NIND UNet Denoiser

Image denoiser trained on the Natural Image Noise Dataset (NIND) from Wikimedia Commons.

Source

• Repository: https://github.com/trougnouf/nind-denoise
• Paper: Natural Image Noise Dataset (CVPR Workshops 2019)
• Dataset: https://commons.wikimedia.org/wiki/Natural_Image_Noise_Dataset
• License: GPL-3.0

Architecture

Standard U-Net encoder-decoder: 3 → 64 → 128 → 256 → 512 → 512 (bottleneck), with skip connections and sigmoid output activation.

ONNX Model

Property	Value
File	model.onnx
Input	input — float32 [1, 3, 768, 768]
Output	output — float32 [1, 3, 768, 768]
Resolution	Static, baked at 768×768
Opset	20
Normalize	[0, 1] range (divide by 255)
Tiling	Yes (attributes.input_sizes: [768])

Notes

• No ImageNet normalization — input is simply RGB in [0, 1] range.
• Output is clamped to [0, 1] by the sigmoid activation.
• Exported with FP32 precision.
• The 768×768 input is baked into the graph so JIT-compiling EPs (CoreML, MIGraphX) only pay the compile cost once. Callers must tile at exactly this size; darktable reads input_sizes from the manifest and locks the runtime tile size accordingly.

Selection Criteria

Property	Value
Model license	GPL-3.0
OSAID v1.0	Open Source AI
MOF	Class I (Open Science)
Training data license	CC BY 4.0 / CC0 (per-image, Wikimedia Commons)
Training data provenance	Real-world noise/clean pairs photographed by authors, published on Wikimedia Commons
Training code	GPL-3.0
Known limitations	None — all components publicly available under open licenses
Published research	Natural Image Noise Dataset (CVPR Workshops 2019)
Inference	Local only, no cloud dependencies
Scope	Image denoising
Reproducibility	Full pipeline (setup, convert, clean, demo)