Okeinla

Okeinla is a Bayer RAW denoising and demosaicing model built on RawNIND and NAFNet.

• Model weights and application packages
• Okeinla dataset

Setup

Use Python 3.13 or newer and uv:

uv sync
cd src/rawnind
export OPENCV_IO_ENABLE_OPENEXR=1

Run training and tools from src/rawnind.

Credits

• RawNIND: Benoit Brummer and Christophe De Vleeschouwer, Learning Joint Denoising, Demosaicing, and Compression from the Raw Natural Image Noise Dataset (2025), paper.
• RawForge / Restorer: Okeinla uses a loss recipe derived from its training code.
• NAFNet: Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun, Simple Baselines for Image Restoration (ECCV 2022), code.

License

Code and model weights: GPLv3. Okeinla and RawNIND datasets: CC BY-SA 4.0. See LICENSE and the NAFNet MIT notice.