AI Low Poly

Low poly is a reduction rather than a coating. The request sent from this panel asks for the entire scene — characters, objects, background — built from a minimal number of flat, visible triangles with sharp angular edges, each facet carrying a solid colour or a barely graded tone, and no smooth curves or intricate texture left anywhere. The target is a minimalist geometric model with a modern or retro video game feel, not a photograph with a mosaic laid over it.

Facets, not a blur with triangles

That distinction is the whole quality question, and the instruction is blunt about it: simplified, faceted, devoid of smooth curves and fine texture, with form defined by the fill of each polygon. A result that looks like a soft blur with lines across it has failed the brief. The optional prompt box, whose hint asks for the subject, style, composition, and important details, is where to settle the scale of the reduction — chunky obvious facets or fine, almost painterly ones — and whether the faces in the image need to stay identifiable. It is also where the colour logic goes: "large flat facets, three tones per face", "hard shadows with no gradient".

Source images that survive a mesh

• High-contrast lighting from one clear direction, because plane changes are what a facet is drawn along.
• A single subject with a strong outline against a simple background.
• Bold colour blocking — a bright jacket against a pale sky.
• Architecture, vehicles, and anything already made of flat panels.

Soft-focus portraits, fine hair, lace, and busy fabric patterns are the losing cases: everything the source image says about texture is what the request throws away, so those subjects come back generic.

Where the polygons give up on a face

Eyes and mouths first: the smallest features, and the ones you recognise a person by, and a facet can flatten them into suggestion. Then hands, which collapse into mittens when every finger becomes a set of flat planes. Then the boundary between subject and background, where a crisp edge turns into a ragged line of colour. Any lettering becomes shapes rather than letters, so check signage and logos deliberately. Then decide whether the overall angularity reads as a style choice or a jagged accident, and set the two files side by side at full size — the effect is easier to judge against what it replaced than alone.

Frequently asked questions

about the low poly effect

Will my face still be recognisable?

Treat that as the thing to verify rather than the thing to assume. The request asks for minimal detail across the whole scene, and identity in a portrait depends on exactly the small gradations around the eyes, nose, and mouth that get flattened. If recognisability matters more than the style, say so in the prompt box and compare two attempts.

Can I control how many triangles the result uses?

There is no density control on the panel, so it is described in words — chunky, large facets versus fine faceting — and then judged in the output. Two runs written at opposite ends of that phrasing is the practical way to find the level you want, since the instruction only promises a minimal number of polygons, not a specific count.

What happens to text and logos in my image?

They get rebuilt as polygons, which means letterforms are invented rather than preserved. A logo can come back as a similar shape in similar colours with unreadable characters inside it. If the frame contains a word you need, this style is a poor choice for it.

Which model does this page send the request to?

The sora-image model, with the low-poly instruction described above written into the request and anything you add in the prompt box going alongside it. Nothing on this page changes the model itself, so the way to steer the result is the wording you supply.