alastorid/RealESRGAN-x4plus-Optimized-CoreML

Published by WaifuFinder. manifest.json pins 9 model files by SHA-256; the app verifies every one of them, plus the licence, attribution and validation files, before installation.

Real-ESRGAN x4plus — FP16 Core ML

Original authors: Xintao Wang, Liangbin Xie, Chao Dong and Ying Shan. Paper: Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data. Original project. RRDBNet reference architecture by the BasicSR contributors. WaifuFinder contributes conversion, optimization investigation, validation and native integration; the original authors created the network and trained weights.

This is the official 23-block x4plus, 64 features and growth 32, with native 4× reconstruction. It is not the anime 6B model, RealESRNet or an x2 cascade. Weights, activation tensors and input/output tensors use FP16. Scalar operator parameters remain in their Core ML-required types. No INT8/INT4, palettization, compression or other weight quantization is used.

Files and native usage

RealESRGANValidated.mlpackage is a portable Core ML ML Program for macOS 15+. Compile it on the destination Mac using MLModel.compileModel(at:), then load with default specialization and MLComputeUnits.all. Compiled .mlmodelc artifacts are machine/OS dependent and are deliberately not distributed.

Input feature input and output restored use NCHW RGB tensors normalized to 0...1. Input shapes are 1×3×128×128, 1×3×256×256, and 1×3×384×384 (default); output width and height are each multiplied by four. Clamp output to 0...1 when encoding pixels. The native implementation is included under reproduce/. Color management, orientation, padding, alpha handling and tiling belong to the consumer; do not substitute x2 pixel-unshuffle preprocessing.

WaifuFinder uses 384px tensors with conservative 160px source context around 64px cores, reflects against the whole source image and crops core pixels into the final RGB bitmap. Small images up to 224px per side use one 384px tensor. It bounds output at 256 million pixels (at most 16,384px per side), checks cancellation between predictions, and reuses its loaded model and input buffer. This is a correctness-validated configuration, not a claim of optimal large-photo performance.

Validation and ANE findings

validation.json contains the original numerical checks, native device comparison, ANE hardware trace summary, and actual local/remote integration checks. Maximum FP32-reference error across the original native tensor checks was 0.006221; worst mean error was 0.000348 on normalized RGB. Integration evidence includes native 4× reconstruction with partial tiles and pixel-identical authenticated remote execution using the actual x4 model. The remote check uses loopback.

The M3 compute plan prefers ANE for all 1,024 executable operations. A noninteractive Core ML Instruments trace recorded 100 RealESRGAN ANE prediction intervals. Default .all was faster than forced CPU+GPU and comparable to CPU+ANE on the recorded fixtures. Explicit fastPrediction specialization was rejected as slower. Recorded warm tensor medians were 46.11 / 187.75 / 430.14 ms for 128 / 256 / 384 inputs. These are earlier tensor measurements, not large-image runtime timings. Sampled process footprint excludes some Core ML/ANE-service allocations. No additional profiling was performed during application integration.

Provenance and reproduction

Official checkpoint: RealESRGAN_x4plus.pth. SHA256: 4fa0d38905f75ac06eb49a7951b426670021be3018265fd191d2125df9d682f1. Real-ESRGAN tag revision: 4c05b20cb30aded73ed3731848a9fbbf70c7c0ac. BasicSR revision: 8d56e3a045f9fb3e1d8872f92ee4a4f07f886b0a; architecture SHA256: 6aa8eb3f77e918b90eeff4ebfaf8745abe0c40eb8fcc03abbc1928693d70ab58.

Use an audited offline Python environment (PyTorch 2.7.0, NumPy 2.5.4, Core ML Tools 9.0) with the included scripts or the WaifuFinder repository:

REALESRGAN_CONVERTER_PYTHON=/absolute/path/to/audited/python \
  tools/setup-native-realesrgan.sh --variant x4plus --convert-official

manifest.json pins each model and reproduction file. It also records conversion options and reference results. Normal native inference and app installation use Apple frameworks and require no Python, checkpoint download or media executable.

Licences and responsible interpretation

Real-ESRGAN weights: BSD-3-Clause (LICENSE). BasicSR reference architecture: Apache-2.0 (BasicSR-LICENSE). Preserve NOTICE and all copyright notices. WaifuFinder source excerpts are MIT; see WaifuFinder-LICENSE. AI reconstruction synthesizes detail and must not be treated as original photographic or forensic evidence. TruFor and SAFIRE in WaifuFinder continue to analyze and display original source pixels.