justanhduc / multar

The official implementation of the ICASSP paper "A simple way of multimodal and arbitrary style transfer"

Multimodal Arbitrary Style Transfer

The official implementation of the ICASSP paper "A simple way of multimodal and arbitrary style transfer".

Requirements

Theano

neuralnet

Usages

Training

Dataset

Download the MS COCO 2017 train and valid sets here. Put all the downloaded folders into a folder.

Next download the Wikiart train/test dataset. Also put the train/test split csv files into the downloaded folder

Training

First, download the trained normalized VGG weight here.

We bootstrap the network from a trained AdaIN. To train an AdaIN model in Theano, please see here. We provide a trained AdaIN weight here. Put all the weight files into the project root.

To train a network using the default settings, use

python train.py path/to/COCO-train-valid-root path/to/WikiArt

Testing

To test the network, prepare a folder of input images, a folder of style images, and a pretrained model. A trained model can be downloaded here. Then execute

python test.py path/to/input/images path/to/style/images/ path/to/a/trained/weight/file

By default, the script will generate 5 images/style. User --help to see more options.

Results

Citation

If you use this implementation in your paper, please kindly cite the following paper

@INPROCEEDINGS{multar,
author={A. {Nguyen} and S. {Choi} and W. {Kim} and S. {Lee}},
booktitle={ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
title={A Simple Way of Multimodal and Arbitrary Style Transfer},
year={2019},
volume={},
number={},
pages={1752-1756},
keywords={Image style transfer;convolutional neural network;deep learning},
doi={10.1109/ICASSP.2019.8683493},
ISSN={2379-190X},
month={May},}

References

• "Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization" by Huang et al.

• For more information, checkout my implementation of AdaIN.