birefnet-lite-onnx

BiRefNet_lite — browser-tuned ONNX export (static 1024², opset 17)

Custom ONNX export of BiRefNet_lite (bilateral reference network for dichotomous image segmentation, Swin-v1-tiny backbone) from the official ZhengPeng7/BiRefNet_lite weights, re-hosted for AllPrivate — where every model runs in the visitor's browser and nothing is uploaded.

Why a custom export

The community export (onnx-community/BiRefNet_lite-ONNX) throws OrtRun std::bad_alloc in ONNX Runtime Web on every EP (fp32/fp16 × wasm/webgpu, tested 2026-07 on an M-series MacBook): it is a dynamic-shape trace whose deformable convolutions decompose into GatherND/ScatterND/Clip chains that fall back to CPU on the WebGPU EP and materialize im2col tensors of hundreds of MB inside the 32-bit wasm heap.

This export replaces DeformableConv2d.forward with a numerically identical per-kernel-tap GridSample decomposition (one bilinear GridSample + 1×1 Conv per tap, accumulated — peak extra memory is one [1,C,H,W] tensor per tap) and traces with a static input shape, so all shape dynamism constant-folds away. It runs to completion on both the wasm and WebGPU execution providers of ONNX Runtime Web (verified 1.26-dev).

Provenance

• Source weights: model.safetensors at pinned revision 7838f1c
  • sha256: 4417d89795250e698c3cb0ae8df15743810065f646f48a694fdfa7ca052d0815
• Exported with torch.onnx.export (PyTorch 2.8.0, TorchScript exporter), opset 17, fp32, constant folding on, post-processed with onnxslim 0.1.94 (constant-folds the Swin attention-mask construction; Gemm fusion disabled) plus a content-hash initializer dedupe (the backbone is traced twice for the multi-scale 'cat' input) — see export_birefnet_lite.py for the exact reproducible script
• I/O: input_image [1, 3, 1024, 1024] (NCHW, RGB, float32, ImageNet normalization: (x/255 - mean) / std, mean [0.485, 0.456, 0.406], std [0.229, 0.224, 0.225]) → output_image [1, 1, 1024, 1024] logits; apply sigmoid for the [0,1] matte
• Graph ops: Add, BatchNormalization, Concat, Conv, Div, Erf, Gather, GlobalAveragePool, GridSample, LayerNormalization, MatMul, Mul, Pad, Relu, Reshape, Resize, Shape, Sigmoid, Slice, Softmax, Transpose, Unsqueeze — no GatherND, no ScatterND, no Clip
• Validated against the unpatched PyTorch reference (torchvision deform_conv2d): GridSample decomposition max abs dev 3.1e-5 vs torchvision; full-model ONNX (onnxruntime CPU) max abs logits dev 1.0e-4, max post-sigmoid dev 1.3e-9
• File sha256: 50a57872cc739192446da2a934159f957c81af8b5a161dfda8e3daa51660ca67

License

MIT, inherited from BiRefNet (© Zheng Peng et al.). This repo only re-packages the officially released weights in ONNX form.

Citation

@article{BiRefNet,
  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}
}