Add BiRefNet_lite web ONNX (768 and 512, fp16 weights) with model card, LICENSE and NOTICE
f01535d verified | birefnet-lite-web | |
| These ONNX files are derived from BiRefNet_lite, created by Peng Zheng, Dehong Gao, | |
| Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang and Nicu Sebe: | |
| "Bilateral Reference for High-Resolution Dichotomous Image Segmentation", | |
| CAAI Artificial Intelligence Research, 2024 (arXiv:2401.03407). | |
| Code and weights: https://github.com/ZhengPeng7/BiRefNet and | |
| https://huggingface.co/ZhengPeng7/BiRefNet_lite, MIT licence, | |
| Copyright (c) 2024 ZhengPeng. | |
| Changes made by clevr.tools (see README.md for details and validation): | |
| - deformable convolution rewritten as a per-kernel-position bilinear gather, so it fits in browser GPU memory | |
| - wide Split operators regrouped to stay within WebGPU's storage-buffer limit | |
| - separate static-shape exports at 768 and 512 pixels | |
| - float32 weights stored as float16 and cast back to float32 when the model loads | |
| - ONNX Runtime graph optimisation (basic level) applied before the weight conversion | |
| The model's weights were not retrained or fine-tuned. | |