# Krea2/loras face consistency — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/9e1a14892e54ec1b

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-11-17 08:35 ingested 2026-11-18 02:39

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1xyf8xr/krea2loras_face_consistency/)

## retained text



```
Hi guys!

I'm currently trying to train some new LoRAs for Krea 2 or Moss using AI Toolkit, but face consistency is really bad most of the time.

I've tried training LoRAs with several real women, but the results are very inconsistent. Only one of them turned out well. It can be really weird

My datasets are quite similar, with around 40 pictures each, featuring different angles and environments.

Have you had the same experience? Should I stop ai toolkit ? Thx !
```



document fetched 2026-11-18 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1xyf8xr/krea2loras_face_consistency
