| license: mit | |
| library_name: mlx | |
| pipeline_tag: image-to-image | |
| tags: [mlx, image-restoration, deblurring, denoising, nafnet] | |
| base_model: megvii-research/NAFNet | |
| # NAFNet width64 (MLX) — Image deblurring | |
| Apple MLX port of **[NAFNet](https://github.com/megvii-research/NAFNet)** (Simple Baselines for | |
| Image Restoration, ECCV 2022). Runs on Apple Silicon via [MLX](https://github.com/ml-explore/mlx). | |
| This checkpoint: **REDS** (Image deblurring). width64. | |
| ## Usage | |
| ```python | |
| from nafnet_mlx import NAFNetConfig | |
| from nafnet_mlx.pipeline import load_model, restore_to_file | |
| m = load_model("model.safetensors", NAFNetConfig.reds_width64()) | |
| restore_to_file(m, "input.png", "output.png") | |
| ``` | |
| ## Validation | |
| Faithful NHWC port (SimpleGate, Simplified Channel Attention, channel-axis LayerNorm2d, | |
| UNet + PixelShuffle). PT-vs-MLX full-model parity on a real image ~1e-6. | |
| Uses NAFNetLocal (TLC) local pooling. | |
| ## License & attribution | |
| MIT. Derived from megvii-research/NAFNet (MIT). See `NOTICE`. | |