World models vs video models

One produces a finished result. The other maintains a situation. Almost every disappointment comes from confusing the two.

By Ivan Flugelman · 8 min read · Updated 2026-08-18

The words get used interchangeably in marketing and they are not interchangeable at all. The difference is not resolution or duration. It is whether the system is producing an artifact or maintaining a situation.

Getting this right saves you from the most common mistake in this space, which is directing a world model as if it were a video model and then concluding that world models do not work.

A video model produces an artifact

You give it a prompt, possibly an image, possibly a camera instruction. It returns a clip. The clip is finished. Everything the clip will ever be is decided at generation time.

Your control surface is the prompt and the re-roll. If the clip is wrong, you do not correct it, you replace it. That is why video work with these tools feels like fishing: you cast, you evaluate, you cast again.

This is not a criticism. For most of what gets made right now, an artifact is exactly what you want. A trailer is a sequence of artifacts. So is a campaign.

A world model maintains a situation

You give it a starting condition, then you act inside it. Your input at second twelve changes what exists at second thirteen. The system is holding something for you between frames.

Your control surface stops being the prompt and becomes behavior: where you look, where you move, what you touch, what you refuse to do. This is closer to directing an actor than to writing a brief.

It also means failure looks different. A video model fails by producing an ugly or wrong clip. A world model fails by forgetting: the corridor you walked down becomes a different corridor when you turn around, an object drifts, a face reorganizes itself. Those are memory failures, not aesthetic ones.

A video model fails by looking wrong. A world model fails by forgetting.

How to tell which one you are actually holding

Marketing language will not tell you. Three tests will:

• The turn-around test. Look at something, turn away, turn back. Is it the same thing? If yes, there is memory. If no, you have a very responsive video model.
• The consequence test. Break something, or move something, then continue. Does the change persist and affect what comes next?
• The boredom test. Stand still and do nothing. A video model keeps generating events because it is producing a clip. A world holds, because nothing is happening.

Why the distinction matters for your craft

With a video model, your world building lives in the prompt and in the edit. You define the rules, you generate against them, you reject what does not belong, and you assemble the survivors in an order that teaches the audience how to read the world.

With a world model, your world building has to live in the setup, because you will not be able to fix it later by re-rolling. The starting condition is the whole brief: the laws, the materials, the light, the vocabulary of objects, and the things that must never appear.

That is a much higher bar for the front end of the process. It rewards exactly the work most people skip.

Common mistakes

1. Prompting a world model like a video model, in one long paragraph, then blaming the model when the world drifts. Long prompts describe an artifact. Worlds want laws and constraints.
2. Judging a world model on a single still. A still cannot show state, memory, or consequence, which are the only things that distinguish it.
3. Assuming persistence you did not test. If you never turned around, you do not know whether the room remembers.
4. Treating real-time as the only thing that matters. A slow system with memory is more useful to a director than a fast one without it.

Sources and further reading

1. DeepMind Genie 3
2. GameNGen
3. Decart Oasis

The method is free. The practice is the Academy.

Everything above is the whole method. The World Building Academy is the part a page cannot carry: the fundamentals on video, a full world built end to end, the Render Stack, motion, editing, and the case studies. Or start with the free Codex.

Read next