ljleb/sd-mecha

SDXL rotation mixes CLIP G attention layers to_q to_k to_v

Opened this issue · 0 comments

ljleb commented

Minor bug that allows rotate to overalign by mixing the neurons of the attention layers. Multiple ways to fix:

  • switch SDXL arch txt2 keys to transformers keys
  • add a way for architecture extensions to manually convert weights to neurons
  • something else