Real-ESRGAN general x4v3 super-resolution model

UseTool file:
  /vendor/ai/realesr-general-x4v3.onnx

Original trained model:
  realesr-general-x4v3 from Real-ESRGAN by Xintao Wang and contributors
  https://github.com/xinntao/Real-ESRGAN
  Licensed under BSD-3-Clause:
  https://github.com/xinntao/Real-ESRGAN/blob/master/LICENSE

Redistributed file:
  4,871,181 bytes
  SHA-256 09b757accd747d7e423c1d352b3e8f23e77cc5742d04bae958d4eb8082b76fa4
  Exported from PyTorch 2.1.2, dynamic input and output dimensions,
  input "input" float32 [batch, 3, height, width], values 0..1,
  output "output" float32 [batch, 3, height*4, width*4], values 0..1.

How this build was verified:
  This ONNX export is redistributed from a public mirror, not from the
  upstream project, which publishes PyTorch checkpoints rather than ONNX.
  To establish that it contains the same trained weights, UseTool compared
  it against the ONNX export published by Qualcomm on Hugging Face
  (https://huggingface.co/qualcomm/Real-ESRGAN-General-x4v3, release
  v0.61.0), whose LICENSE file points at the Real-ESRGAN license above.

  Both models were run on the identical 1x3x128x128 input with
  onnxruntime-node 1.23.2. Over all 786,432 output values:
    mean absolute difference 0.000276
    maximum absolute difference 0.061683
  The remaining difference is export and optimisation noise; the weights
  are the same. The Qualcomm build was not shipped because it fixes the
  input to 128 x 128 and splits the weights into a second file, which
  forces small tiles and triggers known external-data problems in the
  browser runtime.

Modifications by UseTool:
  None. The file is redistributed byte for byte and its SHA-256 is verified
  during every build. The model runs entirely in the browser; no image and
  no inference data is sent anywhere.

Tiling:
  Pictures are processed in 128 x 128 tiles with an 8 pixel overlap; the
  last pixel row and column are extended at the image border.