A Style-Aware Content Loss for Real-time HD Style Transfer

Artsiom Sanakoyeu*, Dmytro Kotovenko*, Sabine Lang, Björn Ommer, In ECCV 2018 (Oral)

Website: https://compvis.github.io/adaptive-style-transfer
Paper: https://arxiv.org/abs/1807.10201

Please click on the image for a high-res version.

Requirements

python 2.7
tensorflow 1.2.
PIL, numpy, scipy
tqdm

Inference

Simplest van Gogh example

To launch the inference on van Gogh style:

Download the pretrained model model_van-gogh_ckpt.tar.gz
Download sample photographs.
Extract the model to ./models/ folder and sample photographs to ./data/ folder.
Run the following command:

CUDA_VISIBLE_DEVICES=0 python main.py \
                 --model_name=model_van-gogh \
                 --phase=inference \
                 --image_size=1280

Stylized photographs are stored in the folder ./models/model_van-gogh/inference_ckpt300000_sz1280/

Additional settings

--ii_dir INPUT_DIR - path to the folder containing target content images.
You can specify multiple folders separated with commas (don't use spaces!).
--image_size SIZE - resolution of the images to generate.
--save_dir SAVE_DIR - path to the output dir where the generated images will be saved.
--model_name NAME - the name of the model (all model should as subfolders in ./models/).

Usage example (inference):

CUDA_VISIBLE_DEVICES=0 python main.py \
                 --model_name=model_van-gogh \
                 --phase=inference \
                 --image_size=1280 \
                 --ii_dir ../my_photographs1/,../my_photographs2/ \
                 --save_dir=../save_processed_images_here/

If your GPU memory is not large enough, set the variable CUDA_VISIBLE_DEVICES="" to use CPU.

Pretrained models

We provide pretrained models for the following artists:
Paul Cezanne,
El-Greco,
Paul Gauguin,
Wassily Kandinsky (Василий Кандинский),
Ernst Ludwig Kirchner,
Claude Monet,
Berthe Morisot,
Edvard Munch,
Samuel Peploe,
Pablo Picasso,
Jackson Pollock,
Nicholal Roerich (Николай Рерих),
Vincent van Gogh.

Download pretrained models: link.
Extract models to the folder ./models/.

Training

Content images used for training: Places365-Standard high-res train mages (105GB).

Style images used for training the aforementioned models: download link.
Query style examples used to collect style images: query_style_images.tar.gz.

For example, Vincent van Gogh style: vincent-van-gogh_road-with-cypresses-1890.tar.gz.
This is the dataset representing a particular artistic period of Vincent van Gogh and was automatically collected using "Road with Cypress and Star, 1890" painting as query.

Download and extract style archives in folder ./data.
Download and extract content images.
Launch the training process (for example, on van Gogh):

CUDA_VISIBLE_DEVICES=1 python main.py \
                 --model_name=model_van-gogh_new \
                 --batch_size=1 \
                 --phase=train \
                 --image_size=768 \
                 --lr=0.0002 \
                 --dsr=0.8 \
                 --ptcd=/path/to/Places2/data_large \
                 --ptad=./data/vincent-van-gogh_road-with-cypresses-1890

Evaluation

How to calculate Deception Score and where to download artist classification model is described in evaluation.

Video stylization

To stylize a video you can use the following script:

# split video on a set of frames
ffmpeg -i myvideo.mp4 -r 25 -f image2 image-%04d.png

CUDA_VISIBLE_DEVICES=0 python main.py \
--model_name=model_van-gogh \
--phase=inference \
--image_size=1280 \
--ii_dir=input \
--save_dir=output

# reassemble the video back from frames:
ffmpeg -i image-%04d_stylized.jpg kktie-out.mp4

Reference

If you use this code or data, please cite the paper:

@conference{sanakoyeu2018styleaware,
  title={A Style-Aware Content Loss for Real-time HD Style Transfer},
  author={Sanakoyeu, Artsiom, and Kotovenko, Dmytro, and Lang, Sabine, and Ommer, Bj\"orn},
  booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
  year={2018}
}

cedspam / adaptive-style-transfer