Perceptual Losses for Real-Time Style Transfer

Unofficial PyTorch implementation of real-time style transfer

Reference: Perceptual Losses for Real-Time Style Transfer and Super-Resolution, ECCV2016

Requirements

• Pytorch (version >= 0.4.0)
• Pillow

Download

• The trained models can be downloaded throuth the Google drive.
• MSCOCO train2014 is needed to train the network.

Usage

Arguments

• --train-flag: Flag for train or evaluate transform network
• --train-content: Path of content image dataset (MSCOCO is needed)
• --train-style: Path of a target style image
• --test-content: Path of a test content image
• --model-load-path: Path of trained transform network to stylize the --test-content image

Train example script

python main.py --train-flag True --cuda-device-no 0 --imsize 256 --cropsize 240 --train-content ./coco2014/ --train-style imgs/style/mondrian.jpg --save-path trained_models/

Test example script

python main.py --train-flag False --cuda-device-no 0 --imsize 256 --model-load-path trained_models/transform_network.pth --test-content imgs/content/chicago.jpg --output stylized.png

Results