https://files.catbox.moe/09kd04.pngWarning, technical bullshit.
Comfy attention couple implementation is annoying because it makes my iterative upscale pass freak out from the new model dimension size.
Work-around: concat prompts into one prompt, re-use base model with lora. Overall okay.
Side-effects: some prompt bleed (only mococo should have heart-eyes, only fuwawa should have blush), weird blur around faces sometimes (probably from the depth of field tag)