Why AI Horror Looks Fake — and How I Push It Toward Realism
AI can generate an impressive monster in seconds. Making that monster look like it was actually filmed is harder.
That difference is where most of my horror experiments live. I am less interested in making a technically perfect creature than making a viewer hesitate for half a second and wonder whether the footage could be real.
The biggest giveaway is usually the camera
Generated horror often fails before the creature even appears. The camera floats too smoothly, reframes itself at exactly the right moment or moves in a way no person holding a phone would move. That polished movement tells the viewer they are watching an effect.
For found-footage style work, I deliberately keep camera behavior simple. Sometimes that means completely stationary framing. Other times it means believable handheld movement with imperfect timing. The camera should react like a camera, not like a movie trailer.
Do not show everything
AI models are very good at producing detail and very bad at knowing when detail hurts the illusion. Horror works well when the viewer has to fill in part of the image. Darkness, fog, distance, compression and partial obstruction can all help because real footage contains uncertainty.
That does not mean hiding every flaw. It means designing the shot so the creature belongs in the environment rather than posing for the viewer.
Natural motion matters more than dramatic motion
A common failure is asking for too much at once: turn, crawl, scream, transform, run and attack in a ten-second clip. The model has to invent transitions between all of those actions, and that is where anatomy and physics usually fall apart.
I get better results when the subject has one clear behavior and one scare beat. A figure can stand still longer than feels comfortable, then move suddenly. A crawler can enter slowly and accelerate once. A ghost can drift, stop and react. Fewer actions often create a more believable result.
Photographic realism is a stack of small choices
Lighting, lens behavior, framing, motion, skin texture, clothing, depth of field and background detail all need to agree with each other. When one part looks like a polished 3D render and everything else looks like phone footage, the illusion collapses.
That is why my prompts tend to spend as much time describing the camera and environment as the monster itself. The creature is only half of the shot.
Headtap is where I document more of these experiments as I work through them. For related work, see how I build realistic ghost projection videos, or watch the finished experiments on YouTube.