← sources

People claiming to train a LoRA in 30 mins or less locally...you're lying right?

source: reddit (reddit:StableDiffusion)published 2026-10-19 22:44ingested 2026-10-20 02:38

retained text

I keep seeing this comment, most recently with regard to MOSS, and it just doesn't compute for me. Over the past 3 years I've trained hundreds of LoRAs across image/video/audio models and I always need to run my image LoRAs for 3k-6k steps (for a dataset between 12-20 images) for a true 1:1 likeness.

I'm left to assume that the people saying this are either 1) Renting frontier hardware, 2) Lying, 3) Settling for an extremely vague likeness after \~500 steps.

Is anyone **actually** able to train LoRAs that fast locally with good results? If so, please share your method!

document fetched 2026-10-20 00:31 · canonical https://reddit.com/r/StableDiffusion/comments/1xabitu/people_claiming_to_train_a_lora_in_30_mins_or