BiRefNet lite for WebGPU (640 × 640)

A re-export of ZhengPeng7/BiRefNet_lite (revision aa62cd8) that runs in ONNX Runtime Web on WebGPU. It is NardoCut's high-quality background removal, run on the person's own computer.

Input: 1×3×640×640, RGB 0…1 (the ImageNet normalisation is inside the graph). Output: the person's matte, 1×1×640×640, 0…1 (the sigmoid is inside the graph).
What changed: the deformable convolution (torchvision.ops.deform_conv2d), which the public exports turn into ops WebGPU cannot run, is written as bilinear grid_sample × modulator + a 1×1 convolution: the same arithmetic (2·10⁻⁷ on the matte). Fixed shape, weights unchanged. Script: tools/birefnet/ in the NardoCut repository.
Speed: ~0.5 s a frame on a laptop RTX 3050 in ONNX Runtime Web 1.30.
License: MIT, as the original (LICENSE, Copyright (c) 2024 ZhengPeng).

Downloads last month: -; Downloads are not tracked for this model. How to track

Model tree for NardoCreations/birefnet-lite-webgpu

Base model

ZhengPeng7/BiRefNet_lite

Quantized

(26)

this model