Pixelization of Natural Image via Cartoonization

発表スライド

References

• CartoonGAN: Generative Adversarial Networks for Photo Cartoonization
• Deep Unsupervised Pixelization

Requirement

• Python 3.5
• PIL
• Numpy
• Pytorch 0.4.0
• Torch

Dataset

Training Dataset (Deep Unsupervised Pixelization)

We collect 3000 portrait images and 900 pixel arts for traning our method. The folders named trainA and trainB contain the portrait images and pixel arts respectively. (aligned celeba, clip arts and pixel arts).

DUP's pretrained model is in ./checkpoints_pixelization.
CartoonGAN's pretained models are in ./cartoonGan/pretrained_model.

Testing Dataset

DUP

Create the folders testA and testB in a certain directory. Note that testA and testB contain the cartoon arts to be pixelized and pixel arts to be depixelized respectively.

CartoonGAN

Input natural images - portrait aligned celeba

Training

• To train a model (DUP):

python3 ./train.py --dataroot ./samples --resize_or_crop resize_and_crop --gpu_ids 0

Testing

• After training, all models have been saved in the directory ./checkpoints_pixelization/.
• To test a model (DUP):

python3 ./test.py --dataroot ./samples --resize_or_crop resize_and_crop --gpu_ids 0 --how_many 1 --which_epoch 150 --loadSize 256

• To test a model (Pixelization of Natural Image via Cartoonization)

python ./main.py --loadSize 320

• Note :
  main.py will cartoonized portrait images in ./input/natural_input and output to ./input/cartoon_input/testA
  Then testA and testB in ./input/cartoon_input will be pixelized and depixelized respectivelty.
  More testing flags in the file ./options/base_options.py.
  All testing results will be shown in the directory ./results_pixelization/.