Fitting 5 character LoRAs in one image: 4 methods tested, best hits 3/5

PoleRacers · reddit · 2026-11-19

A detailed Reddit experiment on the Anima model compares four approaches to rendering five distinct character LoRAs in one image: prompt-only (2/5 correct, features bleed), regional conditioning with global LoRAs (only 3 characters drawn), RS MultiCharacter Sampler with overlapping boxes (blended faces), and non-overlapping boxes plus per-character face pass (best at 3/5, first to render the dark-skinned character). Key takeaways: face passes can't resurrect undrawn characters, strong LoRAs override outfit prompts, and generation takes 7-8 minutes per image.

Original post →

More from Multimodal

Multimodal channel →