# Should I retrain my LoRA from scratch? — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/2ad1a7bafe8efe73

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-09-25 07:49 ingested 2026-09-26 02:37

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1wpq9ww/should_i_retrain_my_lora_from_scratch/)

## retained text



```
I don't know what to do 😭 A while ago, I trained my own LoRA with kohya\_ss on my PC. It took around 3 days to train, and I used about 40–50 images. I'm actually really happy with the result, but I trained it on Illustrious XL, and now I can't generate as many different characters as I'd like.

Should I train a new LoRA from scratch using a different base model, or is there a way to make my current LoRA more flexible? I'm not sure what would be the best approach

[](https://www.reddit.com/submit/?source_id=t3_1wppq2q&composer_entry=crosspost_prompt)


```



document fetched 2026-09-26 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1wpq9ww/should_i_retrain_my_lora_from_scratch
