Can a LoRA make image models generate ordinary-looking people instead of perfect faces?

baben7 · reddit · 2026-07-27

The author wants to train a “real people” LoRA for image generation because many models over-optimize for beauty and produce unrealistically perfect faces and bodies.

They ask whether a larger dataset than the usual 20–30 images is needed, whether training settings should differ from a standard character LoRA, and whether it is even possible to push a model away from its default aesthetic bias toward more ordinary human appearances.

Original post →

More from Multimodal

Multimodal channel →