This workflow can split the image into any number of regions based on a drawn masks. All region are sampled interchangeably allowing each region to use different LoRA models. As a result, LoRA in one region have 0 influence on the others. To ensure that regions stay consistent with each other, OpenPose ControlNet is used. This also allows for interaction between the characters. You can, though don't need to provide your own OpenPose images, you can extract them from photos or generate them on the spot. The main limitation is how well the preprocessor can detect the pose in the image. For that reason providing your own image will still give the best quality, but of course requires more effort.

If all you need is 2 characters, then my Region LoRA workflow is a better option as you only draw 1 mask (instead of 1 mask per region) and the regions are divided by masked/unmasked area, which also results in a pixel-perfect border between them leading to less artifacts

The workflow includes a second pass, an upscaler and many QoL features.

More specific instructions inside

Requires:

• WAS's Comprehensive Node Suite - ComfyUI
• ComfyUI Impact Pack
• ComfyUI Inspire Pack
• ComfyUI-Custom-Scripts
• comfyui_controlnet_aux

(with ComfyUI Manager go to Manager -> Install Missing Nodes)