Why AI images look cheap

I am currently putting together a film for a Runway contest. Stills come from Genspark, animation from Seedance, and every day I look through dozens of frames of which only a few make it into the edit.

Three years ago an image of this level would have cost a shoot day: lighting, location, crew. Now it costs twenty dollars a month. But it also costs twenty dollars for everyone else. So the question has shifted from "how do I generate a good frame" to "how do I spot a cheap one among the generated".

For a long time I could not articulate what exactly I was reacting to. I could see something was off, but not explain it. Now I use three tells, and they cover most of the rejects.

Light from nowhere

A cheap frame has no light source. The face is lit from everywhere at once, shadows are soft or missing entirely. The eye reads this as fake before you have time to think about why.

In an expensive frame you can tell where the light comes from: a window, a lamp, the sun behind someone's back. If you generate your own frames, write the light source into the prompt in plain words. "Side light from a single window" changes a frame more than any style reference you attach to it.

Too much of everything

The model loves to fill the frame. Give it room and it will add reflections, dust in the air, one more cup on the table, one more person in the background. It was trained on millions of images where all of that exists, so "busy" is its default.

Now look at frames from ads you yourself consider expensive. Most of the time there are few objects, and it is immediately clear which one matters.

The person weighs nothing

The subtlest tell. A person in a cheap frame looks pasted onto the background: they do not press into the sofa, do not crease the fabric, and their shadow lives a life of its own. Formally everything is in place, but the body does not sit in the space.

This is still the weakest spot of the generators, so weight is the first thing I check. If it holds up, the frame usually passes the other tells too.

Training the eye

Two habits helped me.

The first is watching slowly. I keep a collection of films and frames I return to, and from time to time I try to explain to myself in one sentence why each of them works. I do not always manage, but the attempt trains more than scrolling a feed.

The second habit is less pleasant: cull without mercy. If you show a client or a team twenty options, you are not giving them a choice, you are handing them your job. Pick yourself. If you get it wrong, at least you will know on what.