How to generate synthetic character data for LoRA with limited originals?

BeautifulStation4 · reddit · 2026-08-27

A user discusses strategies for generating synthetic data to train a high-likeness character LoRA when starting with only 3 original images that vary in style, angle, and lighting. Key questions include the feasibility of synthesis, best tools for the job, the impact of missing multi-angle data on likeness, and whether training data should prioritize consistency or diversity.

Original post →

More from Multimodal

Multimodal channel →