Here is my first attempt at creating a multi-character LoRA without bleeding, using LoRA isolation and Differential Output Preservation, DOP. It´s not 100% bullet proof yet as you can see on a few pictures, hairstyle for example. But that can be fixed by prompt as well. But it nail it 9/10. All characters used are my own creation and can be found separate with the same name in their collections.
Here’s how it works: The main trigger the red section is always included in the prompt (if you want 3 people). After that, you can address each woman individually using only her trigger word (see the green words). I’ve added two wildcards—one for NSFW and one for SFW—so you can easily test them out. If you use wildcard, the trigger is baked in. I haven’t personally generated all the images using the wildcards yet, so I can’t guarantee the results. Watch each version for triggers.
" Lux54woman is an adult woman with copper-red hair. Brat21woman is an adult Persian woman with dark hair. brat25woman is an adult Japanese woman with dark hair. Full-body photograph outside an upscale nightclub at night, all three visible from head to toe, colorful neon reflections on the pavement. Lux54woman wears a fitted emerald satin dress and black high heels, holding a small clutch. Brat21woman wears a burgundy cocktail dress and gold heels, resting a hand on her hip. brat25woman wears a black sequined dress and silver heels, smiling toward the camera. Natural skin texture, realistic proportions, distinct faces, coherent shared scene."
I’m including the entire trigger word, but you can separate the ladies later if you want—or just run one at a time.