Updated BiRefNet models for 2026.2

Flame Tools and Tech
, ,
1

With the release of 2026.2 the way BiRefNet works has changed slightly. The extra matchbox nodes are no longer needed for pre/post processing, and those parameters have migrated into the json file packaged with the .inf file.

I have uploaded several new variations that can be used for 2026.2 and are backwards compatible I believe to 2026 (in older versions you still need the matchboxes, it will just ignore the new json segments).

One new variation is an fp16 model which is smaller an runs faster. All ONNX files come from the official repository, I only re-wrapped them for Flame.

image
image818×200 12.6 KB

The 2026.2 files are the new models. We’ll delete the old one to avoid confusion.

One note - BiRefNet is not temporarily stable, so it’s a mixed bag for video. You need to inspect your matte carefully. But there can still be valid use cases.

I also did some testing with the new ‘TemporalSmoothing’ node added in 2026.2. Unfortunately it doesn’t seem to work very well in this case, so it’s not a definitive workaround to BiRefNet’s limitations.

This is the test case I used:

image
image1022×446 58.7 KB

It tried all the smoothing modes, and ‘weighted average’ gets close, but can still leave matte artifacts.

image
image828×686 25.5 KB

Hope you all find these new models uploads useful (access them via the Logik Portal).

2

I used the BiRefNet_HR-general-epoch_130 model, and I found it very good. In my case, the matte finish was perfect for my scene. I have a question: does this template have a file size limitation? I tested 1080p, and anything larger than that, and the Flame terminal says there’s insufficient memory. Even if I use another GPU with more memory, would it exceed 1080p? Or is that a limitation of the model?

3

There’s no hard size limitation, but bigger sizes require more VRam and you will get this error. Depends on your GPU. And a second GPU doesn’t help, it will only use primary.

In a pinch run it at HD and then upscale the matte. Obviously some detail loss but possibly better than nothing.

4

I tested Sammie running inside Flame, it works well, but the Bifnet model has better result keys.