Alphelo — AI model notice
=========================

Alphelo ships two versions of the same AI model. Each is one ONNX file split into parts for hosting.

1) CPU version (used on most devices, WebAssembly)
   Files: /assets/model/birefnet-lite-512-int8-37fe1e0b/model.part00.bin … model.part02.bin
   SHA-256 of the reassembled file: 37fe1e0b34a3d770e2813ef614e8081aad0af97f3be658e3a28bccb6bf448a52
   Made by the Alphelo project from the fp32 export below (onnx/model.onnx,
   SHA-256 1cb0fb360dadd15af77c639085d77a9df67db0c64315560c3de005f676345ac2):
   graph optimisation, 8-bit dynamic quantization of the Conv, Gemm and MatMul weights (uint8),
   remaining large tensors stored as fp16. This is a modified version of the MIT-licensed model;
   the MIT license and the copyright notice below apply to it.

2) GPU version (used with WebGPU)
   Files: /assets/model/birefnet-lite-512-fp16-eff9216b/modelGpu.part00.bin … modelGpu.part04.bin
   SHA-256 of the reassembled file: eff9216bb2f9d3f023d9c2b7196845a7485739ab1f231593633e4d2344ffc516
   The fp16 export below (onnx/model_fp16.onnx), redistributed unmodified (only split into parts).

What it is
  BiRefNet_lite (Bilateral Reference for High-Resolution Dichotomous Image Segmentation),
  re-exported to ONNX at 512x512 input resolution.

Original model and weights
  Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, Nicu Sebe
  Code:    https://github.com/ZhengPeng7/BiRefNet  (MIT License, Copyright (c) 2024 ZhengPeng)
  Weights: https://huggingface.co/ZhengPeng7/BiRefNet_lite  (license: MIT)
  Full license text: BiRefNet-LICENSE.txt

ONNX export used here
  https://huggingface.co/studioludens/birefnet-lite-512
  (revision 4a3c40c36c94093cc1e724d9ea428b8fa4b57dc7, files onnx/model.onnx and onnx/model_fp16.onnx)
  The model card declares the MIT license; it is a derivative of the MIT-licensed weights above.

Citation
  Zheng, Peng et al. "Bilateral Reference for High-Resolution Dichotomous Image Segmentation."
  CAAI Artificial Intelligence Research, 2024.