Runway's Interface World Model renders working websites as real-time video

Runway shows interfaces generated as live video

Runway has published a research preview of an "Interface World Model," a system that renders fully interactive websites and apps in real time with no code executing behind them. According to dev.to, the model treats an application as video output rather than as a program: a video-generation model paints each frame on the fly as a user clicks and drags, while a language model interprets every action and decides what should happen next.

The division of labor is the interesting part. The video model owns the pixels — what the interface looks like from moment to moment — while the language model owns the intent, deciding how the scene should respond to input. What is absent is the layer that normally connects the two: markup, application logic, and the rendering pipeline that turns written code into a display.

What the demos showed

Dev.to describes three demonstrations: a virtual shirt try-on where a garment is dragged from a rack onto a photo, a salad assembled by dropping ingredients into a bowl, and an interactive combustion-engine simulation. Each was generated one frame at a time as the user interacted, with nothing resembling a conventional codebase underneath.

Preferred by human testers

In blind testing, human evaluators preferred these generated interfaces to pages built by a leading coding model in a majority of head-to-head matchups. Dev.to reports the preference held on two measures: how naturally objects behaved on screen and how well the system followed instructions. The comparison model is not named and no exact win rates were published, so the result should be read as directional rather than definitive.

Where it currently fails

Three limitations stand out in the report. Text gets garbled. Longer sessions drift off course. And the system can produce screens that look entirely convincing while being subtly wrong — arguably the most consequential failure mode, because visual plausibility is no evidence that anything on screen is actually correct. The system is a research preview, not a finished product.

Why it matters

The core idea is an inversion of how the web works. Today, an interface is code that gets rendered; here, the interface is generated media, produced on demand for the person looking at it. If the approach scales, building an interface starts to resemble directing a model rather than writing software — the page exists only while it is being generated.

The timing matters as much as the technique. As dev.to notes, the preview landed at a point where video-generation models have become fast and cheap enough that generating an interface live is a plausible idea rather than an absurd one. The blind-testing result adds weight to that: even with garbled text and drifting sessions, the generated interfaces beat the output of a leading coding model in human preference.

None of the hard questions are answered yet — how you verify a screen that is drawn rather than computed, how state persists across sessions, and what happens to accessibility and semantics when there is no underlying code. But as a signal of direction, it is clear: the gap between "AI writes the code" and "AI is the interface" has now been demonstrated in public.