nvdiffmodeling

Differentiable rasterization applied to 3D model simplification tasks, as described in the paper:

Appearance-Driven Automatic 3D Model Simplification
Jon Hasselgren, Jacob Munkberg, Jaakko Lehtinen, Miika Aittala and Samuli Laine
https://research.nvidia.com/publication/2021-04_Appearance-Driven-Automatic-3D
https://arxiv.org/abs/2104.03989

License

This work is made available under the Nvidia Source Code License.

For business inquiries, please visit our website and submit the form: NVIDIA Research Licensing

Citation

@inproceedings{Hasselgren2021,
  title     = {Appearance-Driven Automatic 3D Model Simplification},
  author    = {Jon Hasselgren and Jacob Munkberg and Jaakko Lehtinen and Miika Aittala and Samuli Laine},
  booktitle = {Eurographics Symposium on Rendering},
  year      = {2021}
}

Installation

Requirements:

Microsoft Visual Studio 2019+ with Microsoft Visual C++ installed
Cuda 10.2+
Pytorch 1.6+

Tested in Anaconda3 with Python 3.6 and PyTorch 1.8.

One time setup (Windows)

Install Microsoft Visual Studio 2019+ with Microsoft Visual C++.
Install Cuda 10.2 or above. Note: Install CUDA toolkit from https://developer.nvidia.com/cuda-toolkit (not through anaconda)
Install the appropriate version of PyTorch compatible with the installed Cuda toolkit. Below is an example with Cuda 11.1

conda create -n dmodel python=3.6
activate dmodel
conda install pytorch torchvision torchaudio cudatoolkit=11.1 -c pytorch -c conda-forge
conda install imageio
pip install PyOpenGL glfw

Install nvdiffrast in the dmodel conda env. Follow the installation instructions.

Every new command prompt

activate dmodel

Examples

Sphere to cow example:

python train.py --config configs/spot.json

The results will be stored in the out folder. The Spot model was created and released into the public domain by Keenan Crane.

Additional assets can be downloaded here [205MB]. Unzip and place the subfolders in the project data folder, e.g., data\skull. All assets are copyright of their respective authors, see included license files for further details.

Included examples

skull.json - Joint normal map and shape optimization on a skull
ewer.json - Ewer model from a reduced mesh as initial guess
gardenina.json - Aggregate geometry example
hibiscus.json - Aggregate geometry example
dancer_brushed_gold_64.json - LOD example, trained against a supersampled reference
dancer_displacement.json - Joint shape, normal map, and displacement map example

The json files that end in _paper.json are configs with the settings used for the results in the paper. They take longer and require a GPU with sufficient memory.

Server usage (through Docker)

Build docker image (run the command from the code root folder). docker build -f docker/Dockerfile -t diffmod:v1 . Requires a driver that supports Cuda 10.1 or newer.
Start an interactive docker container: docker run --gpus device=0 -it --rm -v /raid:/raid -it diffmod:v1 bash
Detached docker: docker run --gpus device=1 -d -v /raid:/raid -w=[path to the code] diffmod:v1 python train.py --config configs/spot.json

pielet/nvdiffmodeling