| | --- |
| | license: bsd-3-clause |
| | tags: |
| | - coreml |
| | - super-resolution |
| | - real-esrgan |
| | - image-upscaling |
| | --- |
| | |
| | # Real-ESRGAN CoreML |
| |
|
| | CoreML conversion of [xinntao/Real-ESRGAN](https://github.com/xinntao/Real-ESRGAN) (RealESRGAN_x4plus) for use in macOS/iOS apps. |
| | |
| | ## Model |
| | |
| | - **Architecture**: RRDBNet (23 RRDB blocks, 64 features, 32 growth channels) |
| | - **Scale**: 4x upscaling |
| | - **Precision**: Float16 |
| | - **Input**: `(1, 3, 256, 256)` — RGB normalized [0, 1] |
| | - **Output**: `(1, 3, 1024, 1024)` — RGB [0, 1] |
| | - **Size**: ~33 MB (mlpackage), ~30 MB (zip) |
| | - **Target**: macOS 15+ |
| | |
| | ## Usage |
| | |
| | Download `RealESRGAN-x4plus.mlpackage.zip`, unzip, then compile: |
| | |
| | ```swift |
| | let compiledURL = try MLModel.compileModel(at: mlpackageURL) |
| | let model = try MLModel(contentsOf: compiledURL) |
| | ``` |
| | |
| | For images larger than 256x256, use tile-based processing with overlap padding. |
| | |
| | ## License |
| | |
| | BSD-3-Clause (same as original Real-ESRGAN) |
| | |
| | ## Credits |
| | |
| | - Original model: [xinntao/Real-ESRGAN](https://github.com/xinntao/Real-ESRGAN) |
| | - Converted for [Fluxforge Studio](https://fluxforge.app) |
| | |
