How to train a LoRa

#11
by DenRakEiw - opened

Has anyone ever tried to train a Playground 2.5 LoRa with Kohya? I've trained one using Kohya_SS and SDXL settings with AdamW optimizer. Unfortunately, the images with the LoRa model didn't turn out well, and the concept wasn't trained either. Do you have any ideas on how to train LoRa models for Playground effectively?

It should not be the method of adding noise as used in DDPM.
You can refer to this training code: https://github.com/huggingface/diffusers/blob/playground-dreambooth-lora/examples/dreambooth/train_dreambooth_lora_sdxl.py

How would I create a finetune or Lora based for this model? The link shared above is 404.

I found some trick to create finetue in Foocus with adding embedding in the prompts and negative prompts, putting vae and some lora that it works well
paneer tikka.png
You can see my showcase in my youtube channet at https://youtu.be/jKpkxmnAKVw

Sign up or log in to comment