# Krea 2 “Normal People” LORA advice? — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/8b042bb677c3c045

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-09-07 17:37 ingested 2026-09-08 02:40

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1w9ysn0/krea_2_normal_people_lora_advice/)

## retained text



```
I typically train on Runpod  using AI Toolkit. I would like to increase the variety and diversity of regular looking people across various, distinct ethnic groups (Italian, Ethiopian, Egyptian, Japanese, Chinese, Thai, etc very specific).

Any suggestions on training settings and dataset size/tagging?

Do you think I could knock this out in one Lora using an equal number of maybe 25-50 high quality images for each ethnicity?  Or should I make a specific LORA for each ethnic group?

I fear if I try to do them all with 1 Lora, the training will try to just find an average across all of them instead of retaining subtle characteristics from each group.

I’m also a bit confused on what Lora Rank I should choose given I’m trying to capture fine details across each ethnic group 
```



document fetched 2026-09-08 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1w9ysn0/krea_2_normal_people_lora_advice
