realesr-general-x4v3-onnx

realesr-general-x4v3 — ONNX export (NCHW, opset 17)

Clean ONNX export of realesr-general-x4v3 (SRVGGNetCompact, 4× general image super-resolution) from the official Real-ESRGAN weights, re-hosted for AllPrivate — where every model runs in the visitor's browser and nothing is uploaded.

Provenance

• Source weights: realesr-general-x4v3.pth from the official Real-ESRGAN v0.2.5.0 release
  • sha256: 8dc7edb9ac80ccdc30c3a5dca6616509367f05fbc184ad95b731f05bece96292
• Exported with torch.onnx.export (PyTorch 2.8.0, TorchScript exporter), opset 17, fp32, constant folding on — see export_realesr.py in this repo for the exact reproducible script
• I/O: input [1, 3, height, width] (NCHW, RGB, float32 in [0,1]) → output [1, 3, 4*height, 4*width]; height/width are dynamic
• Graph ops: Add, Constant, Conv, DepthToSpace, PRelu, Resize — no Clip, no layout transposes, so it runs on both the wasm and WebGPU execution providers of ONNX Runtime Web
• Validated against the PyTorch reference: max abs deviation ≤ 5.5e-6 across 64×64, 256×256 and 100×37 inputs

Unlike the original network's common community exports, the output is not clipped to [0,1] inside the graph; clamp at the consumer (u8 conversion clamps naturally).

License

BSD 3-Clause, inherited from Real-ESRGAN (© Xintao Wang). This repo only re-packages the officially released weights in ONNX form.