File size: 2,279 Bytes
effebc8 57e8f83 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 | ---
license: cc-by-nc-sa-4.0
library_name: coreml
pipeline_tag: image-to-image
tags:
- super-resolution
- esrgan
- coreml
- non-commercial
---
# Community ESRGAN upscalers for Core ML — NON-COMMERCIAL
Two community ESRGAN models offered by [`upscale`](https://github.com/arraypress/swift-upscale) as
opt-in variants. **Both are licensed CC BY-NC-SA 4.0: non-commercial use only**, with attribution, and
anything you make from the models themselves must carry the same licence. The full licence text is in
`LICENSE`. Pictures you upscale are your own; the restriction is on the models.
| Package | Model | Author | Licence |
|---|---|---|---|
| `RealESRGAN-UltraSharp.mlpackage` | 4x-UltraSharp | Kim2091 | CC BY-NC-SA 4.0 |
| `ESRGAN-MangaJPEGLQ.mlpackage` | MangaJPEGLQ | bunzero | CC BY-NC-SA 4.0 |
Licences as listed by the authors and on [OpenModelDB](https://openmodeldb.info) (4x-UltraSharp,
1x-MangaJPEGLQ). The upstream Core ML mirror these come from
(`LocalMuseAI/coreml-realesrgan-x4plus` at `4f6c21f7c431ae4e52ae45d11d6195fd96d93fcc`, converted by
VincentGOURBIN/RealESRGAN-CoreML) labels them BSD-3; that is wrong for these two.
## Changes
None to the models: each archive matched the SHA-256 Hugging Face listed upstream and is republished
unzipped. Attribution: 4x-UltraSharp © Kim2091; MangaJPEGLQ © bunzero; Core ML conversion by
VincentGOURBIN. This repository is shared under the same licence (CC BY-NC-SA 4.0).
## MangaJPEGLQ on Core AI (`ESRGAN-MangaJPEGLQ.aimodel`, swift-upscale 2.1+)
`upscale --variant manga` crashed in swift-upscale 2.0 / upscale 0.3.0: **MangaJPEGLQ is a 1× model**
(it cleans JPEG blocks out of greyscale manga without enlarging — OpenModelDB lists `1x-MangaJPEGLQ`,
bunzero, CC BY-NC-SA 4.0), and the Core ML build was run as if it were 4×. This Core AI asset is
exported from bunzero's original weights (`1x-MangaJPEGLQ.pth`, OpenModelDB's file host) with
`export_esrgan.py`: a 256-px tile, float16, RGB in → Rec. 601 luma → the 1-channel model → copied to RGB.
Parity against the original PyTorch weights on six stock-photo tiles: **min 56.6 dB, mean 61.0 dB**
(`parity.py`). `upscale` now runs it as a clean-up pass before the default 4× model.
The Core ML `.mlpackage` builds above stay for older versions.
|