AssetFurnace/birefnet-lite-onnx

BiRefNet-lite (ONNX, native DeformConv)

BiRefNet-lite by Zheng Peng et al., converted to ONNX for AssetFurnace, which uses it to cut the subject out of a picture before turning it into a 3D model.

The weights are unchanged. The difference is the export: this file uses ONNX's built-in DeformConv layer (opset 19) instead of breaking each deformable convolution into many small steps. On an Apple Silicon Mac with ONNX Runtime 1.30 (CPU), one 1024 x 1024 cut-out:

File	Peak memory	Time
Common BiRefNet-lite ONNX export	9.5 GB	~5 s
This file	4.0 GB	1.8 s

Same output: on our test pictures the two masks agree to 99.8% or better.

• Input input_image: float32 [1, 3, 1024, 1024], ImageNet-normalised RGB.
• Output output_image: [1, 1, 1024, 1024] raw scores; apply a sigmoid to get the mask (the same contract as rembg's birefnet-general-lite, so it is a drop-in).
• Needs ONNX Runtime 1.20 or newer (native DeformConv).
• sha256 a97176abb6fe55525b11113bf5bead5a9b5866b2c0d4ac058186cabb86cadfff, 180666557 bytes.

Converted from ZhengPeng7/BiRefNet_lite revision aa62cd87eafb9cc43056d08ef3615a14628b831d. MIT licence, as the original (see LICENSE). All credit for the model to the BiRefNet authors.