LayerD (BiRefNet) – ONNX fp16

ONNX export of cyberagent/layerd-birefnet, the matting model from LayerD: Decomposing Raster Graphic Designs into Layers (CyberAgent, ICCV 2025, code). Given a design image, it predicts the alpha of the top-most layer (text, shapes, decorations).

This file is used by Paperwell for its "Magic Layers" feature (desktop, DirectML).

How it was made

The LayerD weights were placed into the BiRefNet ONNX graph from onnx-community/BiRefNet-ONNX (revision 534d3c8) and converted to fp16. No Python code from LayerD or BiRefNet was run. Each constant in the ONNX graph was matched to the official BiRefNet weights (ZhengPeng7/BiRefNet 66f7f80) by value, and the same transform was applied to the LayerD weights (cyberagent/layerd-birefnet 679f743). The script is scripts/layerd_onnx.py.

Usage

  • Input input_image: float32 [1, 3, 1024, 1024], RGB, ImageNet mean/std normalised.
  • Output output_image: logits [1, 1, 1024, 1024]; apply sigmoid to get the alpha.
File Size SHA-256
onnx/model_fp16.onnx 489,657,211 bytes 1f232994e123095ed5ec7383addd9f4c071526e22fc08e9877758cda96af6739

License

Apache-2.0, the same as the LayerD weights. The BiRefNet architecture is MIT-licensed. All credit for the model goes to the LayerD authors (CyberAgent) and the BiRefNet authors. If you use it, please cite LayerD as described in its repository.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for Soap26/layerd-birefnet-onnx

Quantized
(1)
this model