birefnet-lite-1024-webgpu

BiRefNet-lite 1024px ONNX (WebGPU tier)

Verbatim mirror of jiabins0303/birefnet-lite-1024-webgpu (MIT), pinned for the BRPOD print-on-demand decorator's client-side background removal. Fixed 1024x1024 input, graph patched to run fully on onnxruntime-web's WebGPU execution provider (needs an adapter with maxStorageBuffersPerShaderStage >= 8). Load with { device: 'webgpu', dtype: 'fp32', model_file_name: 'model_fp16' }; it does not run on the WASM backend. Output is single-channel logits (transformers.js applies sigmoid).

Upstream weights: ZhengPeng7/BiRefNet_lite (MIT). The ONNX export methodology and files come from the source repo above; nothing was retrained or modified. Both config.json and preprocessor_config.json are byte-identical to the source.

Do not modify this repo in place. Publish changes as new commits and bump the pinned revision in BRPOD's backgroundRemovalService.ts, which loads this repo by commit sha.

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}
}