Flame 2026 BiRefNet - Better Results

[allklier]

With the release of Flame 2026 come some extra Matchboxes that make BiRefNet work with the inference node.

Here are some updated tests:

The setup requires bracketing the Inference node with PreBRN and PostBRN matchboxes which come with Flame 2026.

This is one of the test images from the BiRefNet team:

Here some random tests on various stock footage:

These three stock clips are all UHD resolution, ProRes files. On my Linux Flame with A6000 GPU it renders at 0.7fps which isn’t too bad.

These tests were done on the 130 epoch of the general ONNX version from the BiRefNet site: GitHub - ZhengPeng7/BiRefNet: [CAAI AIR'24] Bilateral Reference for High-Resolution Dichotomous Image Segmentation

I’m working on uploading this model to LogikPortal.

And many thanks to the ADSK team (@fredwarren @frenetj) for staying on the cause to make these models work.

[hBomb42]

Are these temporally coherent? And are they nvidia-only?

[cristhiancordoba]

@hBomb42

Are these temporally coherent? And are they nvidia-only

No and no.

[hBomb42]

Ah, ok. I am less envious now. Thank you!

[allklier]

@hBomb42 To answer your question: not entirely, but it’s better than some other tools.

Here’s a frame range of each of these shots with the matte as an overlay: Frame.io

You can see that it loses areas of low contrast. But in the areas where it latches on, it is surprisingly stable and has no or very little flicker. And in some areas, like the bridge deck, it’s missing some inner parts of the matte which could be augmented easily.

So will depend on shot and material, but it might still give you a decent headstart.

[hBomb42]

Thanks for sharing the moving tests, Jan. Agreed!

[TimC]

thanks for sharing. this is a good deal better than our previous ML matchboxes for human body extraction and really helpful for quick mockups and rough roto