This is code of "A Semi-Procedural Convolutional Material Prior" Project | Paper
set up environment using environment.yml or build up your own environment
To run using our provided data and pattern, run this script:
python batch_optim2.py --load_option rand --name_pf $filename --ckpt_dir $foldername --loss TD+32L1 --total_iter 2000 --scale_opt --H_inten 10 --lr 0.005 --run_option opt
When running this, images in data folder will be loaded per class, $filename and $foldername are the saved file and folder path, results will be save to the directory like below:
$foldername/stone/stone_wall_2/TD+1e-132L1_3l16c5k_5in5o_$filename_rand_0
for TD+1e-132L1_3l16c5k_5in5o_tt_rand_0: "TD+1e-132L1" means loss; "3l16c5k" mean architecture (3 layer pixconv, 16 channel, 5 kernel size chaconv)
To run on your customized data, use this script:
python cust_optim.py --name_pf $filename --ckpt_dir $foldername --loss TD+32L1 --total_iter 2000 --scale_opt --H_inten 10 --lr 0.005 --run_option opt --in_img_path $imgpath --in_pat_path $patpath
where $imgpath and $patpath are the specified path of image and patterns, all the patterns in this directory will be used as input to the network
If you find this work useful for your research, please cite:
@inproceedings{zhou2023semi,
title={A Semi-Procedural Convolutional Material Prior},
author={Zhou, Xilong and Hasan, Milos and Deschaintre, Valentin and Guerrero, Paul and Sunkavalli, Kalyan and Kalantari, Nima Khademi},
booktitle={Computer Graphics Forum},
year={2023},
organization={Wiley Online Library}
}
Please contact Xilong Zhou ([email protected]) if there are any issues/comments/questions.