Awesome
<div align="center"> <h2>ECCV 2024</h2> <h1>Every Pixel Has its Moments: Ultra-High-Resolution Unpaired Image-to-Image Translation via Dense Normalization</h1>Ming-Yang Ho<sup>1</sup>, Che-Ming Wu<sup>2</sup>, Min-Sheng Wu<sup>3</sup>, Yufeng Jane Tseng<sup>1</sup>
<sup>1</sup>National Taiwan University, <sup>2</sup>Amazon Web Services, <sup>3</sup>aetherAI<br>
[Paper (arxiv)
] [Paper (official)
] [Project Page
]
<br><br><image src="./images/teaser.jpg"/>
<br><image src="./images/framework.jpg"/>
Get started with an example
We provide a simple example (one image from the Kyoto summer2autumn dataset) for you to translate an UHR image with our DN.
Download example data
$ ./download.sh
$ unzip simple_example.zip
Environment preparation
- Please check your GPU driver version and modify
Dockerifle
accordingly - Then, execute
$ docker-compose up --build -d
- Get into the docker container
$ docker exec -it dn-env bash
Inference
- In the docker container, please execute
$ python3 transfer.py -c data/japan/config.yaml
- Then, you can see a translated image at
experiments/japan_CUT/test/IMG_6610/combined_dn_10.png
- To see the image conveniently, you can leverage the provided
visualization.ipynb
. The setup of jupyter notebbok can be achived by- a. modify a port mapping setting in
docker-compose.yml
; e,g,- 19000:8888
- b. install
jupyter
in the container - c. run your jupyter notebook by
nohup jupyter notebook --ip=0.0.0.0 --port=8888 --allow-root &
- d. open the jupter notebook service on your port (
19000
here)
- a. modify a port mapping setting in
Datasets
real2paint
Dataset
For the real domain, please download the UHDM dataset from its official website. For the painting domain, we have curated a dataset of high-resolution Vincent van Gogh paintings, which can be downloaded at link1 or link2. Please note that we do not own these images; users should ensure their use does not trigger legal issues.
Kyoto-summer2autumn
Dataset
Please download it at link.
ANHIR
Dataset
Please download it at link. Please note that we do not own these images; users should ensure their use does not trigger legal issues.
ACROBAT
Dataset
Please download it at link. Please note that we do not own these images; users should ensure their use does not trigger legal issues.
Train your model
The training of I2I model is the same as KIN. DN is a plugin for any I2I model with InstanceNorm layers.
Citation
@InProceedings{10.1007/978-3-031-72995-9_18,
author="Ho, Ming-Yang and Wu, Che-Ming and Wu, Min-Sheng and Tseng, Yufeng Jane",
title="Every Pixel Has Its Moments: Ultra-High-Resolution Unpaired Image-to-Image Translation via Dense Normalization",
booktitle="Computer Vision -- ECCV 2024",
year="2025",
publisher="Springer Nature Switzerland",
address="Cham",
pages="312--328",
isbn="978-3-031-72995-9"
}