Real-ESRGAN (RRDBNet) — ONNX
ONNX exports of the standard Real-ESRGAN RRDBNet upscalers, used by
SceneWorks to run the image_upscale
job on Apple Silicon via onnxruntime + the CoreML execution provider — so the macOS
build needs no Python/PyTorch (epic 3482 "Python Eradication", sc-3489).
| file | factor | params | input | output |
|---|---|---|---|---|
real_esrgan_x2.onnx |
2× | RRDBNet 23-block | input [1,3,h,w] f32 RGB [0,1] |
output [1,3,2h,2w] |
real_esrgan_x4.onnx |
4× | RRDBNet 23-block | input [1,3,h,w] f32 RGB [0,1] |
output [1,3,4h,4w] |
- Opset 17, dynamic height/width (the consumer tiles at 512 px with 16 px pad).
- fp32. CoreML EP runs these directly; CPU EP is the fallback.
Provenance
Exported 1:1 from the canonical Real-ESRGAN weights — RealESRGAN_x2plus.pth /
RealESRGAN_x4plus.pth (mirror nateraw/real-esrgan,
originals from xinntao/Real-ESRGAN releases)
— with no weight changes. The export reproduces the exact RRDBNet architecture the
SceneWorks torch path uses; verified parity against the torch reference (onnxruntime-CPU
PSNR ≈ 101 dB, CoreML PSNR ≈ 72–74 dB / visually identical).
Reproduce: scripts/spikes/sc3489_export_reference.py.
sha256 real_esrgan_x2.onnx 7115ba92e8a1bfa63d68558ef006ef3d91273a068d321b1439f8bb1c9179002c
sha256 real_esrgan_x4.onnx 5c586662929cbc686c1a5c38d9c060dbdb4ea5863a1f7672b8c0761e6b89c033
License
BSD-3-Clause, following the upstream Real-ESRGAN weights.