>>11338103
Last multi arm hyper muscle cyclops for now, don't want to be spamming lol
>>11338115
>i though if i specify in the dataset
>that there is text tag, then the base model "would fill the gaps" by putting the said tag:text
NTA but imo the problem with text and logos is that it's super hard for the model to differentiate what kind of text is actually present in the dataset. In your case negging text would probably still render watermarks and other shit. So it's best practice to have the dataset images as 'clean' as possible. Same with censoring but in that case the models are indeed pretty decent at filling in the gaps in my experience.
>i want to star making my own loras
If you have the hardware just go for it! It's pretty rewarding when it works lol