← sources

Krea 2 raw? Turbo? what do I train on?

source: reddit (reddit:StableDiffusion)published 2026-09-29 09:50ingested 2026-09-30 02:41

retained text

Hey guys, so I kinda missed the boat when Krea 2 was the hot topic, but now that the dust has settled it's pretty clear it's the best open source model right now, and I want to have a go at training on it.

The problem is that every guide I find says something different. Some people say train on Raw and generate on Turbo, others don't, and then there's this Turbo adapter training stuff on top of that. So what's actually the go-to approach here? Any tips or tricks for Krea 2 LoRAs would be really appreciated too.

I've also been looking at the different trainers: AI Toolkit, Musubi Tuner, and Fizgig, which caught my eye recently. I did one run on Raw with AI Toolkit using the default settings, only bumping the resolution to 1024, and honestly it didn't turn out great. Most of my LoRAs so far have been on Anima with kohya sd-scripts and multi-GPU, so I'd kinda prefer a trainer that does that. AI Toolkit only lets you use one GPU per training run, which is a bit of a bummer.

Last thing: what should natural language captions look like for Krea 2? Are there any quirks I should know about, like length, style, or how to handle trigger words?

document fetched 2026-09-30 00:35 · canonical https://reddit.com/r/StableDiffusion/comments/1wt5vyq/krea_2_raw_turbo_what_do_i_train_on