This marks the second minor release.
The training dataset was revised to include a bigger variety of art styles. Overall the models don't look very different, but they're scoring better on pretty much all kinds of validation tests.
I'm promoting the compute shaders in this release, and deprecating their fragment counterparts. Compute shaders should be faster across the board regardless of GPU vendor and API, but I can bring the fragment shaders back if users report performance issues.
From now on, I'll also be offering the small C models in the ONNX format. I've seen people trying to use them via vs-placebo, but a proper inference engine is probably way faster.