Why Does the AI Keep Inventing a Different Room?

Updated 2026-10-27·3 min read·Consistency

Why it happens

A location still goes to the model as one reference image. One image tells the model what the place looks like from one angle. It doesn't tell it what's behind the camera, or what the ground looks like from a different height. So when the angle changes, the model invents those parts.

We hit this on a manor's front drive. The location still showed a big circular emblem set into the gravel. Its written description talked about a landscaped island with a tree instead, and never mentioned the emblem. The one feature that made the drive recognizable lived only in the picture, so every shot invented its own ground: flowers in one, bare gravel in another, nothing in a third.

There were more causes in the same scene:

• Two locations, one piece of ground. The front door's description said rose beds sat at the bottom step. The drive's description said gravel. Whichever location a shot named decided the floor.
• Features in the wrong place. Putting the emblem into text sent with every shot of that location made it show up at the foot of the steps in a close two-shot, at the wrong size.
• No scale. "Gravel" told the model nothing about size, so a turning circle for cars kept coming back as a garden path.

The rule behind all of it is that a reference image beats words about the same thing. A close-up insert kept coming back as a wide room shot while the room's reference was attached; dropping the room got it first try. A phone stayed pink because its prop image was pink.

How to fix it

1. Look at the reference, then rewrite its description to match. A description that disagrees with its own picture is a trap for every shot after it.
2. Name the identifying features in words, with position: "a large compass rose inlaid in the gravel at the center of the drive."
3. Give the scale as well as the material. "A broad circular gravel drive, wide enough for a car to turn around on."
4. Keep the always-on text to what every angle sees. The surface and scale go with every shot; a feature in one part of the place goes only in the shots that show it.
5. Make neighboring locations agree where they meet.
6. Describe what IS there. "Bare gravel from her knees to the bottom step" beats "no flowers."
7. When a picture contradicts the prompt, list the references first. For a tight insert, you may not want the room reference at all.

Example

A detective drama has a diner booth. Shot 1 is a wide; shot 4 is a close-up of a coffee cup. If shot 4 also sends the whole diner's reference, it may come back as another wide. Send the cup and the booth, describe "a chipped white mug on a red formica table, booth seat edge out of focus," and leave the room out.

How ShowMaker handles it

In ShowMaker, locations have reference stills and shots cite them as @slugs, so the same room goes to the model on every shot that names it, and you can see exactly which locations a shot uses. When one frame drifts, Shoot with AI redraws it from a plain note. More in keeping AI locations consistent and the cast and assets docs.