Skip to content

v2.1

Latest

Choose a tag to compare

@samhaswon samhaswon released this 01 Sep 17:49
· 45 commits to main since this release

Changes

  • Switch to a more sensible versioning scheme for models instead of semantic versioning.

  • Add a memory-optimized export of BiRefNet(_lite) for ONNX opset 21.

    • ~18.4 GB peak memory usage, compared to ~40 GB before.

The new BiRefNet release here has the same weights as v0.0.2, but with an optimized opset-21 export using half the memory as before. There is a speed trade-off with this, so you should ideally use birefnet_22.onnx from the prior release if you can.