Add rawdenoise-nind model: UtNet2 raw denoiser (Bayer + linear Rec.2020)

[andriiryzhkov]

Summary

Adds a new model package – rawdenoise-nind – providing the raw-denoise task. Two UtNet2 variants ship in a single type: multi package, auto-dispatched by darktable based on the sensor CFA pattern:

• Bayer – joint denoise + demosaic for Bayer sensors. 4ch packed RGGB input → 3ch camRGB at 2× spatial resolution (the PixelShuffle output head upsamples by 2). Output is in the camera's native RGB space at arbitrary learned gain; the consumer scalar-gain-matches and applies the ColorMatrix.
• Linear – pure denoise for non-Bayer sensors (X-Trans, Foveon) and anything already demosaicked. 3ch linear Rec.2020 → 3ch linear Rec.2020, same spatial, arbitrary gain.

Both are the canonical base variants from the graph_denoise_models_definitions.yaml map in the upstream training code, with funit=32, activation=LeakyReLU, and match_gain: output (training deliberately leaves the output at an arbitrary scale so the consumer gain-matches at inference).

Source & provenance

• Training code: https://github.com/trougnouf/rawnind_jddc (GPL-3.0)
• Paper: Learning Joint Denoising, Demosaicing, and Compression from the Raw Natural Image Noise Dataset – Brummer & De Vleeschouwer, 2025
• Training dataset: RawNIND – real-world raw noise/clean pairs captured by the authors, published on Wikimedia Commons and UCLouvain Dataverse (CC BY 4.0 / CC0 per-image)
• Selected checkpoints:
  • Bayer: DenoiserTrainingBayerToProfiledRGB_4ch_2024-02-21-bayer_ms-ssim_mgout_notrans_valeither_-4 (iter 4350000)
  • Linear: DenoiserTrainingProfiledRGBToProfiledRGB_3ch_2024-10-09-prgb_ms-ssim_mgout_notrans_valeither_-1 (iter 3900000)

OSAID / MOF status

Fully open: model weights GPL-3.0, training code GPL-3.0, training data CC BY 4.0 / CC0, training-run logs published. Qualifies as OSAID v1.0 Open Source AI and MOF Class I (Open Science).

[TurboGit]

Thanks!

[TurboGit]

@andriiryzhkov : One of the CI check is failing, need some more love :) TIA.

[andriiryzhkov]

@TurboGit : Fixed - demo part was too heavy for CI.

[TurboGit]

Thanks!