Runway’s Solaris Generates App Interfaces Frame by Frame, but It Isn’t Ready for Public Use

Runway is testing a different boundary for software: instead of opening a prebuilt interface, a user would interact with a screen that is generated anew as they click, drag or type. Its newly introduced Solaris model renders those interfaces frame by frame, making the visual scene itself the application rather than a display sitting above conventional interface code.

Runway introduced Solaris on August 31 as the first member of what it calls its Interface World Models family. The system is built on the company’s Gen-4.5 video model and GWM-1 world model, extending video generation toward a use case where the output must respond continuously to a person’s actions.

An interface rendered instead of implemented

Solaris does not use an intermediate code representation to produce the interface. It generates visual frames sequentially, and actions such as clicks and drags condition what appears next. That is a meaningful architectural change from an interface whose screens and behaviors have been defined before a user arrives.

The division of labor matters. A separate language model interprets the request, decides how a scene should change and steers the renderer. Solaris then creates the visual response one frame at a time. Runway says autoregressive generation, step distillation and self-training make that loop fast enough for real-time rendering at 720p.

61%Instruction following

In Runway’s study, 250 participants made nearly 7,500 pairwise judgments. Solaris was preferred over a coded result based on Claude Opus 5 in 61% of instruction-following comparisons.

71%Natural interaction

The same study found Solaris preferred in 71% of comparisons over the coded result when participants judged whether the interaction felt natural.

The evidence is promising, but company-run

Runway evaluated Solaris in a reconstruction benchmark against GPT-4o, Gemini 2.5 Pro and Fable 5, using SSIM and DINOv3 similarity measures. The company also compared the system with a coded Claude Opus 5 result in its user study. Those tests support a case for visual fidelity and interaction quality, but they are Runway’s own evaluations rather than independent production evidence.

Four problems that cut against production use

• Text rendering remains unstable, a direct problem for interfaces that depend on clear, persistent labels and information.
• The system can produce inaccuracies, so a visually convincing interface is not necessarily grounded in correct information.
• Consistency across long interactions is still a stated limitation, even though Solaris is intended to react over a continuing session.
• Accessibility support is incomplete, including integration with accessibility APIs used by assistive technology.

A research system, not a software platform

Solaris is not publicly available. Runway is accepting early-access requests and working with launch partners, but it has released no code, model weights or API details. The immediate question is therefore not whether developers can substitute Solaris for a web stack today, but whether its visual interaction can hold up once these reliability and access constraints are addressed.

For now, Solaris is best understood as a wager that an interface can be generated like an interactive video rather than authored as fixed screens. It establishes a technical direction—one where a language model chooses the next behavior and a world model renders it—but Runway’s own limits show how far that direction remains from dependable everyday software.