Help with character bleed when restaging with likeness
When restaging a character/identity in a new photorealistic scene with other characters, there tends to be a lot of identity bleed—even when I clearly prompt a very different looking secondary character. I would appreciate any help on how to reduce that. Thanks
I was able to clean it up a bit by adjusting my prompting. However, it does seem to still happen fairly often when using a multi-image character sheet as a likeness reference
Thanks for the detailed report, and good that prompting helps some. Honest state of things: identity separation when restaging alongside other characters is the hardest limitation right now, and it is more pronounced with a multi-image character sheet as reference. Character sheets are something v1.2 does train on now, but they are still a bit wonky and will improve in future versions. Part of that improvement is just data: right now I hand-author all of the character-sheet training data myself in my spare time, so more of it directly means better results. If you, or anyone reading this, would be willing to contribute character reference sheets, that would genuinely help push this forward. Really appreciate you pushing on it.
I'd be happy to share the kind of sheets I'm using. I can generate a ton if that's helpful? They're just 5 photo images (4 headshots plus a body shot). If that's something you are interested in, let me know the best way to get them to you!
This is exactly what would help most, thank you. Yes please, I am very interested. Character reference sheets are the single biggest data gap right now, and the fact that you can generate a lot of them is fantastic.
Just the consistent sheets on their own are already hugely valuable. We can generate the edited versions from them ourselves. That said, if you are also able to include target images, the same character actually restaged into a scene or an edited version, that is even better, since real before/after pairs are higher quality supervision than anything we generate. Either way works, whatever is easiest for you. And the more varied the better: different characters, ages, styles, poses.
Easiest way to share is a Hugging Face dataset (public under a permissive license like MIT if you are comfortable, or private with access granted just to me, whichever you prefer). I have added you to the contributors shout-out list. Really appreciate this.