--- license: mit language: - en pipeline_tag: image-to-image tags: - coreml - mlpackage - coremltools - ios - macos - apple-silicon - super-resolution - realesrgan - img2img - image-enhancement - upscale - denoise - xcode - neural-engine --- A curated collection of CoreML (`.mlpackage`) models These Models Can be used easily inside of Xcode projects targeting across apple products If you have link/download issues, my GitHub collection is here: https://github.com/marshiyar/CoreML-Model-Zoo (downloads in Releases) ### RealESRGAN x4
Model contract - **Predictions:** 1x3x256x256 -> 1x3x1024x1024 - **Filename mapping:** <model_id>.mlpackage
File sizes are based on the zipped .mlpackage archives. | Model ID | Size | Input | Output | Profile | | ----------------------------------- | ------- | ----- | ------ | ------- | | [RealESRGAN_x4 ⬇️](https://huggingface.co/marshiyar/RealESRGAN_x4_CoreML/resolve/main/RealESRGAN_x4.mlpackage.zip) | 31 MB | | | base | | [RealESRGAN_x4_Q-8 ⬇️](https://huggingface.co/marshiyar/RealESRGAN_x4_CoreML/resolve/main/RealESRGAN_x4_Q-8.mlpackage.zip) | 13.6 MB | | | Q-8 | | [RealESRGAN_x4_pal-4 ⬇️](https://huggingface.co/marshiyar/RealESRGAN_x4_CoreML/resolve/main/RealESRGAN_x4_pal-4.mlpackage.zip) | 7.66 MB | | | pal-4 | ### Performance Benchmarks (RealESRGAN x4) Tested on MacBook Pro (Apple Silicon M4 Max, 36 GB) running macOS 15.7.2 using Xcode Performance tab with Compute Units: All.
Part 1 — Model + device info | Model File Name | Model Type | Model Size | Device Type | Software Version | Compute Units | |:--------------------|:-------------|:-------------|:--------------|:-------------------|:------------------------------| | RealESRGAN_x4 | ML Program | 34.1 MB | MacBook Pro | macOS 15.7.2 | All (CPU, GPU, Neural Engine) | | RealESRGAN_x4_Q-8 | ML Program | 17.6 MB | MacBook Pro | macOS 15.7.2 | All (CPU, GPU, Neural Engine) | | RealESRGAN_x4_pal-4 | ML Program | 9.2 MB | MacBook Pro | macOS 15.7.2 | All (CPU, GPU, Neural Engine) |
Part 2 — Operations | Model File Name | Neural Engine Operations | CPU Operations | GPU Operations | Total Operations | |:--------------------|---------------------------:|-----------------:|-----------------:|:-------------------| | RealESRGAN_x4 | 955 | 0 | 71 | 1,026 | | RealESRGAN_x4_Q-8 | 955 | 0 | 71 | 1,026 | | RealESRGAN_x4_pal-4 | 955 | 0 | 71 | 1,026 |
Part 3 — Performance (Prediction / Load / Compile) | Model File Name | Metric | Average | Median | Minimum | Maximum | Uncached | |:--------------------|:-----------|:----------|:---------|:----------|:----------|:-----------| | RealESRGAN_x4 | Prediction | 198.62 | 196.61 | 191.82 | 251.10 | 251.10 | | RealESRGAN_x4 | Load | 82.63 | 83.41 | 78.56 | 84.40 | 1,946.85 | | RealESRGAN_x4 | Compile | 199.57 | 197.59 | 195.36 | 205.77 | 195.36 | | RealESRGAN_x4_Q-8 | Prediction | 201.30 | 199.74 | 192.94 | 236.59 | 236.59 | | RealESRGAN_x4_Q-8 | Load | 94.53 | 94.92 | 90.93 | 96.17 | 2,399.90 | | RealESRGAN_x4_Q-8 | Compile | 222.40 | 222.53 | 220.22 | 224.45 | 220.22 | | RealESRGAN_x4_pal-4 | Prediction | 200.81 | 199.31 | 193.33 | 259.98 | 259.98 | | RealESRGAN_x4_pal-4 | Load | 90.99 | 91.74 | 85.77 | 93.65 | 2,103.96 | | RealESRGAN_x4_pal-4 | Compile | 211.95 | 211.87 | 211.81 | 212.18 | 212.18 |
## Author Arshiya Rahgozar [GitHub](https://github.com/marshiyar) ## Attributions Thanks to the authors of Real-ESRGAN Real-world image super-resolution. Original project: [xinntao/Real-ESRGAN](https://github.com/xinntao/Real-ESRGAN) (BSD-3-Clause).