This LoRA contextually fills in the green area in image1 with head from image2.
Basic prompt:
fill in the green area in image1 with head from image2
Sometimes it helps to expand the basic prompt a bit:
fill in the green area in image1 with head from image2, maintain correct anatomy (maintain identity, from back)
Mix with other face swap or consistence LoRA sometimes helps sometimes not. The green area should be specifically RGB(0,255,00). You guys can do it with Mask Editor in ComfyUI (latest), but personally I convert mask to area with RGB(0,255,0).
Best resolution for image2 (source head) is 1024 x 1024. There are no rules about background, zoom nor pose of the source head.
Workflow i'm using:
https://civarchive.com/models/2307457/aint-another-head-swap-workflow?modelVersionId=2596183
Description
Added 768x768 images to training, huge number of steps, better consistency, still not what i want, but a step forward.
FAQ
Comments (1)
Struggling with the concept i've got some observetions:
1) QIE2511 replaces head pretty well by itself, but has trouble in some specific poses and with proportions.
2) The closer look at face, the better identity is preserved.
3) Training LoRA on RTX5090 with quantization reduces the quality.
So, my final attempt assumes: light LoRA training just to correct poses and proportions; create a workflow with local upscale and stitch back the head area to the original image (seems to be hard because QIE like to sligthly decompose the original image).