>>44550733it occured to me the other day when someone on /h/ asked how to stop images wider than 768 from putting in 2girls even though it has 1girl in the tags and 2girls in the negatives. (too newfag to crossboard link so I'll just spoiler it here)
Preface: this is not how the ai actually works, this description just happens to coincidentally match how the ai appears to behave. This is a way of thinking about your composition, not an explanation of the process.
Image your image space having a 512*512 box sliding over it. As it slides the ai looks at the box and goes "is there 1girl in there, is there blonde, is there covered nipples?" And it checks everything in the box for all your positives and negatives and makes the content of the entire box fit your request. The ai cannot see the entire composition at once, just what's in the box. It also doesn't have object permanence. If there is something outside the box, it doesn't know it exists, even if it saw it earlier in the pass. It also requires that everything you asked for is in the box when it looks at it.
When your image is 700 wide, the box overlaps itself enough that you get 1 girl. At 700, there's still 300px of overlap between the farthest left and farthest right box. Girl is couple hundred pixels wide, all is good.
But when your image is 1024 wide, well now there's vast areas without that overlap so it puts a girl in, slides on over "I don't see 1girl here" puts another girl in. If the girls are small enough it might even be able to slide over some more and put in 3 or 4 girls into an image with 1024 width.
It's this lack of object permanence and the narrow scope of the ai's view that causes this problem.
Solution: latent couple lets you put 1girl in the area you want the girl and then you leave background tags in every other region. It will prevent this spawning behavior because latent couple allows you to update what parameters the box needs to fit based on where in the image it is. 