AI Image & Video

Make AI Images Stop Looking Fake: Three Lighting Steps

I assumed AI images looked fake because the skin was too smooth. After comparing 43 prompt approaches and rating 369 images blind, upscaling, grain and film halation barely helped. What made the difference was describing the light clearly. Spend one minute on these three steps before your next generation; no new tools needed.

10/28/2026 ｜ 3 min read

What This Note Gives You

Before your next AI image (or video first frame), spend one minute rewriting the prompt with these three steps, and it becomes much harder to spot as AI. No new tools required.

One Core Idea

A fake-looking image usually isn’t about skin detail. The light has no source and no direction. By default AI gives you light that is equally bright everywhere, which reads like a studio catalog shot.

I had AI gather 28 lighting-method sources and compared 43 prompt approaches, rating each image blind, 369 ratings in total. Seventeen parameter and post-processing comparisons (upscale and redraw, sampler swaps, grain, film halation, color grading) were almost all ties or losses. In one-sentence-difference pairs, adding “sunlight cuts in at an angle through the window, with crisp-edged light patches” won, while adding “a light haze in the air and a golden rim light” turned the image gray and lost.

How To Do It

Step 1 | Name the object the light comes from

Pick one key light: a window, a desk lamp, a phone screen, a streetlight, the sun. Don’t write “beautiful light” or “cinematic lighting”; the model can’t act on adjectives like that.

Example: “the phone screen is the only light on him.”

Step 2 | State direction and edge

Which side the light comes from and at what angle; what it hits, where shadows fall, and whether edges are hard or soft.

Example: “sunlight cuts in at an angle through the window on the left, leaving crisp-edged patches on the sofa.”

Bonus: put the light behind and to the side of the subject (backlight). It reads far more three-dimensional than front light. My first version lost precisely because I wrote the sun as front-on.

Step 3 | Cut haze, glow and whole-frame filters

Words like “haze”, “golden glow”, “cinematic color grade” and “film look” make the whole image grayer and flatter. Let color come from the light source itself.

Exception: if a spotlight-type source is visible in a dark frame, you can describe that beam.

One extra rule for night scenes

Only light that actually exists in the scene should reach the subject; keep everything else dark. If the whole room is lit, it reads as daytime.

Pitfalls You Will Hit

1. Naming a lamp you don’t want lit: to keep a room dark I wrote “only the pendant lamp lights the table”, and the model drew the lamp and lit the room more. If you don’t want a source lighting things, don’t mention it.
2. Too many rules: stuffing a whole lighting glossary into the prompt scored worse than a three-sentence version (1.58 vs 1.83). Pick the two or three lines that matter most.
3. Videos that brighten as they play: dark scenes handed to image-to-video often light up the walls within seconds. Telling the prompt to keep lighting constant doesn’t work; after generation, pull every frame’s brightness back to the first frame (brightness keyframes in any editor). That cut brightness drift from +10 to +2.

Honest Result

The version I installed scored 1.75 on unseen validation scenes, tying the best external method, and beating no method (1.50) and my previous tool (1.42). Not an outright win, but “check the light before fixing the skin” has already changed how I generate images.

When You Get Stuck, Ask an AI

Paste this into Claude or ChatGPT with your own prompt:

I'm generating AI images and they look fake and flat. Here is my prompt: (paste). Only rewrite the lighting part: state which object in the scene the light comes from, which direction it enters from, and where shadows fall; remove haze, glow and whole-frame filters. Leave everything else word for word.

One-Line Summary

Check the light before the skin: name what the light comes from, its direction and edge quality, then cut the haze and glow.