Consistent people across AI renders: why most tools fail and how we fix it

Why single images are easy and campaigns are hard

Ask any image tool for a family on a balcony and you get a good family on a good balcony. Ask again for the same family in the kitchen and you get a different family. Sometimes a different balcony.

For a single post that does not matter. For a campaign it is fatal. A buyer who follows a project for months notices that the residents keep changing, and the campaign stops reading as one place. It reads as stock.

What “consistent” has to mean

Consistency in a property campaign has three layers, and a tool has to hold all three:

Layer	What must stay the same	What usually breaks
Architecture	Facade, proportions, materials, landscape	The tool redraws the building from a text prompt
Cast	The same residents or presenter, face, age, build, clothing logic	Every generation invents a new person
Style	Camera, lens, light, colour grade	Each image comes out in a different photographic style

Most tools solve the first layer badly and the second not at all.

How we hold the cast

The cast is defined once, before the first scene, as a set of reference views of each character: frontal, three-quarter, profile, a few expressions. Every scene is then generated from that set, so the character is a reference rather than a description. A description drifts; a reference does not.

For a virtual presenter the set also includes a voice. One voice per presenter, kept for every clip and every language version, so the presenter in the Arabic cut is recognisably the presenter in the English one.

How we hold the architecture

Every scene starts from the client’s render. The building is the input. We never ask a model to draw it. That is why the facade in frame ten is the facade in frame one: it came from the same file. The Brunnhalde project shows the renders and the scenes built from them on one page.

This has a limit. One render gives ten to fifteen distinct scenes before the angles repeat. A bigger campaign needs more renders. We say that before the order.

How we hold the style

Before production we agree two or three reference images or videos with the client. Lens, light and grade follow them for the whole series. The presenter video is deliberately made in the register of a phone-shot walkthrough, because that is what performs on Reels and TikTok, and because a glossy commercial look is what buyers have learned to skip.

Why this is a production problem

None of the above is a prompt. It is an order of operations: define the cast, lock the style, take the render as source, then generate. Skip a step and the campaign drifts, and no amount of prompt editing brings it back.

That is also why the client’s approval sits before production, on the shot list and the script. Fixing a generated scene point by point is hard. Deciding what the scene should be, before it exists, is cheap.

What you get

A campaign where one family moves from the entrance to the terrace, one presenter fronts every launch, and every frame is unmistakably the same building. See how a story film is made, browse the campaign work, or send a brochure and two renders for a demo on your project.