I'm willing to be proven wrong, but I don't think that idea would work. They're two totally different models (and with different modalities too), so any embeddi

I'm willing to be proven wrong, but I don't think that idea would work. They're two totally different models (and with different modalities too), so any embeddings you might produce with one model will be nonsensical to another unless translated in some way.