Guidance needed for Full finetuning of SDXL

[Anustup900]

I am looking to full finetune the base SDXL model, not Dreambooth / LoRa approach. I want to finetune using Kohya SS training script
So asking for best practices for those who had did this before for :

• what should be a good Learning rate to keep ?
• choice of scheduler
• choice of optimizers
• and steps needed per image
  Looking forward for some guidelines here

[maybleMyers]

How big is your dataset? Is it realistic or anime?

[Anustup900]

Hey @maybleMyers, the dataset building is in progress but it will have near to 5 to 10 k images, It is realistic

[dill-shower]

https://github.com/kohya-ss/sd-scripts/blob/sdxl/README.md#tips-for-sdxl-training

Read this please. As someone who also trains models I recommend you to avoid sdxl for full finetune. Sdxl is very slow to train. Consider using sd3 instead, dit provides much faster training and high quality results