If you have two dolls, what is more human than making them kiss? This is a simple workflow for generating two distinct characters interacting, using regional prompting to reduce prompt bleed. It works as follows: 1) A layout image is generated using a model that is good at composition2) CLIPSeg is used to identify the characters by a descriptor (e.g. man, woman)3) Regional conditioning is applied based on a mask4) Image-2-image is used with a second model5) Upscale and cleanup using the second modelV1: Because of the use of CLIPSeg it performs better with opposite-sex characters where gender can be used as the descriptor. However images can be very successfully generated using alternative descriptors with some thought and patience.V2: Uses a GGUF version of Chroma that is quite friendly to low-VRAM setups as the layout model. Chroma can handle positioning characters in the frame quite reliably, allowing the characters to be identified based on whether they are to the left or to the right, and the use of a yolov detector to identify the character regions.As with my previous workflows, none of this is particularly original or groundbreaking, but I've tried to present it in a tidy and understandable way in the hope that it will be helpful to others!

Kissing Dolls -- Multistep Regional Couple Workflow

Uses GGUF Chroma model for first stage, replacing Illustrious. Uses yolov8 person model for character region model, replacing CLIPSeg.

Files with SHA256: 5c7f22643f842f8926e007351bd0e2ccb6eea8c84ac50124a0914f7e9660d304