← sources

Questions about finetuning a Krea2

source: reddit (reddit:StableDiffusion)published 2026-10-28 10:36ingested 2026-10-29 02:40

retained text

Hey,

  
I have done a bunch of loras from SDXL to Krea2 locally, slow, yet doable. Kohya, Onetrainer, now fizgig which seems to be the king in this place.

Recently, I have tried to use runpod for the first time and I am very much pleased. Quick and not very expensive for an hour's training of a lora.

And there I began to think...

If the cost isn't 'too much', how about finetuning the entire model to include all my datasets already prepared, captioned, masked and used for creating loras. Like.... 20 persons or more datasets to finetune one Krea2 base model, without conflicting each lora in a stacked usage.

  
So I am asking. How much cost should I expect from doing that, realistically?

I mean, it all differs from gpu core/vram to storage size and system ram, and then the running time itself.

How many hours do you normally spend in finetuning such a large model? How much have you paid to fine-tuning a model?

Can anyone shed a light in this regard?

document fetched 2026-10-29 00:33 · canonical https://reddit.com/r/StableDiffusion/comments/1xhki9a/questions_about_finetuning_a_krea2