/vt/ - Virtual Youtubers

>>41513385
Ah yeah, I got pretty good at herding the AI over time. Not sure if being on CPU for a while made me craftier or what.

My method for this kinda stuff is:
- Prompt both characters somewhat separately (first describe one then the other). I specifically Okayu a bit thiccer so prompted accordingly.
- The first results will be gacha. You might get them fine with proper features and details on a first gen with or without high res fix. That one happened to be a single gen without edits or high res fix. If your pic is good, you are done.
If the picture happens to be a gestalt entity of the two, be amused and try again.
- If not, send it to inpaint and mask over one character, cut the other character from the prompt and process at "inpaint at full resolution", you should get proper accuracy then while preserving the pose. This fixes stuff like Korone with cat ears or Okayu with braids.
- Now do it for the other character.
- Correct anything via manual edits or inpaint, I've done it both ways (that's how I got accurate Sanas when all we had was my embed)
- Restore original prompt when satisfied, then do a final pass of DDIM, 10 steps, very very low denoise strength. That should hopefully iron out any irregularities, give it a try or two until satisfied.

I should add this kinda thing to my LORA rentry since you could technically do the same by hotswapping LORAs allowing you to do any weird character crossovers you can imagine. I'm working on the "how to make obscure characters" section during this week, so maybe I'll add these tricks to a subsection or start a separate rentry for it.