GANime

This program implements several deep generative models for generating colorized anime characters based on sketch drawings. There are three main models used in this project, including Neural Style Transfer, Conditional GAN (Pix2Pix), and CycleGAN. [Paper] [Poster]

Demo

This section displays some outputs of the program. Each example contains an input image (left), a ground truth image (middle), and a generated image (right).

Installation

The project requires the following frameworks:

• TensorFlow 2.1.0: https://www.tensorflow.org

• TensorBoard: https://www.tensorflow.org/tensorboard

• TensorFlow Hub: https://www.tensorflow.org/hub

• NumPy: https://numpy.org

• Matplotlib: https://matplotlib.org

• Kaggle API: https://github.com/Kaggle/kaggle-api

Download and preprocess data

This project uses the Anime Sketch Colorization Pair dataset from Kaggle. To get the data, go to https://www.kaggle.com/ktaebum/anime-sketch-colorization-pair, download the dataset, and unzip it in the data folder. Another option is to run the following command (with Kaggle API installed):

python utils/download_data.py

After that, run the following command to preprocess the data:

python utils/preprocess_data.py --model <model>

Train the models

• Neural Style Transfer

python train.py --model neural_style_transfer --epochs 1000 --content-path </path/to/content/image/>  --style-path </path/to/style/image/> --output-path </path/to/output/image/>

• Fast Neural Style Transfer

python train.py --model fast_neural_style_transfer --content-path </path/to/content/image/>  --style-path </path/to/style/image/> --output-path </path/to/output/image/>

• Pix2Pix

python train.py --model pix2pix --epochs 150 --lr 2e-4 --batch-size 32 --data-path </path/to/dataset/> --resume --output-path </path/to/outputs/> --checkpoint-path </path/to/checkpoints/>

• CycleGAN

python train.py --model cyclegan --epochs 150 --lr 2e-4 --batch-size 8 --data-path </path/to/dataset/> --resume --output-path </path/to/outputs/> --checkpoint-path </path/to/checkpoints/>

Test the models

• Neural Style Transfer

python test.py --model neural_style_transfer --data-path </path/to/dataset/> --output-path </path/to/outputs/>

• Fast Neural Style Transfer

python test.py --model fast_neural_style_transfer --data-path </path/to/dataset/> --output-path </path/to/outputs/>

• Pix2Pix

python test.py --model pix2pix --data-path </path/to/dataset/> --output-path </path/to/outputs/> --checkpoint-path </path/to/checkpoints/>

• CycleGAN

python test.py --model cyclegan --data-path </path/to/dataset/> --output-path </path/to/outputs/> --checkpoint-path </path/to/checkpoints/>

Evaluate the models

The program implements two evaluation metrics, including the Structural Similarity (SSIM) Index and the Frechet Inception Distance (FID). To evaluate the models, run the following command:

python evaluate.py --model <model> --metric <metric> --output-path </path/to/outputs/>

Authors

• Tai Vu - Stanford University

• Robert Yang - Stanford University

References

• Gatys, Leon A., et al. “A Neural Algorithm of Artistic Style.” ArXiv:1508.06576 [Cs, q-Bio], Sept. 2015. arXiv.org, http://arxiv.org/abs/1508.06576.

• Ghiasi, Golnaz, et al. “Exploring the Structure of a Real-Time, Arbitrary Neural Artistic Stylization Network.” ArXiv:1705.06830 [Cs], Aug. 2017. arXiv.org, http://arxiv.org/abs/1705.06830.

• Isola, Phillip, et al. “Image-to-Image Translation with Conditional Adversarial Networks.” ArXiv:1611.07004 [Cs], Nov. 2018. arXiv.org, http://arxiv.org/abs/1611.07004.

• Zhu, Jun-Yan, et al. “Unpaired Image-to-Image Translation Using Cycle-Consistent Adversarial Networks.” ArXiv:1703.10593 [Cs], Nov. 2018. arXiv.org, http://arxiv.org/abs/1703.10593.

• “Tutorials | TensorFlow Core.” TensorFlow, https://www.tensorflow.org/tutorials.

• Tensorflow/Examples. 2018. tensorflow, 2020. GitHub, https://github.com/tensorflow/examples.

• “Machine Learning Mastery.” Machine Learning Mastery, https://machinelearningmastery.com/.