Comparison with LTX-2.3 Audio VAE
Hello!
Can you provide the comparison with LTX-2.3's Audio VAE? I think this will be valuable especially in the realm of joint audio-video generation.
Thanks
Hi, thanks for the suggestion — joint audio-video generation is definitely an important use case.
We left LTX-2.3's audio VAE out of this release mainly because it operates at a noticeably lower sample rate (24 kHz vs. our 48 kHz), which makes a direct metric comparison tricky: we would either have to upsample their output (which flatters it on band-limited metrics and penalizes it on full-band ones) or downsample ours (which discards exactly the range our model is optimized for).
That said, the request is fair — we plan to include it in the next release, with the resampling protocol stated explicitly so the numbers are interpretable.
@funnylittleman Thank you for the answer! Looking forward to Kandinsky 6.0 crushing Minimax and co. :)