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 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 (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.