This is official implementation of the paper "Reference based sketch extraction via attention mechanism"
Amirsaman Ashtari*, Chang Wook Seo*, Cholmin Kang, Sihun Cha, Junyong Noh (*equally contributed and listed alphabetically)
Journal: ACM TOG
Conference: SIGGRAPH ASIA 2022
$ python main.py --mode train_deep \
--scope ref2sketch \
--name_data examples \
--dir_data ./datasets \
--dir_log ./log \
--dir_checkpoint ./checkpoint \
--direction a2b
--gpu_ids 0
- For settings from original paper, use 'train' for --mode. However we recommend to use 'train_deep' for producing better quality of model with deeper convolution layers.
- To understand hierarchy of dataset, see Dataset directories structure below.
$ python test.py --name_weight [weight_file] \
--name_data examples \
--direction a2b \
--cuda
- Download the pretrained model from google drive https://drive.google.com/drive/folders/1OvU4pPVTFk07Kp5qKS6MNlsKhoUUckcR?usp=sharing
- Pretrained model "ref2sketch" produce results same as experiments on paper, however we recommend to use "ref2sketch_deep" for better quality output.
- Due to the regulation of twitter, we share each url of twitter dataset used in our paper. You can find twitter_dataset.json file from dataset folder.
Twitter regulation
https://developer.twitter.com/en/developer-terms/agreement-and-policy
ref2sketch
+---[dir_data]
| \---[name_data]
| +---test
| | +---a
| | | +---test_input1.png
| | | +---test_input2.png
| | +---b
| | | +---test_groundtruth1.png #not necessary for testing
| | | +---test_groundtruth2.png #not necessary for testing
| | +---c
| | | +---style1.png
| | | +---style2.png
| +---train
| | +---a
| | | +---train_input1.png
| | | +---train_input2.png
| | +---b
| | | +---train_groundtruth1.png
| | | +---train_groundtruth2.png
#Paired input and groundtruth should have same filename
