# Lora bleeding into unrelated prompt — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/c5d16553037477be

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-10-30 11:34 ingested 2026-10-31 02:38

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1xjcx54/lora_bleeding_into_unrelated_prompt/)

## retained text



```
Hi everyone, i'm looking for some advice since on the internet and via LLMs i couldn't find a certain answer.  
I'm making a LoRA, it's my first experience with this so i'm surely lacking some knowledge.  
I'm using kohya locally to train, on a dataset with 26 images created using this guide ( [https://huggingface.co/AICU/SDXL-LoRA?utm\_source=chatgpt.com#english](https://huggingface.co/AICU/SDXL-LoRA?utm_source=chatgpt.com#english) ).

The results are actually great, my character ( a girl ) shows when using the LoRA and the drawing style and face/hair/body features are all there, so i'm actually kinda happy.

The only problem, as the title says, that she shows up even if i'm not using her activation token.

The base model is this: [https://civitai.red/models/827184/wai-illustrious-sdxl?modelVersionId=2883731](https://civitai.red/models/827184/wai-illustrious-sdxl?modelVersionId=2883731)

This is all the tries i did:  
\- training with <my\_token>, 1girl, solo, ...  
\- training with <my\_token>, ... ( basically removed 1girl, solo )  
\- training with <my\_token>, ... ( removed 1girl, solo ) and adding 26 regularization images.

also i'm training locally on an RTX 3060ti, 8GB VRAM, batch 1 and rank 8 ( i can do rank 16 but it's slower so for testing i went for 8, switching to 16 if needed when finding good results ).  
If you need more infos about something i'm more than happy to share.

I'm kinda confused because the results are there but it would be perfect, of course, if the character didn't bleed in every picture when the tag is not used.

Any advice?

EDIT: SOLVED!!  
Thanks to u/[Massive-Health-8355](https://www.reddit.com/user/Massive-Health-8355/) for the hint on Fizgig!  
I've been able to train a small test LoRA which achieved exactly what i wanted.  
here's a result ( mind that my caption is poor, maybe even the dataset, it was just a test )

*Processing img dn4abg02joyh1...*

From left to right: base model only, lora without activation token, lora with activation token.  
All 3 with the same prompt ( apart from the token ) and settings.

Many thanks to everyone!
```



document fetched 2026-10-31 00:31 · canonical https://reddit.com/r/StableDiffusion/comments/1xjcx54/lora_bleeding_into_unrelated_prompt
