Published by WaifuFinder. manifest.json pins 3 model files by SHA-256;
the app verifies every one of them, plus the licence, attribution and validation files,
before installation.
Real-ESRGAN x2plus — optimized Core ML
Self-contained framewise 2× upscaler for WaifuFinder. WaifuFinder downloads these files directly: no Python, no source checkpoint, no conversion step.
Contents
| Path | What |
|---|---|
RealESRGANValidated.mlpackage |
Converted, original-model-validated model |
manifest.json |
Architecture, original fingerprints, SHA-256 of every file |
validation.json |
Original-model parity, tile and product checks |
LICENSE |
Real-ESRGAN weights, BSD 3-Clause |
BasicSR-LICENSE |
Reference architecture, Apache-2.0 |
NOTICE |
Attribution and conversion summary |
The app compiles the package once at install time, on the machine that will run it, because compiled Core ML is tied to the local compiler and macOS version.
What "optimized" means
The network is the unmodified Real-ESRGAN x2plus. The conversion:
- fuses the first convolution into the following stage;
- expresses 256, 384 and 522 pixel inputs as enumerated Core ML shapes, so a small tile is upscaled at its own size instead of being padded to 522;
- records the measured parity against the original PyTorch network in the model metadata, and the app refuses to use a package that is not that validated one.
validation.json records the measurements: maximum absolute error against the
original network stays at or below 0.0069 across every supported input size,
against a 0.01 acceptance bound.
Provenance
Weights: xinntao/Real-ESRGAN,
RealESRGAN_x2plus.pth. Architecture reference:
XPixelGroup/BasicSR revision
8d56e3a.
Licence
BSD 3-Clause (weights). See LICENSE, BasicSR-LICENSE and NOTICE.
- Downloads last month
- 14