---
license: mit
library_name: libreyolo
tags:
  - denoising
  - image-to-image
  - nafnet
pipeline_tag: image-to-image
---

# LibreNAFNetl-restore-sidd

NAFNet SIDD real-image denoising weights (width-64), repackaged for LibreYOLO.
This gives the LibreYOLO `denoise` restore alias a real model.

## Source

Weights derived from [megvii-research/NAFNet](https://github.com/megvii-research/NAFNet)
(`NAFNet-SIDD-width64`). Copyright (c) 2022 megvii-model. MIT (code) +
Apache-2.0 (BasicSR utilities). Trained on the Smartphone Image Denoising
Dataset (SIDD), which is MIT-licensed.

## Modifications

State-dict metadata-wrap only: keys and learned parameters are unchanged; the
checkpoint is wrapped in the LibreYOLO v1.0 schema (`task=restore`,
`degradation=denoise`, `dataset=SIDD`). Conversion is bit-exact vs the upstream
model (`max_abs_diff == 0`, fp32), so the upstream reported SIDD sRGB validation
figure (PSNR 40.3045 dB, SSIM 0.9614) applies. See
`weights/convert_nafnet_weights.py` in the
[LibreYOLO source repository](https://github.com/LibreYOLO/libreyolo).

## Usage

```python
from libreyolo import LibreYOLO

model = LibreYOLO("LibreNAFNetl-restore-sidd.pt")
res = model.predict("noisy.jpg")
res[0].save("denoised.png")
```

## License

MIT (with Apache-2.0 BasicSR utilities). See the [`LICENSE`](./LICENSE) and
[`NOTICE`](./NOTICE) files in this repository.