goc-liaozihan/neumip

NeuMIP: Multi-Resolution Neural Materials

SIGGRAPH 2021

https://cseweb.ucsd.edu/~viscomp/projects/NeuMIP

Setup and prerequisites

The code was developed on Ubuntu 18.04, using Python 3.7 and PyTorch 1.7. Although, we don't use any unique features of those systems, running on different configuration might require some small adjustments.

First you need to setup your data path, where all the datasets, models, etc. reside. Go to path_config.py and at line 9 change: path_base = "$YOUR_PATH to your path.

License

The code is licensed under BSD license.

Dataset

Here's the link to the datasets and models: https://drive.google.com/drive/folders/1EbzMlpmE7f49jTs2IY7oMQKZAGAfl3M5

Training and Running

To train:

./rung.sh 0 ./neural_rendering.py --dataset DATASET_NAME \
     --max_iter 30000 --outm  MODEL_NAME \
     --experiment  StandardRawLongShadowMaskOnly \
     --batch 4  --loss comb2

To visualize it:

./run.sh neural_rendering.py --inm wool2.pth --vis 

The main code is in the neural_rendering.py file. Beware, the code supports different configurations/architectures. Please, use -experiment StandardRawLongShadowMaskOnly For loss, use l1 or comb2, which is better for specular.