Sub-pixel CNN (ESPCN, 2016) super-resolution model

UseTool file:
  /vendor/ai/sub-pixel-cnn-2016.onnx

Exact upstream file:
  ONNX Model Zoo, validated/vision/super_resolution/sub_pixel_cnn_2016
  https://github.com/onnx/models/tree/main/validated/vision/super_resolution/sub_pixel_cnn_2016
  File super-resolution-10.onnx

Upstream license:
  The onnx/models repository is licensed under Apache License 2.0.
  https://github.com/onnx/models/blob/main/LICENSE

Origin of the weights:
  Trained with the PyTorch super-resolution example (BSD-3-Clause,
  https://github.com/pytorch/examples/tree/main/super_resolution) on crops
  from the 200 training images of the BSD300 dataset. The architecture is
  the efficient sub-pixel convolutional network of Shi et al., 2016.

Redistributed size:
  240,078 bytes

SHA-256:
  85f36ff88cc504a24af5e0602148bc56a8aa09a58eca8c0da2756f3e8186035e

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

Model input and output:
  input  "input"  float32 [batch, 1, 224, 224], values 0..1
  output "output" float32 [batch, 1, 672, 672], values 0..1
  The network works on the luminance channel only. UseTool enlarges the
  colour of each tile by ordinary resampling and applies the network's
  luminance on top, which is the intended use of this model.

Measured against the larger model:
  Per megapixel of output, this network needed 81 ms where Real-ESRGAN
  needed 960 ms on the same machine (onnxruntime-node 1.23.2, CPU),
  a factor of 11.8.