Models served by bg.alftech.space
=================================

Both run in your browser with onnxruntime-web (MIT, ../vendor/ort/LICENSE).
Nothing here is fetched from anyone else's server.

u2netp.onnx  (4,574,861 bytes)
  sha256 309c8469258dda742793dce0ebea8e6dd393174f89934733ecc8b14c76f4ddd8
  U^2-Net, small version ("U2-Netp"): Xuebin Qin, Zichen Zhang, Chenyang Huang,
  Masood Dehghan, Osmar R. Zaiane, Martin Jagersand. "U^2-Net: Going Deeper with
  Nested U-Structure for Salient Object Detection", Pattern Recognition 2020.
  https://github.com/xuebinqin/U-2-Net  - Apache License 2.0 (LICENSE-u2net.txt)
  ONNX file as published by the rembg project (MIT):
  https://github.com/danielgatis/rembg/releases/download/v0.0.0/u2netp.onnx

isnet-general-use-fp16w.onnx  (90,653,491 bytes)
  sha256 e319d72e6ad79c51bb92e22d3422a8837d35f646fb27ee4734df1f242eb122b5
  IS-Net, "general use" weights: Xuebin Qin, Hang Dai, Xiaobin Hu, Deng-Ping Fan,
  Ling Shao, Luc Van Gool. "Highly Accurate Dichotomous Image Segmentation",
  ECCV 2022.  https://github.com/xuebinqin/DIS  - Apache License 2.0 (LICENSE-isnet.txt)
  Source ONNX: https://github.com/danielgatis/rembg/releases/download/v0.0.0/isnet-general-use.onnx
  (178,648,008 bytes). Changed here: every large weight is stored as float16
  and cast back to float32 when the model loads (tools/fp16_weights.py), which
  halves the download; the network still computes in float32. Against the
  original file the masks agree with IoU 0.9998 or better (tests/README).

Training data: both models were trained by their authors on research datasets
(DUTS-TR for U^2-Net, DIS5K for IS-Net) that carry their own terms of use. The
weights themselves are published under Apache-2.0, which allows commercial use.