Q-P1NG/birefnet-lite-512

BiRefNet lite 512 (ONNX), as the Q-P1NG app downloads it

An unchanged copy of the 512 x 512 ONNX export of BiRefNet lite (Peng Zheng et al., MIT) made by Studio Ludens: studioludens/birefnet-lite-512 at commit 962e67969ec6acec9c19fab76acee4f5e22e990c (onnx/model.onnx here as model.onnx). It is kept here so that the Q-P1NG app's cut-out pack downloads from a place that does not change; the app checks every file's SHA-256 before using it.

File	SHA-256	Bytes
model.onnx	1cb0fb360dadd15af77c639085d77a9df67db0c64315560c3de005f676345ac2	191877254
preprocessor_config.json	a79663b311d8404dfb329f7b45bc9f970a0c155aeedae8ac15b6b77d76789613	389
LICENSE	b5cd5faac84a0acb8a6bdb8aeef0919e528891be27216df84eab398d7c8bd2aa	1106

Input and output

• Input input_image: float32 [1, 3, 512, 512], RGB resized to 512 x 512, rescaled by 1/255 and normalised with ImageNet's mean [0.485, 0.456, 0.406] and deviation [0.229, 0.224, 0.225].
• Output output_image: float32 [1, 1, 512, 512] logits; a sigmoid gives the alpha matte, to be brought back to the picture's size.

Full precision weights only: the float16 export of the same repository computes in float16 throughout, which ONNX Runtime's CPU does not run on every phone.

License

MIT, as the original: see LICENSE (Copyright (c) 2024 ZhengPeng, Copyright (c) 2026 Studio Ludens V.O.F.).

Citation

@article{zheng2024birefnet,
  title={Bilateral Reference for High-Resolution Dichotomous Image Segmentation},
  author={Zheng, Peng and Gao, Dehong and Fan, Deng-Ping and Liu, Li and Laaksonen, Jorma and Ouyang, Wanli and Sebe, Nicu},
  journal={CAAI Artificial Intelligence Research},
  year={2024}
}