← sources

What does a good LoRA character dataset contain?

source: reddit (reddit:StableDiffusion)published 2026-10-20 15:59ingested 2026-10-21 02:38

retained text

I know, there is tons of information out there on this topic, but while some of it is years old, others is just completely contradicting each other.

Also I assume that it might depend on which model should be trained on.

**So here is the question: If you train a character LoRA in 2026, what does your dataset contain? Headshots only? Portraits only? Upper body, full body? A mix of all? And how much of each?**

Example: I have a dataset of 33 images, headshots / portraits only.

I've trained LoRAs for SDXL and ZIT with that which turned out quite well. Training a Moss-Image LoRA with that results in the generation focussing very much on the head of the person, even when prompting for a full body shot (head disporportionally big, person leaning forward to make the head bigger, etc). Is that because we need to include more diverse dataset pictures or is the LoRA simply overbaked then? Or is it a captioning problem?

  
Thanks for any input.

document fetched 2026-10-21 00:30 · canonical https://reddit.com/r/StableDiffusion/comments/1xaw0kk/what_does_a_good_lora_character_dataset_contain