Could you disclose some model training details?

#8
by eeyrw - opened

Such as batch size and steps for different resolution stage or other tricks.

I'd love to hear more as well! It's encouraging to see an independently-trained SDXL model. Knowing more about its training would help us know how to use it effectively.

  • What sorts of prompts was this trained on?
  • Do you always send the same prompt text to both of SDXL's text encoders?
  • Was it trained at different aspect ratios? How did you make use of SDXL's micro-conditioning fields, original_size, target_size, and crop_coords?

Sorry, we're not planning to disclose our training recipe at this stage. Thanks for the questions though!

ehsanakh changed discussion status to closed

Sign up or log in to comment