Add NAFNet MLX weights (faithful port of megvii-research/NAFNet)

NOTICE

@@ -0,0 +1,13 @@
+nafnet-mlx
+Apple MLX port of NAFNet (Simple Baselines for Image Restoration).
+
+This work is licensed under the MIT License.
+
+Derived from:
+- megvii-research/NAFNet (MIT) — official PyTorch implementation and pretrained weights.
+  Chen et al., "Simple Baselines for Image Restoration", ECCV 2022 (arXiv:2204.04676).
+  TLC (test-time local converter) from Chu et al., arXiv:2112.04491.
+- Built on the BasicSR framework conventions (Apache-2.0).
+
+Pretrained weights (REDS / SIDD / GoPro width64) are the official megvii-research releases,
+converted to MLX safetensors. Original weights are MIT-licensed.

README.md

@@ -0,0 +1,30 @@
+---
+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 denoising
+
+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: **SIDD** (Image denoising). width64.
+
+## Usage
+```python
+from nafnet_mlx import NAFNetConfig
+from nafnet_mlx.pipeline import load_model, restore_to_file
+m = load_model("model.safetensors", NAFNetConfig.sidd_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.
+
+
+## License & attribution
+MIT. Derived from megvii-research/NAFNet (MIT). See `NOTICE`.

config.json

@@ -0,0 +1,24 @@
+{
+  "img_channel": 3,
+  "width": 64,
+  "middle_blk_num": 12,
+  "enc_blk_nums": [
+    2,
+    2,
+    4,
+    8
+  ],
+  "dec_blk_nums": [
+    2,
+    2,
+    2,
+    2
+  ],
+  "local": false,
+  "train_size": [
+    1,
+    3,
+    256,
+    256
+  ]
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:62c6825ef8224e81a89e3c2ff35e9cde2367815ef12648dd4837006b5b86a44c
+size 463995842