TxST (Arbitrary text driven artistic style transfer)

If the above does not work, try the following one.

Text-driven image style transfer (TxST) that leverages advanced image-text encoders to control arbitrary style transfer.

By Zhi-Song Liu, Li-Wen Wang, Wan-Chi Siu and Vicky Kalogeiton

This repo only provides simple testing codes and pretrained models.

Please check our paper (30.2Mb) or compressed version (3.2Mb).

@article{liu2022name,
  title={Name Your Style: An Arbitrary Artist-aware Image Style Transfer},
  author={Liu, Zhi-Song and Wang, Li-Wen and Siu, Wan-Chi and Kalogeiton, Vicky},
  journal={arXiv preprint arXiv:2202.13562},
  year={2022}
}

Requirements

• Ubuntu 20.04 (18.04 or higher)
• NVIDIA GPU

Dependencies

• Python 3.8 (> 3.0)
• PyTorch 1.8.2 (>= 1.8)
• NVIDIA GPU + CUDA 10.2 (or >=11.0)

Or you may create a new virtual python environment using Conda, as follows

conda create --name TxST python=3.8 -y
conda activate TxST
conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch-lts -y

Installation

First, install additional dependencies by running

$ pip install -r requirements.txt

Second, install CLIP

Please use the following command for installation, as we have modified the model.py for intermediate features.

$ pip install ./lib/CLIP

Testing

Download Model Files

1. Pre-trained Models

You can simply run the following commands:

wget --load-cookies /tmp/cookies.txt "https://docs.google.com/uc?export=download&confirm=$(wget --quiet --save-cookies /tmp/cookies.txt --keep-session-cookies --no-check-certificate 'https://docs.google.com/uc?export=download&id=1lQm5MGpPV1154MbtvGQDZlCMx2D8beHr' -O- | sed -rn 's/.*confirm=([0-9A-Za-z_]+).*/\1\n/p')&id=1lQm5MGpPV1154MbtvGQDZlCMx2D8beHr" -O models/tmp.zip && rm -rf /tmp/cookies.txt
unzip ./models/tmp.zip -d models
rm ./models/tmp.zip

The files under "./models" are like:

├── models
│   ├── readme.txt
│   ├── texture.ckpt
│   ├── wikiart_all.ckpt
│   └── wikiart_subset.ckpt

Or you can manually download them from here.

2. Pre-trained VGG Model

You can simply run the following commands:

wget --no-check-certificate 'https://docs.google.com/uc?export=download&id=19ZbeHK2UxzzTNeDMcWfE1TbyFkBUurns' -O pretrained_models/vgg_normalised.pth

The files under "./pretrained_models" are like:

├── pretrained_models
│   ├── readme.txt
│   └── vgg_normalised.pth

Or you can manually download them from here.

Artist Style Transfer using Reference Images

put your content images under "data/content" and put your style images under "data/style"

then run the following script.

$ python eval_ST_img.py

the results are saved at "output" folder, like

Artist Style Transfer using Texts

run

You can find some artists' names from wikiauthors.txt file.

$ python demo_edit_art_style.py --content %path-to-your-content-image% --style %artistic-text%

# Example
python demo_edit_art_style.py --content data/content/14.jpg --style vangogh

the results are saved at "output" folder, like:

Texture Style Transfer using Texts

run

$ python demo_edit_texture_style.py --content %path-to-your-content-image% --style %texture-text%

# Example
python demo_edit_texture_style.py --content data/content/14.jpg --style grid

the results are saved at "output" folder, like:

Visualization

Here we show some cases on Wikiart style transfer using just texts as style description. We first compare with state-of-the-art CLIP based approach CLIPstyler. We have better artistic stylization and consistent style changes.

We also use more artists's names for style transfer.