shenfalong / styletransfer

Real-time neural style transfer via meta networks

Neural Style Transfer

This repository is for [Neural Style Transfer Via Meta Networks]. The meta network takes in the style image and generated an image transformation network for neural style transfer. The fast generated model is only 449KB, which is able to real-time execute on a mobile device. For more details please refer and cite this paper

@inproceedings{shen2018style,
  author = {Falong Shen, Shuicheng Yan and Gang Zeng},
  title = {Neural Style Transfer Via Meta Networks},
  booktitle = {CVPR2018},
  year = {2018}
}

Installation

This library is based on Caffe. CuDNN 7 and NCCL 1 are required. Please follow the installation instruction of Caffe.

Meta Network Architecture

Examples

The size of image transformation network for the following images is 7MB.

The size of image transformation network for the following images is 449KB.

Scripts

Python code. Please execute the scripts in Python folder. Meta model is very huge while the generated model is very small.

• pretrained meta models
  Meta model train_8 (130M), generated model is 449KB.
  Meta model train_32 (968M), generated model is 7MB.
  

Put these models into python/model/ and modify the model name in demo.py.

Pytorch implementation

https://ypw.io/style-transfer/