Okeinla 1.0

Okeinla jointly denoises and demosaics Bayer RAW images. Trained this release using RawNIND and the Okeinla dataset. The selected checkpoint is 2940000.

Training code: https://github.com/AyedaOk/okeinla

Files

File	Use
pytorch/iter_2940000.pt	PyTorch model state dictionary
pytorch/iter_2940000.pt.opt	Matching Adam state dictionary for training resume
onnx/model_bayer.onnx	Production float32 ONNX model, opset 20
onnx/model.json	RAW application adapter metadata
darktable/okeinla.dtmodel	darktable model package
training/args.yaml	Minimal architecture settings for export
training/configs/	Three training-stage configurations
SHA256SUMS	File checksums

License and credits

The PyTorch weights, matching optimizer state, and ONNX model under GPLv3; see LICENSE. Okeinla's source code also uses GPLv3. The RawNIND and Okeinla datasets use CC BY-SA 4.0. Neither dataset is included in this model repository.

• RawNIND: Benoit Brummer and Christophe De Vleeschouwer, Learning Joint Denoising, Demosaicing, and Compression from the Raw Natural Image Noise Dataset (2025), paper, code, dataset.
• NAFNet: Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun, Simple Baselines for Image Restoration (ECCV 2022), code. See licenses/NAFNet-MIT.txt for the architecture's MIT notice.
• RawForge / Restorer: training code supplied the loss recipe. Okeinla's final stages increased the perceptual-loss weight from 0.025 to 0.035. PyTorch/torchvision supply VGG16 and its ImageNet weights separately.