Then execute the following command in the root of the repository. If you want to use pretrained models to perform style transfer, please download the pre-trained models in Google Drive and put the downloaded experiments directory under the root of this repository. Optionally, if you are a conda user, you can execute the following command in the directory of this repository to create a new environment with all dependencies installed. The code in this repository solves this problem. When we continuously perform style transfer with a style transfer algorithm, the produced result will gradually lose the detail of the content image. We also train a model with the FFHQ dataset as the content and Metfaces as the style to convert a portrait photo into an artwork. Style Transfer ExamplesĪrtistic Portrait Style Transfer Examples ArtFlow adopts a projection-transfer-reversion scheme instead of the encoder-transfer-decoder to avoid the content leak issue of existing style transfer methods and consequently achieves unbiased style transfer in continuous style transfer.

Jie An *, Siyu Huang *, Yibing Song, Dejing Dou, Wei Liu and Jiebo LuoĪrtFlow is a universal style transfer method that consists of reversible neural flows and an unbiased feature transfer module. Official PyTorch implementation of the paper:ĪrtFlow: Unbiased Image Style Transfer via Reversible Neural Flows

Download artflow