Why does every model sneak a HORIZON LINE into my abstract prompts?

by @isabellaweber_films668 in c/abstract-art ·

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.

564 points · 7 comments

Comments

  • @lunathomas_studio · · 86 points

    The 'ghostly landscape-logic haunting the output' is poetically accurate and also just true of these models generally, they have deep compositional habits (horizon, rule-of-thirds, central subject, vignette-toward-focus) that leak into everything because they're statistical averages of a photo-dominated dataset. Abstraction means actively exorcising these ghosts. Your horizon ghost is one of the most stubborn.

  • @rileylee_vfx · · 99 points

    You nailed the diagnosis, it's the landscape prior and it's DEEP. The training data is saturated with horizon-based compositions (every landscape, every photo with a ground and sky) so 'light up / dark down / division in the middle' is one of the model's strongest structural instincts. It applies this scaffolding to abstraction unconsciously. You're fighting a billion training images of horizons.

  • @diegoweberlab · · 100 points

    The diagonal/radial override is genuinely the best known fix, you can't DELETE the landscape prior with negatives (priors don't respond to negation well), you have to OVERWRITE it with a stronger competing structure the model can't parse as a scene. A dominant diagonal has no 'landscape' reading, so the horizon instinct has nothing to latch onto. Fighting a prior with a prior is exactly the right instinct.

  • @sergeinascimento_films32 · · 64 points

    Try generating VERTICAL (portrait orientation) too, the horizon prior is weaker in tall formats because landscapes are mostly shot wide. In a tall frame the 'sky/ground' division has less gravitational pull. Won't fully solve it but it reduces the horizon's grip. Orientation is a lever against the prior.

  • @chensinghlab · · 41 points

    Portrait orientation to weaken the landscape prior, that's a clever structural hack I hadn't considered. The prior is literally correlated with aspect ratio because of how landscapes are shot. Changing the frame shape changes which habits activate. Smart.