search

mkshing / ziplora-pytorch

Implementation of "ZipLoRA: Any Subject in Any Style by Effectively Merging LoRAs"
MIT License
507 stars 34 forks source link

Figure 3 in the Paper #17

Open scuwyh2000 opened 10 months ago

scuwyh2000 commented 10 months ago

Hi, such a great work!

I kindly want to ask the similarity in figure3. In the paper, you said "In Fig. 3 we show the mean cosine similarity values for each layer of the last U-net block for a particular content-style pair before and after applying ZipLoRA."

My confusion is which one is the last U-net block for SDXL v1.0. The last CrossAttnBlock? And why there are 34 pairs similarity in the figure3?

Thanks!