Background-remover model used by https://capypix.com/background-remover/

Model:  IS-Net "isnet-general-use"
Paper:  Highly Accurate Dichotomous Image Segmentation, Xuebin Qin et al., ECCV 2022
Source: https://github.com/xuebinqin/DIS (code Apache-2.0; the code licence text is in DIS-code-LICENSE.txt)

Input file: isnet-general-use.onnx as published in the v0.0.0 release of the rembg project
(https://github.com/danielgatis/rembg), 178,648,008 bytes,
sha256 60920e99c45464f2ba57bee2ad08c919a52bbf852739e96947fbb4358c0d964a.

Changes made by Capypix to that file:
  1. Kept only the main output "output_image"; the 11 side outputs used for training were removed.
  2. Weight tensors with 256 or more values are stored as float16 and converted back to
     float32 by Cast nodes when the model loads, so all computation is still float32.
  3. MaxPool ceil_mode changed from 1 to 0. At the fixed 1024x1024 input every pooled
     tensor has even height and width, so the result is the same.
  4. The resulting file (90,424,005 bytes, sha256
     7a929b2e0a155ffd81806f21d9328c1c05831bd5e9e985c3041643afbdd842d5) is split into
     part0.bin ... part11.bin; manifest.json lists the size and sha256 of every part.
     Concatenating the parts in order gives the ONNX model.

Input: 1x3x1024x1024 RGB, values x/255 - 0.5. Output: 1x1x1024x1024 mask, 0..1.