World Models and the Missing Input Layer

Last week I wrote that Context Catalogs are useful today for better briefs, prompts, and product-in-context creative.

But they become critical as generative video, real-time production, and world models start to change how we experience advertising.

World model research is often referenced in terms of robotics, self-driving and simulation use cases. This makes sense. NVIDIA Cosmos, for example, is focused on helping AI systems understand and predict the physical world. But similar research and resulting model releases (coupled with chip progress) will also shape advertising creative.

• More realistic products
• More realistic environments
• More realistic motion
• More control over scenes, use cases, and product interactions
• More ability to leverage commercial context in real time instead of throwing spaghetti at the wall and hoping platforms find a match

Ethan He, who worked on Cosmos and later Grok Imagine, described the video side of world models as “real-time long horizon video generation.” If he’s right and video becomes more real-time, interactive, and context-conditioned, then the question for advertisers changes.

It is no longer only: can we generate a good ad?

It becomes: what should the system know before it starts generating?

• → What is true about the product?
• → Where can it appear?
• → What claims are allowed?
• → What scenes are off-limits?
• → What audience moment are we speaking to?
• → How should we weigh performance learnings so they shape generation?
• → What brand-side context can we expand so the platform’s private user-side signals have more room to explore and find the right match?

You can already see pieces of this future forming.

Decart is talking about live product placement in real-time generated video. Runway describes frames being “synthesized in the moment, conditioned on context.” WPP and NVIDIA have worked on AI production systems for brand-compliant, product-accurate content at scale.

This is why I think Context Catalogs are not just useful, but necessary.

AI is accelerating the creative workflow. But as generation moves inside systems, those systems will need better inputs.