>>69807682>>69808217The actual effect I think is still debated (in terms of quality) but yes it's supposed fill the token limit and then start it again which gives more emphasis to whatever comes after the break.
This has a more substantial effect in XL models since you can use them to do multi-subject gens.
