Why does every model sneak a HORIZON LINE into my abstract prompts?
Maddening recurring problem. I ask for pure non-objective composition, color and form, no subject, and the model keeps giving me a faint HORIZON, a subtle division two-thirds up where 'sky' meets 'ground'. Even when I say 'no horizon, no landscape, no ground', there's this ghostly landscape-logic haunting the output.
I think it's because the training data is OVERWHELMINGLY landscape-oriented, up is light/sky, down is dark/ground, and that compositional prior is baked so deep the model reaches for it even in abstraction. It's like the model has a gravitational pull toward 'this is a scene with a horizon'.
Anyone cracked this? My current hack is deliberately prompting a STRONG competing structure (a dominant diagonal, a central radial burst) to override the horizon prior with something it can't read as landscape. Fighting a prior with a stronger prior. But I'm curious if there's a cleaner fix than tricking it out of its landscape habit.