Runway unveils Solaris, which redraws the screen with every click

According to the announcement, Solaris doesn't rely on interfaces built in advance from code. Instead, every time a user taps or drags on the screen, it draws the response frame by frame in real time. Runway is billing this as the world's first "interface world model."

To break that down: a world model is technology originally built for video-generation AI, predicting what comes next and rendering the following frame accordingly. Runway has swapped the target of that prediction from video to computer screens, treating a user's click or drag as a signal telling the system how the next screen should change.

Traditional software requires a designer to pre-write code so that pressing a button leads to a predetermined screen. Runway says Solaris does away with that blueprint entirely. In its introductory video, the company summed up the approach with the line: "The image itself is the application."

Moving video-generation tech into screen generation

World models originally developed as a technique in video and image generation, predicting the next scene and rendering the corresponding frame. In a study METAL LAB covered last August, we saw world models like Sora and Genie-3 used to simulate physical states. Runway has taken the same underlying concept and redirected its prediction target — not video frames, but the computer screen itself. Every single user action becomes a factor that determines what the next screen will look like.

A click isn't a command — it's a signal for "what comes next"

Runway compared how Solaris handles clicks and drags to how video-generation models handle text. Just as a video model treats words not as a lookup command but as a signal for the next scene to render, Solaris treats user actions the same way — not as a database query, but as a signal to generate the next frame. Because of this, there's no fixed, predetermined screen state; the same action can produce a different screen depending on context, according to the company.

Category	Traditional software	Solaris
Screen generation	Fixed screens called from prewritten code	Generated frame by frame in real time on every click
Layout	Designed in advance by developers	No predetermined layout — the image itself is the screen
Handling user input	Interpreted as a fixed command	Interpreted as a signal to generate the next frame

Performance claims and use in agent training

Runway says that when it comes to generating new interfaces, Solaris outperforms frontier LLMs on both structural similarity and information retention. The company also said Solaris opens up a new way for AI agents to learn how to operate screens, not just a new experience for human users. Runway's outlook is that the need for fixed, pre-rendered operating systems and screen structures will disappear altogether.

What you can sign up for now

Runway says it's taking early access applications for Solaris through its announcement page. The release didn't specify a launch timeline, pricing, or supported platforms — for now, the only confirmed step is applying to be considered for access.

Editor's view

Last month Runway released Ruby, a tool that converts SDR footage into broadcast-ready HDR. This time, the target isn't video at all — it's software itself. The direction is clear: Runway is taking the frame-prediction technology it has built up and pushing it beyond video. While other companies are moving toward having LLMs write code to assemble a screen, Runway is skipping the code step altogether and generating the screen directly.

If you've worked hands-on with video-generation models, this approach will feel familiar. The output varies slightly every time — and along with that comes the difficulty of enforcing a fixed, predictable flow. For parts of an interface that need to behave identically without a single deviation, like a login screen or a checkout flow, that variability could become a liability rather than a feature.

There are two things Korean service planners and startups should pay attention to here. One is that this looks genuinely appealing for quickly generating multiple screen concepts during prototyping. The other is that it's still too early for production use. The very nature of screens being generated slightly differently each time makes it impossible to run QA and accessibility checks the way they're currently done.

Whether the output can move beyond prototype quality will be the first real test of how much practical value this technology actually delivers.