bgremover-onnx

bgremover-onnx

ONNX export of BiRefNet_lite, prepared to run in the browser via onnxruntime-web on WebGPU.

Used by BackgroundRemover, a free and open-source background remover that does all inference client-side.

Files

What is different from a stock export

BiRefNet's ASPP blocks use deformable convolutions. A normal export emits DeformConv nodes, and onnxruntime-web implements no DeformConv kernel on any execution provider, so the graph fails at session creation:

Could not find an implementation for DeformConv(22) node

This export rewrites them as the mathematically equivalent grid_sample plus a 1×1 convolution per kernel tap. Verified against torchvision.ops.deform_conv2d to a relative error of ~1e-15 across padding, stride, dilation and kernel-size variations; the resulting graph matches a stock export to max|diff| = 0.00022.

Reproduce with:

python scripts/export_onnx.py --model birefnet-lite --web

Do not quantise this to int8

int8 is smaller (159 MB), faster, and matches fp32 to IoU 0.9948 — on native CPU. On onnxruntime-web's WebGPU path it returns a near-all-foreground mask: measured fg = 0.974 against an expected 0.443. fp32 is what ships.

Input / output

• Input input, shape [batch, 3, 1024, 1024], float32. ImageNet normalisation: mean [0.485, 0.456, 0.406], std [0.229, 0.224, 0.225].
• Output output, shape [batch, 1, 1024, 1024], float32 logits. Sigmoid is not baked in — apply it yourself.

Licence

MIT, inherited from the upstream model. BiRefNet_lite was trained on DIS-TR, an open dataset, and is cleared for commercial use.

Upstream: ZhengPeng7/BiRefNet · paper