NVidia AI Denoiser command line tool

This is a simple implementation of NVidia AI denoiser. You can find a pre-built windows distribution either on my website here or in the releases tab of this repro. To build you will need to install the CUDA took availible from here and the OptiX 7/8 SDK availible here.

You will require an Nvidia driver of at least 465.84 or higher and an Nvidia GPU of Maxwell architecture or newer to use the OptiX denoiser.

Building

python3 -m pip install conan
conan profile detect --force
conan install . --build=missing -s:h compiler.cppstd=20 -s:b compiler.cppstd=20
cmake -S . -B build -DCMAKE_POLICY_DEFAULT_CMP0091=NEW -DCMAKE_TOOLCHAIN_FILE=build/generators/conan_toolchain.cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=./out
cmake --build build --config Release
cmake --install build --config Release

Usage

Command line parameters

• Command line parameters
• -v [int] : log verbosity level 0:disabled 1:simple 2:full (default 2)
• -i [string] : path to input image
• -pi [string] : path previous denoised result (optional, required for temporal denoising)
• -aov%d [string] : path to additional input AOV image to denoise
• -oaov%d [string] : path to additional AOV output image to denoise
• -o [string] : path to output image
• -os [string] : output suffix appended to input filename to create output image filename
• -a [string] : path to input albedo AOV (optional)
• -n [string] : path to input normal AOV (optional, requires albedo AOV)
• -mv [string] : path to motion vector AOV (optional, required for temporal denoising)
• -pid [string] : path to prevoius denoiser-internal per-frame data (optional, required for temporal denoising)
• -oid [string] : path to current denoiser-internal per-frame data (optional, required for temporal denoising)
• -b [float] : blend amount (default 0)
• -hdr [int] : Use HDR training data (default 1)
• -gpu [int] : Select which GPU to use for denoising (default 0)
• -repeat [int] : Execute the denoiser N times. Useful for profiling.
• -h/--help : Lists command line parameters

You need to at least have an input and output for the app to run. If you also have them, you can add an albedo AOV or albedo and normal AOVs to improve the denoising. All images should be the same resolutions, not meeting this requirement will lead to unexpected results (likely a crash).

For best results provide as many of the AOVs as possible to the denoiser. Generally the more information the denoiser has to work with the better. The denoiser also prefers images rendered with a box filter or by using FIS.

Examples

Here is a quick example scene that uses the images that can be found in the image folder of this repository.

Noisy image

Denoised output

Simple sequence batch script

As it has been widely requested here is a very simple batch script for denoising sequences until I have time to implement something proper into the application itself. It will do the most simple denoising without any feature AOVs. Save the following code into a file named Sequence.bat and place it into the directory where your images are saved. Running this script will denoise all files image files that match the chosen file extension in the folder. There are three parameters that you will need to edit in the script,

• FILE_EXTENSION – the file extension of your image
• PATH_TO_DENOISER – the full directory of the Denoiser.exe
• OUTPUT_PREFIX – a prefix which is prepended to the name of the image to create the output name. I.e. with the prefix denoised_ the image test.jpg will become denoised_test.jpg

SET FILE_EXTENSION=jpg
SET PATH_TO_DENOISER=D:\Projects\NvidiaAIDenoiser\Denoiser_v2.0
SET OUTPUT_PREFIX=denoised_

for /r %%v in (*.%FILE_EXTENSION%) do %PATH_TO_DENOISER%\Denoiser.exe -i "%%~nv.%FILE_EXTENSION%" -o "%OUTPUT_PREFIX%%%~nv.%FILE_EXTENSION%"

cmd /k

License info

This project is shared under the MIT License.