SCUNet — Core ML
Core ML conversion of SCUNet scunet_color_real_psnr, a blind real-image denoising model by Kai Zhang et al. This repository distributes the converted FP32 model for use on macOS.
Original model and code: cszn/SCUNet. Original weights: KAIR v1.0 release, scunet_color_real_psnr.pth.
Download
The archive is hosted unchanged at the root of the main branch. It contains SCUNet.mlpackage/ at the top level. Downloads are public and do not require authentication.
- Filename:
SCUNet.mlpackage.zip - Size: 66,794,969 bytes
- SHA-256:
f60c1b9e2e1e52a08876f7e6acb4848c52373d564531a30c12573b4e73ff24bf
Model interface
| Property | Value |
|---|---|
| Format | Core ML ML Program (.mlpackage) |
| Compute precision | FP32 |
| Minimum deployment target | macOS 15 |
| Input | image: Float32 MLMultiArray, NCHW [1, 3, 512, 512] |
| Output | denoised: Float32 MLMultiArray, NCHW [1, 3, 512, 512] |
| Color encoding | sRGB-encoded RGB, range 0–1 |
| Output range | Clamped to 0–1 in the model |
The tile size is fixed at 512×512. Tile larger images with overlap; the consuming app uses 32 pixels of overlap. Extract the archive and compile/load the package with Core ML. This artifact is a Core ML conversion, not an MLX checkpoint.
Conversion notes
Converted from the original SCUNet weights using a trace-friendly implementation. The export replaces stride-2 2×2 downsampling convolutions with equivalent space-to-depth and 1×1 convolutions to work around a Core ML CPU backend discrepancy. FP32 was selected for this distribution.
Credits and license
SCUNet was developed by Kai Zhang and collaborators. Please credit the original authors and their repository when using this model.
Zhang et al., Practical Blind Image Denoising via Swin-Conv-UNet and Data Synthesis, Machine Intelligence Research, 20, 822–836 (2023). Paper.
The original SCUNet code and weights are licensed under Apache-2.0. This converted model is distributed under the same license. See the upstream license.