How to perform the SDXL → SD1.5 zero-shot direction adaptation

[lattimorelin]

Hi🙂
I’m trying to reproduce it, but my target model is SD1.5
In the paper you mention “zero-shot adaptation of our semantic directions obtained on SDXL to SD1.5”, could you explain how this adaptation is actually performed?
Do I just take the ViT-L chunk of the SDXL Δe (first 768 dims) and plug that into target model SD1.5?
Really looking forward to your reply, thanks a lot!

[stefan-baumann]

Hi!

Do I just take the ViT-L chunk of the SDXL Δe (first 768 dims) and plug that into target model SD1.5?

Yes, that's exactly what I did. Except for having to adjust the range of scales I'm using (it tended to often need slightly higher delta scales after transfer), that's literally all you have to do.

Best,
Stefan