Cool Idea - Not sure it works that well with people or is better than Ace++ or Redux

[sruckh]

Image attached for reference

I tried changing steps, and different inferences, but I never got the "people" to look anything like the source images. I tried different prompts, including the position of the tags, but nothing gave me better results than what I could get from a straight text2image workflow.

I appreciate the opportunity to test and to see the great effort in the AI Image generation space, but it seems like this is better for inpainting-type use cases rather than trying to composite people into photos with accuracy.

I did not try to just put a single person in and then use that generated image to add a second person, something to test later.

Thank You.

[sborse3]

Hey @sruckh, are any of ACE++ or redux able to do multi-person composition well? Could you point me to any other models for this task.

[sruckh-kubra]

I think I got a workflow from Civitai, don't have the exact link near me, that basically does that. It basically creates a 50/50 white/black image the size of your latent. It uses that image for the mask. It goes through the first iteration using specific LORAS and text, and then creates an inverse of that mask and runs through a second run with different LoRas and prompts. Then it stiches the two together. It even has hand and face detailers (which if I remember correctly also have their own LORas) after that. Finally at the end going through a Super Upscale. It took quite a long time for a final composite image, and it was still a 50/50 bet if the Image was crap or not. Sometimes the stiching was poor, and others the Character Lora did not seem to have enough influence on one side or the other. Plus, sometimes an exact 50/50 split did not make the most sense. So in theory, it seemed like a good approach, but the results were far from being perfect.

[sruckh-kubra]

On a side note I have also taken a manual approach. I create two different images using the same seed, but changing LORA's and text to describe each character LORA. Then I take both those images (to Photoshop) and manually blend them by typically just swapping one face our the other. Again, it sometimes take a few generations even when keeping the seed the same to get images that would blend without too much work..

[sborse3]

Thanks @sruckh-kubra for your inputs!