Skip to content

PIL × DAT - Pillow extension for AI-based image upscaling.

License

Notifications You must be signed in to change notification settings

omgold/pillow-dat

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

26 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

PIL × DAT

PIL × DAT - Pillow extension for AI-based image upscaling.

License


Installation

For PyPI:

pip install pillow-dat

For Poetry:

poetry add pillow-dat

Get started

from PIL.Image import open

from PIL_DAT.Image import upscale

lumine_image = open(".github/lumine.png")
lumine_image = upscale(lumine_image, 2)

Remark: We strongly advocate for the utilization of DAT light models owing to their streamlined design and outstanding speed performance. However, should you opt for alternative models, please note that *.pth model weights can be accessed via Google Drive.

Example

Input (lumine.png) DAT light (x2) Bicubic (x2)
Input (lumine.png) DAT light (x2) Bicubic (x2)

Benchmarks

Speed

Performance benchmarks have been conducted on a computing system equipped with an Intel(R) CORE(TM) i7-9750H CPU @ 2.60GHz processor, accompanied by a 2 × 8 Go at 2667MHz RAM configuration. Below are the recorded results:

In seconds 320 × 320 640 × 640 960 × 960 1280 × 1280
DAT light (x2) 16.1 65.3 146.8 339.8
DAT light (x3) 14.3 61.7 - -
DAT light (x4) 14.0 63.0 - -

The results were compared against the renowned OpenCV library, utilizing its EDSR model known for delivering superior image quality.

In seconds 320 × 320 640 × 640 960 × 960 1280 × 1280
EDSR (x2) 25.6 112.9 264.1 472.8
EDSR (x3) 24.3 112.5 - -
EDSR (x4) 23.6 111.2 - -

Remark: All benchmark results presented here are reproducible. For detailed implementation, please refer to the following files: benchmark_speed_dat_light.py and benchmark_speed_edsr.py.

Quality

DAT light (x2) EDSR (x2)
DAT light (x2) EDSR (x2)

Remark: All benchmark results presented here are reproducible. For detailed implementation, please refer to the following files: example.py and benchmark_quality_edsr.py.

Alpha-channel-awareness

Input DAT light (x2) EDSR (x2)
Input DAT light (x2) EDSR (x2)

Remark: All benchmark results presented here are reproducible. For detailed implementation, please refer to the following files: benchmark_alpha_channel_awareness_dat_light.py and benchmark_alpha_channel_awareness_edsr.py.

Contribution

Please install Python.

Please install Poetry via pipx.

Please install VSCode and its extensions:

  • Black Formatter
  • isort
  • Python
  • Pylance
  • Even Better TOML

To have your Python environment inside your project (optional):

poetry config virtualenvs.in-project true

To create your Python environment and install dependencies:

poetry install

To run unit tests:

pytest

To publish package:

poetry publish --build -u __token__ -p <pypi_token>

Acknowledgement

This library is founded upon the pioneering research paper, "Dual Aggregation Transformer for Image Super-Resolution".

@inproceedings{chen2023dual,
    title={Dual Aggregation Transformer for Image Super-Resolution},
    author={Chen, Zheng and Zhang, Yulun and Gu, Jinjin and Kong, Linghe and Yang, Xiaokang and Yu, Fisher},
    booktitle={ICCV},
    year={2023}
}

About

PIL × DAT - Pillow extension for AI-based image upscaling.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%