Comment by runeblaze

each text token is often subword unit, but in VLMs the visual tokens are in semantic space. Semantic space obviously compresses much more than subword slices.

disclaimer: not expert, on top of my head