Home

Awesome

Neural Prefiltering for Correlation-Aware Levels of Detail

Alt text

Source code of the paper "Neural Prefiltering for Correlation-Aware Levels of Detail " by Philippe Weier, Tobias Zirr, Anton Kaplanyan, Ling-Qi Yan and Philipp Slusallek at SIGGRAPH 2023.

The implementation is based on the Real-Time Path Tracing Framework from Intel.

You can find the original paper and supplemental viewer on the project webpage.

Building the project

We tested our implementation with the following configuration:

Required Cuda Compute Capability

Our implementation depends on the FullyFusedMLP implementation of TinyCudaNN which requires at least Compute Capability 7.5. Make sure you set the variable CMAKE_CUDA_ARCHITECTURES="your_gpu_compute_capabability" appropriately for your GPU. A list with the compute capability associated with every GPU can be found here

Linux

From the rptr project directory configure and build the project as follows:

mkdir build && cd build

cmake -GNinja -DCMAKE_CXX_COMPILER=g++ -DCMAKE_C_COMPILER=gcc -DENABLE_NEURAL_LOD=ON -DCMAKE_BUILD_TYPE=Release -DCMAKE_CUDA_ARCHITECTURES="compute_capabability"..

ninja

If you are running into a posix-related error while re-building the simplest solution is to clean the CMakeFiles as follows:

rm -rf CMakefiles/ && rm CMakeCache.txt

Once compiled you can run a scene with a given configuration file from the build folder as follows:

./rptr ../../neural_lod/scenes/arcade/arcade.vks --config ../../neural_lod/scenes/arcade/arcade_world_final.ini --img 1024 1024

Getting the scenes used in the paper

The repository only ships with the Arcade scene. The other scenes used in the paper can be downloaded separately from the University of Saarland Cloud Server along with the pre-trained network weights. Simply download and copy the scene folders next to the arcade scene folder in neural_lod/scenes/. The other pre-trained scenes can then be rendered similarly, for example, to run the pre-trained Bay Cedar scene from the Moana Scene data-set use:

./rptr ../../neural_lod/scenes/isBayCedarA1/isBayCedarA1.vks --config ../../neural_lod/scenes/isBayCedarA1/isBayCedarA1_world_final.ini --img 1024 1024

Script for rendering all levels of detail

The script located at neural_lod/results/scripts/render_learned.py will launch the application, render the references as well as the neural prefiltered level of details and finally output a comparison figure between the reference and our approach at the end of the process. Note: the application still responds to user inputs while rendering so make sure to not change the camera while rendering. The Python script requires the following packages to be installed on your system (via pip install for example):

For an example usage, the following command renders all neural and reference LoDs for the arcade scene and stores the resulting images at neural_lod/results/arcade/:

python3 render_learned.py ../../neural_lod/scenes/arcade/arcade.vks --config_path ../../neural_lod/scenes/arcade/arcade_world_final.ini --result_dir ../../neural_lod/results/arcade/ --ref_spp 512 --spp 4096 --skip_present

To render other scenes only the scene name needs to be adapted. The previous script renders the scene with a constant background of 1, optionally, render_learned_envmap.py can be used to render the scene with a given HDR environment map, for example:

python3 render_learned_envmap.py ../../neural_lod/scenes/isBayCedarA1/isBayCedarA1.vks --config_path ../../neural_lod/scenes/isBayCedarA1/isBayCedarA1_world_final.ini --result_dir ../../neural_lod/results/isBayCedarA1_envmap/ --ref_spp 1024 --spp 16384 --envmap_file syferfontein_6d_clear_puresky_2k.exr --envmap_rotation 60 --skip_present

Learning visibility and appearance

If a scene is opened without any pre-trained weights you will need to optimise visibility and prefiltered appearance in the Throughput Optim and Visibility Optim tabs of the Neural Lod Learning window. After configuring the network input encoding and MLP as desired, start the optimisation by pressing Restart Throughput/Visibility Optimization. To visualise the result make sure to select the appropriate renderer variant in the Renderer window.

Create and render a custom scene

To render and learn the prefiltered version of a scene of your choice a few steps are necessary.

VKS/VKT Scene format export

To convert a scene with associated textures we provide a few helper scripts:

The libvkr project provides, among other helpful scripts, a way to check the metadata of a .vks scene file which also shows the expected path of the textures.

Scene Voxelisation

Our prefiltering technique requires a voxelised scene as input. This is provided in the form of a point cloud obj file where every vertex corresponds to the centre of a voxel. We generate this obj point cloud file using a slightly modified version of the cuda-voxeliser library located at neural_lod/cuda_voxelizer. The library also depends on the trimesh library. To build those two libraries, make sure to check out the original README of the respective projects.

Once the cuda_voxelizer is successfully compiled you can use the neural_lod/results/scripts/voxelize.py script to convert a .vks scene to a point cloud file. Assuming a test_scene test simply use:

python3 voxelize.py ../../scenes/test/test.vks ../../scenes/test/test

You should now be able to open your custom test scene using ./rptr ../../neural_lod/scenes/test/test.vks and learn the prefiltered visibility and appearance as described above.

Citation

@article{10.1145/3592443,
author = {Weier, Philippe and Zirr, Tobias and Kaplanyan, Anton and Yan, Ling-Qi and Slusallek, Philipp},
title = {Neural Prefiltering for Correlation-Aware Levels of Detail},
journal = {ACM Trans. Graph.},
issue_date = {August 2023},
volume = {42},
number = {4},
month = {jul},
articleno = {78},
numpages = {16},
year = {2023},
url = {https://doi.org/10.1145/3592443},
doi = {10.1145/3592443},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
}