Artifact is a deep learning project that uses convolutional neural networks in tandem with classical image processing techniques to create unique styles.
We begin by using classical image processing techniques to generate modified images from our original image set. The effects chosen are:
- Blur
- Kuwahara
- Posterize
Next, we train a CycleGAN model using our modified and original image sets. Hyperparameters to note, which may differ from the original CycleGAN paper, are as follows:
- The generator encoder and decoder use 32, 64, 128 filters instead of 64, 128, 256. This change was made to satisfy GPU space constraints.
- The discriminator loss is not scaled by 0.5. Better results were found when the discriminator loss was left unscaled.
- The number of residual layers is 9, as is recommended by the original paper for images of size 256 by 256.
- The number of training iterations is 160,000. The number of validation iterations is 40,000. Training is split into 8 epochs.
Finally, once training is complete, we feed a modified image into our model to obtain a new (fake original) image. Feel free to make observations of artifacts in your output image and tune the model as you like. Common artifacts are tiling, light and dark patches, and color distortion. Please note the following observations from training on the blur effect:
- Using random pairs of images rather than associated pairs yields better
results. See
--shufflebelow. - Validation loss is not a clear indicator of the degree to which artifacts are present.
- Increasing the number of residual layers to 12 seems to worsen artifacts.
- Scaling the discriminator loss by 0.5 seems to remove some of the color distortion at the cost of extra tiling and extra color distortion at the edge of light and dark patches.
- High-contrast, noisy images yield the best results. Likewise, images with flat coloring yield the worst results.
The following images are generated from a larger image and cropped after processing. Therefore, it may prove effective to apply reflection padding, symmetric padding, or wrap padding to an image prior to processing in order to reduce the effect of unwanted artifacts.
First, install the required dependencies:
pip install -r requirements.txt --upgrade
Once installation is complete, download and unzip the 2017 COCO image set (19GB):
python download.py
python unzip.py
Download ImageMagick. Please note that the commands below will take a very long time to complete for the full set of 123,403 images.
Before processing, it is highly recommended to first resize and crop the entire set of images:
magick mogrify -resize 256x256^ -gravity center -extent 256x256 "data/COCO/*.jpg"
Next, copy the images into new folders:
cp -r data/COCO data/COCO_blur_2
cp -r data/COCO data/COCO_kuwahara_2
cp -r data/COCO data/COCO_posterize_8
Lastly, apply effects to the copied images:
magick mogrify -blur 0x2 "data/COCO_blur_2/*.jpg"
magick mogrify -kuwahara 2 "data/COCO_kuwahara_2/*.jpg"
magick mogrify -posterize 8 "data/COCO_posterize_8/*.jpg"
To train your model, run one of the following commands:
python train.py --modified data/COCO_blur_2 --prefix blur_2 --shuffle
python train.py --modified data/COCO_kuwahara_2 --prefix kuwahara_2 --shuffle
python train.py --modified data/COCO_posterize_8 --prefix posterize_8 --shuffle
Please consult the help documentation for more information:
python train.py --help
Once training is complete, you can use a checkpoint and a processed image to generate a new (fake original) image:
python evaluate.py --checkpoint checkpoints/blur_2_epoch=007.ckpt --input /PATH/TO/INPUT/IMAGE --fake
python evaluate.py --checkpoint checkpoints/kuwahara_2_epoch=007.ckpt --input /PATH/TO/INPUT/IMAGE --fake
python evaluate.py --checkpoint checkpoints/posterize_8_epoch=007.ckpt --input /PATH/TO/INPUT/IMAGE --fake
Additionally, you can view the parameters of your model like so:
python evaluate.py --checkpoint /PATH/TO/MODEL/CHECKPOINT
Please consult the help documentation for more information:
python evaluate.py --help
You can view log data from your training session using TensorBoard:
tensorboard --logdir lightning_logs/version_0
Run the following script for information about PyTorch and your graphics card:
python info.py
MIT License
@inproceedings{CycleGAN2017,
title={Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks},
author={Zhu, Jun-Yan and Park, Taesung and Isola, Phillip and Efros, Alexei A.},
booktitle={Computer Vision (ICCV), 2017 IEEE International Conference on},
year={2017}
}