How to Make AI Photos Look Real — Grain, Light, Imperfection

Everyone can spot a fake AI photo, but almost nobody can tell you why it's fake. The skin is too smooth. The light comes from nowhere. Every surface is clean, every line is straight, everything is centered. The image has no history — and that's what the brain flags before you've consciously processed a single pixel.

Making AI photos look real is therefore not about more quality. It's about deliberate imperfection. Real photos are full of small failures, and you have to prompt those failures back in.

Grain Is Not Optional

Digital-perfect smoothness is the number one tell. Film grain does two jobs at once: it adds the texture the brain associates with real photography, and it physically masks the too-perfect rendering of skin and fabric that gives generations away.

This is why "grainy aesthetic" appears in virtually every prompt in our archive. It's not a style preference — it's a realism tool. Variants that work: "shot on 35mm film," "visible film grain," "slightly soft focus." If your generations look like smartphone HDR ads, grain is the first fix.

Light Has to Come From Somewhere

Fake images are lit like video games: everything visible, nothing committed. Real photographs are lit by one dominant source with real consequences — shadows fall away from windows, faces go half-dark under streetlights, backgrounds blow out behind a subject in front of a bright sky.

Prompt the source and direction explicitly: "soft window light from the left," "late sun low behind me, lens flare, my face partly in shadow," "single warm table lamp, rest of the room falling dark." Then accept the consequences — a partly shadowed face is more believable than a perfectly exposed one. If every part of your image is well lit, none of it is credible.

Break the Composition, Slightly

AI defaults to centered subjects, level horizons, and full-body symmetry. Phone photos taken by humans have none of those. Ask for the human version: "candid, shot from slightly below eye level," "subject off-center," "mid-stride, caught between movements," "framed slightly loose with extra space on the right."

The word "candid" is doing enormous work across the archive for exactly this reason. It shifts pose, gaze, and framing all at once — from portrait mode to caught-in-passing.

Give Surfaces a History

Rendered worlds are new; the real world is used. Add age to every material your scene touches: "weathered teak decking," "worn leather seats," "scuffed marble floor," "linen slightly wrinkled from sitting." One or two patina words per scene are enough.

The same goes for your outfit. "White shirt with natural wrinkles at the elbows, sleeves pushed up unevenly" beats a pressed-flat render every time. Clothes that have been worn for four hours look different from clothes that were generated four seconds ago — describe the four-hour version.

Imperfect Humans Are Real Humans

Request natural texture in the subject: "natural skin texture," "slight stubble," "hair slightly moved by wind." And keep expressions mid-state — "relaxed, about to smile," "glancing off-camera" — rather than full grins, which models render with uncanny precision.

Your reference photo sets the ceiling here: a clean, well-lit, honest reference produces a face the model can render naturally. A filtered selfie produces a filtered-looking output, and no prompt can fix it.

The Scroll Test

The bar is not "undetectable under forensic zoom." The bar is: does a thumb pause and keep scrolling without a flag going up? Grain, directional light, off-center candid framing, and used surfaces get you past that bar reliably.

Every prompt in the PROMPTMVSTR archive already carries these techniques — the grain, the lighting conditions, the candid framing, the guardrails — because each one was iterated until the output passed exactly this test. Copy a prompt, keep the imperfections in, and resist the urge to clean them up. The flaws are the realism.