An entire interface is generated in real time with Solaris

With Solaris, Runway is experimenting with an Interface World Model capable of generating a visual interface in real time, frame by frame, based on the user’s actions and instructions.

An application that no longer relies on a sequence of predefined screens, but transforms while it is being used. That is the premise behind Solaris, the first Interface World Model introduced by Runway. Every click, drag, typed message, or voice command conditions what appears next as the interface is continuously generated as an interactive scene.

Runway describes Solaris as a new kind of operating system. The term does not refer to a complete replacement for Windows, macOS, or Android. For now, it is a research prototype responsible for producing the visual and interactive part of an application without first building its screens from coded components.

In a conventional interface, buttons, menus, forms, and behaviors are defined by developers. Solaris instead begins with an initial state, such as an image, product, or branded environment, and then generates what should appear on screen. The entire image becomes the interface.

A virtual store could allow users to grab an item of clothing and place it on a representation of themselves, move products around, or rearrange the store. In another example, ingredients can be dragged into a salad bowl. An educational demonstration could illustrate combustion, while a painting could become a workspace where users draw in its style.

These interactions do not necessarily need to be defined individually. Their meaning depends on the scene and the user’s request. Clicking on a cat might change the texture or color of its fur, while the same action applied to a scientific object could trigger an explanation or transformation. A single environment can therefore support different behaviors without relying on a fixed list of programmed features.

The system separates reasoning from rendering. A language model interprets the request, determines which changes or behaviors should occur, and produces instructions to guide the scene. Solaris then generates the corresponding visual result. This structure would allow the model responsible for understanding user intent to be changed without necessarily replacing the one generating the interface.

Solaris builds on the research behind Gen-4.5 and GWM-1. Runway says it adapted its video system to understand user actions and respond with sufficiently low latency to create the impression of direct interaction. The stream targets a resolution of 720p.

Generation is autoregressive, meaning that each new frame depends on the previous frames and the actions performed in between. To accelerate the process, Runway condensed a procedure that originally required many denoising steps into a much smaller number. The model was also trained on its own outputs to improve stability during longer sessions. These technical choices are outlined in the official Solaris research post.

This approach is also intended to make the interface feel less static. Reflections, objects, and environmental details can continue to change even when the user is not clicking. The application begins to resemble a simulated environment rather than a collection of pages connected by transitions.

Runway also sees Solaris as a potential training environment for computer-use agents. These systems generally learn to navigate existing software with relatively predictable interfaces and structures. An Interface World Model could instead expose them to changing or entirely unfamiliar situations and measure their ability to adapt.

The company presents two evaluations alongside its research. The first concerns reconstructing an interface from a screenshot. Multimodal models including GPT-4o, Gemini 2.5 Pro, and Claude Fable 5 were asked to produce coded versions of 30 interfaces with varying levels of visual complexity.

The results were compared using measures of structural similarity and visual information retention. According to Runway, the accuracy of coded reconstructions declines as scenes become more visually complex. However, this experiment evaluates the ability to reproduce an image as a coded page. It does not assess the full functionality, security, or maintainability of an application.

The second evaluation compares Solaris with an interface produced by Claude Opus 5 from the same starting image and interaction request. Runway says it collected nearly 7,500 pairwise judgments from 250 participants across 30 examples. Solaris was preferred in 61% of responses for instruction following, compared with 24% for the coded interface. For natural behavior, the respective results were 71% and 21%.

These findings come from a study designed and published by Runway. They do not constitute an independent evaluation or demonstrate that Solaris outperforms language models at building complete software. The comparison focuses on selected visual interactions rather than data accuracy, information persistence, external service integration, security, or resistance to errors.

The company acknowledges several limitations. Producing stable, readable text remains difficult within a continuously generated scene. Runway is therefore considering hybrid systems that combine conventional elements with generated areas, particularly when precise information needs to remain visible.

Trust is another challenge. A visually convincing interface could display incorrect information or perform a different action from the one requested. Runway plans to ground generations more firmly in starting images and verified reference material, but has not yet detailed how outputs would be validated.

Maintaining coherence over long sessions also remains unresolved. A scene may gradually alter its objects, lose information, or drift away from its initial state. Integration with screen readers and the accessibility APIs used by existing software has yet to be solved. A fully visual application does not automatically provide the semantic structure those technologies require.

Cost represents another significant difference from conventional interfaces. A traditional website or application displays resources that have