Transformed to CoreML NAFNet model for deblurring. Normalization, preprocessing and postprocessing were integrated to network graph. FP8 quantized model is provided weights/nafnet_reds_64_fp8.mlmodel (66.04 MB). Original model is NAFNet-REDS-width64.
There is a known issue with artifacts on the edges of the image. It is caused by the model inference with FP16 and lower. If you're encountering this issue, try to use FP32/FP16 model. You can download those models from here. Also, model needs to be inferenced in FP32 too, so try to use CPU on the device, because Apple mobile NPU/GPU is not supported for FP32.
Model accepts given sizes:
[
(256, 256), (512, 512), (768, 768), (1024, 1024),
(1280, 1280), (1536, 1536), (1792, 1792), (2048, 2048), (2304, 2304),
(2560, 2560), (2816, 2816), (3072, 3072), (3328, 3328), (3584, 3584),
(3840, 3840), (4096, 4096), (4352, 4352), (4608, 4608), (4864, 4864)
]Execution script pads provided images to the nearest size listed above.
- Create Python virtual environment:
virtualenv -p python3.9 .venv && source .venv/bin/activate- Install all requirements:
pip install -r requirements.txt- Use the project 🎉
To run use command below:
python run_nafnet_deblur.py --data-root imagesModel inference will work only on MacOS. Also, model loading will take a LOT of time. Generally, it is not suitable for mobile devices...
Model input is RGB Image. Model output is float32 image of shape (H, W, 3). Each pixel value of output varies from 0.0f to 255.0f.
Input node name is image, output - result.
- coremltools - The NNs inference framework
- OpenCV - Images processing framework
This project is licensed under the MIT License – see the LICENSE file for details