What is DimFusion ?
#9
by blanchon - opened
Looking at the diffusers implementation it seems that many text layers hidden dim are kept from the SmolLM3 (not just the last one).
And each one is projected into different block of the DiT.
Is this what you called DimFusion.
I would be interested to know more about this, what's the main purpose ?
Thanks for the great question! 😊
Yes - that’s exactly the mechanism we refer to as DimFusion. We’ll be releasing a detailed paper very soon that dives deep into how it works and why we designed it that way. I’ll make sure to share it here once it’s out!
MishaF changed discussion status to closed