Question about training details

ToTheBeginning / PuLID

[NeurIPS 2024] Official code for PuLID: Pure and Lightning ID Customization via Contrastive Alignment

Apache License 2.0

2.64k stars 186 forks source link

Question about training details #49

Open Exuan148 opened 5 months ago

Exuan148 commented 5 months ago

Hello, when you calculated the layout loss and layout-sem loss, which cross-attention layers are the QKV features from? Do you use the features from all the cross-attention layers? I'm looking forward to your reply, thanks!

Exuan148 commented 5 months ago

QK features, sorry.