Adaptive Instance Normalization

Unofficial Pytorch Implementation "Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization"

Reference: Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization, ICCV 2017

Requirements

• torch (version: 1.2.0)
• torchvision (version: 0.4.0)
• Pillow (version: 6.1.0)
• matplotlib (version: 3.1.1)

Download

• The trained model can be downloaded through the Google drive
• Content dataset: MSCOCO train2014 is needed to train the network.
• Style dataset: Painter by numbers is needed to trian the network.

Usage

Arguments

• --train: Train flag to learn the style transfer network
• --content-dir: Content dataset path for learn the style transfer network
• --style-dir: Style dataset path for learn the style transfer network
• --imsize: Size to resize the shorter side of the image (maintaining the aspect ratio)
• --cropsize: Size to crop the image
• --cencrop: Flag for crop the center region of the image (default: randomly crop)
• --style-weight: Style loss weight (If you want to enhance the style, increase this value to train the network)
• --gpu-no: Device no (-1: cpu, 0~N: gpu)
• --content: Source contenet image
• --style: Target style image
• --mask: Mask to generate masked stylized image
• --style-strength: Trade off factor between content and style (1.0: style, 0.0: content)
• --interpolation-weights: Interpolation weights of multiple styles
• --load-path: Model load path
• --layers: Layer Indices to extract content and style features
• --preserve-color: Flag for color preserved stylization

Example Scripts

Training

python main.py --train --imsize 512 --cropsize 256 --content-dir ./coco2014/ --style-dir ./painter_by_numbers/

• Training loss

Generat the stylized image

python main.py --imsize 512 --content ./imgs/content/lena.jpg --style ./imgs/style/abstraction.jpg --load-path ./check_point.pth --style-strength 1.0

Generat the stylized image with multiple styles

python main.py --imsize 512 --content ./imgs/content/chicago.jpg --style ./imgs/style/abstraction.jpg ./imgs/style/starry_night.jpg --load-path ./check_point.pth --style-strength 1.0 --interpolation-weights 0.5 0.5

Generat the stylized image with multiple styles and masks

python main.py --imsize 512 --content ./imgs/content/blonde_girl.jpg --style ./imgs/style/mondrian.jpg ./imgs/style/starry_night.jpg --mask ./imgs/mask/blonde_girl_mask1.jpg ./imgs/mask/blonde_girl_mask2.jpg --load-path ./check_point.pth --style-strength 1.0 --interpolation-weights 1.0 1.0

Generat the color preserved stylized image

python main.py --imsize 512 --content ./imgs/content/blonde_girl.jpg --style ./imgs/style/mondrian.jpg --load-path ./check_point.pth --preserve-color

Results

Stylized image with single style

Stylized image with multiple styles

Stylized image with multiple styles and masks

Color preserved stylized image