search

derrian-distro / LoRA_Easy_Training_Scripts

A UI made in Pyside6 to make training LoRA/LoCon and other LoRA type models in sd-scripts easy
GNU General Public License v3.0
1.05k stars 103 forks source link

content_or_style: balanced # content, style, balanced #238

Closed zethfoxster closed 5 days ago

zethfoxster commented 2 weeks ago

does this param exist in easy lora scripts? or can it be added as an extra arg?

Jelosus2 commented 2 weeks ago

Do you know where this lies in the code of the sd scripts or do you know any trainer that has it?

Jelosus2 commented 2 weeks ago

Ok, so, for what I read in the issues page of the sd scripts repo it seems that the mode "balanced" of the option content_or_style (which is in AI-toolkit) is what sd scripts applies by default and then there's an argument called timestep_sampling which has a new value called "shift" that shifts the timesteps (which is apparently what AI-toolkit does with the other two options). The latest commit of the sd3 branch of sd scripts was pulled 7h ago so make sure to update the trainer before you try :)

zethfoxster commented 2 weeks ago

i train flux on the weekends as the trainings always take the whole day :) ill be sure to give it a try on a fresh pull. btw does this mean its being added to the gui in some form? or is it to be added as a extra arg?

Jelosus2 commented 2 weeks ago

I'm not sure if it will be added to the UI but you can use it in the extra args section

zethfox commented 2 weeks ago

Alright thanks for the heads up, as long as it works. I know you said what it does above but can you explain in layman term why that would shift the model towards content or style? Do you know?

Jelosus2 commented 2 weeks ago

I don't know the math behind but I know is because of the timestep distribution