the include/engine/AppConfig.hpp.

Built project files are in the out/build/<build-target> folder.

Startup arguments may vary for different modifications in different branches.

Project can be run in benchmark mode to store performance and quality metrics in the out/build/<build-target>/output/ folder. In order to start project in the benchmark mode you need to set the respective startup argument to 1.

OutputAnalysis/MetricPlotting.ipynb notebook can be used to reproduce plots from my thesis using data received from benchmarking.

Branches

Different branches contain different modifications of the base NRC implementation:

• master contains unmodified Jan Spindler's base NRC implementation
• train_filter_* contain several train filtering options
• input_* contains various options for additional input to neural network
• new_transmittance_estimator contains implementation of unbiased and biased raymarching transmittance estimators from this paper
• combined contains combination of modifications from train_filter_adaptive, input_density and new_transmittance_estimator branches
• approximate_mie implements importance sampling by phase function from this paper but it is more noisy than Henyey-Greenstein so it is not used in the combined branch