U²-Net (Qin et al., "U²-Net: Going Deeper with Nested U-Structure for Salient Object Detection", 2020) as the standard u2net.onnx used by rembg — Apache-2.0. All credit to the U²-Net authors.

Mirrored for QtMeshEditor (epic #764), where it isolates the subject before single-image 3D reconstruction: qtmesh generate3d --remove-bg, the GUI "Remove background" checkbox, and the remove_bg MCP arg.

[1,3,320,320] float NCHW → [1,1,320,320] saliency in [0,1].

QtMeshEditor's usage notes: composite the masked subject over gray-128 for TripoSR (white reconstructs as a wall) but over white for TripoSG (its reference pipeline), then crop/re-pad to a 0.85 foreground ratio.

Apache-2.0 (upstream U²-Net). Credit: Xuebin Qin et al.