Multi-character consistency in a single image is one of the hardest scenarios in AI generation. Pure text essentially cannot solve it. This article gives three progressive solutions.
Three solutions
Solution 1 · Regional Prompter / Couple
WebUI plug-in "Regional Prompter" splits the canvas into zones, each with its own prompt. Left half = character A, right half = character B. Success rate around 60%.
Solution 2 · Multi-LoRA + regional
Stack two character LoRAs with regional prompting so each LoRA applies to one zone only. Success rate 70–80%, but the two LoRAs must not conflict.
Solution 3 · ControlNet compositing + inpaint
- Generate A and B separately as two clean images
- Composite into a single layout in Photoshop
- Pass the comp through img2img at low denoise to unify lighting
- Inpaint transitions as needed
Success rate 90%+ at the cost of more manual work. Commercial projects typically take this path.
Pitfalls
Two LoRAs bleed
Both loaded simultaneously can swap features. Use weights 0.6–0.7 each.
Imprecise zone borders
Regional Prompter can produce "two characters sharing an eye". Plan zones generously.