Xclusiv whole-body cut-out (BiRefNet lite, ONNX 576 x 1024)
The whole-body person cut-out used by the Xclusiv video editor (Green screen, Paper cut-out, Vertigo and other "you in front" effects).
- Model: BiRefNet lite by Peng Zheng et al., https://github.com/ZhengPeng7/BiRefNet (code MIT), weights https://huggingface.co/ZhengPeng7/BiRefNet_lite (MIT).
- What we changed: nothing in the weights. Exported to ONNX for a fixed portrait input of 576 x 1024 (1 x 3 x 1024 x 576, ImageNet mean/std, RGB), output 1 x 1 x 1024 x 576 logits. The deformable convolution is written with standard operators (grid_sample and a matrix multiply), giving the same numbers (max difference 5e-5 against PyTorch on a real frame), so it runs on onnxruntime 1.21 without the DeformConv operator. No retraining.
- File:
birefnet-lite-576x1024.onnx, 181,530,162 bytes, SHA-25665ecf3f8511e5cb87227a08aa9040f1414f7137cd5a6a9b0afb708ba63680ef8.
Citation
Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, Nicu Sebe. "Bilateral Reference for High-Resolution Dichotomous Image Segmentation." CAAI Artificial Intelligence Research, 2024.