| cutout-birefnet: BiRefNet general (ZhengPeng7/BiRefNet at revision | |
| e2bf8e4460fc8fa32bba5ea4d94b3233d367b0e4), MIT, Copyright (c) 2024 ZhengPeng. | |
| See LICENSE. | |
| Changed for Core AI, in export_birefnet.py: | |
| - torchvision's deform_conv2d is computed from gathers and matrix products | |
| (identical result: max |alpha difference| 4.6e-6 in float32 on the CPU). | |
| - The ImageNet normalisation and the final sigmoid are inside the graph; only | |
| the last, full-size prediction is returned. | |
| - Weights and arithmetic are float16; the input and output stay float32. | |
| The trained weights are otherwise the authors', unmodified. | |