Runway's Solaris explores interfaces generated frame by frame
Illustration, not documentary evidence of the event.
Runway unveiled Solaris, a research system that generates interface images frame by frame in response to clicks, drags and voice commands, according to the-decoder.com in its September 1, 2026 report, which cites Runway. The company calls it an “Interface World Model.” Rather than producing code for an interface, Solaris renders the interface directly.
The reported architecture combines a language model that determines interface changes with a world model that renders them, building on Runway’s Gen-4.5 video model. Runway proposed uses including virtual clothing try-ons, furniture rearrangement and visual lessons. These are company-described applications, not independently verified results.
For teams considering experimentation, the immediate limitation is basic usability. The report identifies unreliable text rendering, absent screen reader support and unresolved questions about reliability over long sessions. At the time of the report, Solaris remained a research effort, with Runway seeking launch partners and offering an early-access form.
The useful distinction is between an interface that can illustrate a request and one that can reliably support a task. A changing furniture scene could help someone explore an idea. A reservation interface would also need to make the selected details and outcome dependable. The supplied reporting does not establish that Solaris can bridge that gap.
Runway’s broader ambition is to replace fixed apps with interfaces generated for each user. That raises a design question beyond rendering quality: how much change helps? Adapting a scene to a request may remove navigation steps, but moving familiar controls could undermine muscle memory. Teams evaluating this approach should ask whether users can repeat actions and recognize mistakes as readily as they can request changes.
Runway also proposed Solaris as a training environment for AI agents. That is a separate use from asking people to rely on generated interfaces for everyday work. The relevant question there is whether practice with generated variations helps agents handle unfamiliar websites. The report provides no evaluation establishing that benefit.
Correction, September 19, 2026: This article was revised to attribute reported claims, remove unsupported statements and clarify the limits of the available evidence.