What happens when you press a button? Until now, that answer had to be written in code ahead of time. Solaris, released by Runway on August 31, 2026, skips that step entirely. It generates the screen frame by frame as the user interacts with it. This is the first model in a new family the company calls Interface World Models.
Not Drawing a Screen, but Generating It Continuously
In software development today, a designer's visuals are translated into an intermediate representation: code. That translation step means the interface can only respond to interactions someone anticipated in advance, and the precision of the original design gets trimmed away by implementation constraints.
Solaris has no intermediate representation. It takes clicks, drags, and typed instructions directly as conditioning for the next frame, producing the interface the way a video model produces footage. Runway's example is a virtual clothing store: pick up a shirt from a rack, drag it onto a photo of yourself to try it on, or rearrange the display. Rather than moving between pages, you reach into the environment in front of you.
The company describes three properties. Solaris is entirely visual, with nothing hidden beneath what the user sees. It is alive, continuously rendered so that lighting and reflections keep moving. And it is open-ended, able to respond to behaviors no developer defined in advance.
Turning a Video Model Into a Real-Time Engine
The technical foundation is Gen-4.5, Runway's video generation model. Solaris follows the path the company opened with GWM-1, its general world model, and rebuilds it to run fast enough for interaction.
The conversion happened in three stages. First, the model was taught to generate autoregressively, with each frame depending only on what came before. Next, the denoising process, normally dozens of steps, was distilled down to a few. Finally, the fast model was trained on its own outputs so image quality holds up over long sessions. Interaction stops feeling interactive somewhere past half a second of delay, and a standard video diffusion model that takes seconds or minutes per clip cannot meet that bar.
The division of labor is distinctive as well. A language model decides what should happen; the world model decides how it looks. The language model interprets the request, judges whether an interaction should modify the current scene or move to a new one, and writes the prompts that guide rendering. Reasoning and rendering are separated.
Runway also calls this a redefinition of the mouse. Because interactions are described in natural language rather than programmed, their meaning can change scene by scene. Click a cat and its fur color and texture transfer to whatever you touch next; click a painting and you start drawing in its style.
Preferred Over Coded Interfaces by 250 Evaluators
Runway presents two evaluations.
The first measures information lost in translation. State-of-the-art multimodal language models, including Claude Fable 5, were asked to recreate 30 different interfaces from a single screenshot, and the results were compared against the originals using SSIM and DINOv3 features. Every model lost information, with natural images suffering most, because visual density does not survive being described in language.
The second is a direct comparison. Solaris and a state-of-the-art language model (Claude Opus 5) were given the same starting image and the same interaction requests, and 250 participants made roughly 7,500 pairwise judgments across 30 interaction examples. On following the requested instruction, Solaris was preferred in 61 percent of comparisons against 24 percent for the coded result, with 13 percent rated equivalent. On behaving naturally within the scene, Solaris took 71 percent against 21 percent, with 6 percent equivalent.
The second gap is the wider one. Code can reproduce the requested change, but it treats that change as an isolated update to the screen. A model that has already learned how objects, materials, and environments behave can generate a response that stays coherent with its surroundings. That is Runway's argument.
The Remaining Problems Are Not Small
The stated weaknesses are specific.
Text is the hardest of them. Stable, legible text is one of the toughest problems in video generation, and interfaces depend on text more than almost anything else. Runway suggests a hybrid approach as the practical path: image models render text-heavy views where a brief pause is acceptable, while video models handle continuous interaction.
Trust is another. In instructional or commercial settings, a convincing wrong answer is worse than no answer. For now, the system stays anchored by building its opening frame from real product imagery and reference material, but feeding verified information into generation as a session unfolds is still an open research area. Holding coherence across long sessions, and connecting to assistive technologies such as screen readers, also remain unresolved.
Solaris is not a generally available application, an API, or an open-weight model. Runway is working with partners toward a public launch, and access today runs through an early-access request form.
Summary
Solaris removes the assumption that a screen must first be translated into code. It was preferred over a coded implementation both for following instructions and for behaving naturally, but the parts that matter most in practice, such as legible text and coherence over long sessions, are still untouched. Before talking about a future where the app dissolves as a unit, the thing to watch is what comes out of this early-access phase.
