license: bsd-3-clause
library_name: coreml
pipeline_tag: image-to-image
tags:
- super-resolution
- real-esrgan
- coreml
- apple-silicon
Real-ESRGAN for Core ML (upscale)
The Real-ESRGAN builds the upscale tool and its
library fetch on first use, as plain .mlpackage folders: 256×256 → 1024×1024 tiles, ×4.
| Package | Model | Size | Use |
|---|---|---|---|
RealESRGAN-x4plus.mlpackage |
RealESRGAN_x4plus | 33 MB | upscale --fast: the quick general model |
RealESRGAN-x4v3.mlpackage |
realesr-general-x4v3 | 2.4 MB | small and fast, slightly softer |
RealESRGAN-x4v3-denoise.mlpackage |
realesr-general-wdn-x4v3 | 2.4 MB | x4v3 that also removes noise and JPEG blocks |
Provenance
Original models: xinntao/Real-ESRGAN (Xintao Wang et al.),
BSD-3-Clause — the licence is copied in LICENSE. Core ML conversions by
VincentGOURBIN/RealESRGAN-CoreML, as mirrored
by LocalMuseAI/coreml-realesrgan-x4plus at 4f6c21f7c431ae4e52ae45d11d6195fd96d93fcc. Each archive
matched the SHA-256 Hugging Face listed there and is republished here unzipped and unchanged.
Not mirrored here: the UltraSharp and MangaJPEGLQ builds that repository also carries — they are
CC BY-NC-SA 4.0, not BSD-3 as the upstream card says, and live in arraypress/upscale-nc with their
real licence.
Measured
On an M3 Max, x4plus upscales 1024×1024 → 4096×4096 in 4.5–4.6 s (Core ML, CPU + Neural Engine).
In a ×4-shrink-and-back test on 5 generated and 5 stock photos it scored LPIPS 0.195 / 0.236
(generated / real) — upscale's default is now SwinIR-L (arraypress/upscale-swinir-l-real), which
scored 0.185 / 0.208; x4plus stays as the fast option.