Help: LoRA training fails to capture consistent character details

puskur · reddit · 2026-08-26

A user is facing issues training a LoRA using Ostris AI Toolkit. Despite using a dataset with detailed captions (pose, expression, attire), the output (at 3150 steps) differs noticeably from the reference and misses key details like a neck mole. The issue persisted across different step counts and even with Flux.2. The user is seeking insights into dataset or parameter settings.

Original post →

More from Apps

Apps channel →