Runway introduces Solaris, an Interface World Model for real-time UI generation

• Solaris generates an interactive app or website as a live sequence of frames, with clicks, drags and typed input influencing what appears next, according to Runway’s Solaris announcement.
• The system splits the work between a language model that decides how a scene should change and a world model that renders it, a division detailed in Runway’s technical introduction and illustrated by a Solaris interaction demo.
• Runway reported that Solaris won its own comparisons against a coded interface on instruction-following and naturalness, a result the company summarized as outperforming frontier language models in its launch post.
• Access remains limited to an early-access request for partners. One early reaction noted that Solaris was not open for public testing.

A virtual showroom can turn a clothing image into a draggable try-on experience, while an x-ray can produce different outcomes from the same drag in Runway’s examples. The company is pitching the same mechanism as an agent-training environment whose layouts never have to repeat.

Solaris starts with a scene or brand environment, then streams a new visual state as the person interacts. Runway calls the result an Interface World Model: no HTML, CSS or application logic sits between the generated frame and the interface the user touches.

Its launch explanation divides the interaction style into three properties:

• Visual: a scene itself can be the application, such as a showroom where a shirt is dragged onto an avatar.
• Alive: lighting, objects and layouts can keep evolving instead of waiting for a page transition.
• Open-ended: a scene can respond to behaviors that were described in a prompt rather than hard-coded during development.

The distinction is useful for creative interface work because the interaction is part of the generated image. Runway’s examples include putting salad ingredients directly into a bowl and manipulating a combustion demonstration, rather than selecting controls from a conventional menu.

Solaris builds on Runway’s Gen-4.5 video model and follows its earlier GWM-1 research. The company adapted the model to treat user actions as conditioning signals, then paired it with a language model for intent and scene transitions.

Runway describes the real-time conversion in four stages:

1. Clicks, drags and other past actions condition the next frame.
2. Frames are generated autoregressively, each based only on prior frames.
3. A multi-step video-diffusion process is distilled to a few steps.
4. The faster model is trained on its own outputs to hold visual quality over longer interactions.

The language model interprets requests, decides whether to alter the current scene or move to another one, and writes the prompts that steer rendering. Solaris handles the pixels and responses; the language model supplies the behavioral plan. Runway’s interface vision

Runway tested the premise in two ways. Its reconstruction evaluation asked multimodal language models to rebuild 30 interfaces from a screenshot, then measured structural similarity and DINOv3-based region similarity. Runway says fidelity fell as visual complexity rose, especially for natural images.

For interaction, the company compared Solaris with a coded interface generated by Claude Opus 5 from the same starting image and requests. A 250-person user study across 30 examples collected nearly 7,500 pairwise judgments:

• Following the requested interaction: Solaris 61%, coded result 24%, equivalent 13%.
• Behaving naturally in the scene: Solaris 71%, coded result 21%, equivalent 6%.

Those are Runway’s own study results, not an independent benchmark. The comparison measures scene-level response quality, which favors Solaris’s specific claim that objects should retain physical and visual coherence through an interaction.

Runway identifies four limits that keep Solaris in research territory:

• Text: stable, legible text remains difficult at real-time video speed. Runway proposes a hybrid approach in which an image model handles text-heavy views when a short pause is acceptable.
• Trust: a persuasive but incorrect answer is risky in commercial or instructional work. The current system is grounded by its starting frame, with richer verified context still an active research area.
• Long sessions: visual and semantic consistency can still drift during extended open-ended use.
• Accessibility and integration: generated interfaces need to work with screen readers, accessibility APIs and the wider software stack.

Runway says it is working with key partners before a public launch and directs interested companies to the early-access form. A [reply to the launch](Runway’s Solaris page) asking to try GWM-1 captures the immediate constraint: neither product availability nor a general-release date was announced.