# Any Qwen 2.1 LoRA training tips so far? — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/9b509e297708a07e

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-09-21 12:46 ingested 2026-09-22 02:38

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1wmbsuz/any_qwen_21_lora_training_tips_so_far/)

## retained text



```
Probably too soon to ask or post this. But been fighting Qwen-Image 2.1 LoRAs in ai-toolkit for a few hours now and wondering if anyone else has it figured out or if we're all just suffering.

My experience: character LoRAs come out okay-ish, but anatomy is a constant fight melted hands, legs, feet, double mouths and chin, waxy skin, *chernobyl* mutation, extra limbs in certain poses. Even did one run at bf16 with zero quantization and it still produced body horror around the 200-1500-1750 step mark before samples slightly recovering and then returning to shit. Never had this with any other new base model since SD and SDXL.

Currently testing a bigger "corrective" setup (heavy dataset, high-res buckets, quantized base, guidance loss + blank prompt preservation on) and it's actually very slightly improving, which makes me think the base priors are just damaged and you have to force-feed it. Or I'm doing it all wrong, which is why I'm asking. And seeing not many LoRA posted yet, got me thinking if everyone is facing this.

What settings have actually worked for you guys? Rank, lr, dataset size, resolutions, anything. Curious if anyone has it training clean with a normal small dataset or if everyone's out here performing surgery.
```



document fetched 2026-09-22 00:31 · canonical https://reddit.com/r/StableDiffusion/comments/1wmbsuz/any_qwen_21_lora_training_tips_so_far
