Runway unveils Solaris, a world model that generates interfaces in real time, frame by frame

Runway has unveiled Solaris, the first model in a new family of AI systems the company calls "Interface World Models" (world models for interfaces). The core idea is as simple as it is radical: instead of an operating system or an application displaying pre-designed screens that stay fixed until the next update, Solaris generates the interface itself in real time, frame by frame, responding continuously to every click, drag or user interaction. There is no pre-built app or website waiting to be used: the interface is synthesized on the fly, as if it were an interactive video generated while you interact with it.

The underlying reasoning Runway offers is this: all current software requires a "translation" from visual design into an intermediate representation (typically code) before that design can do anything. That conversion step, according to the company, is inevitably a lossy compression: every possible behavior has to be programmed in advance, and the result is an interface frozen in time, unable to adapt to interactions its developers did not anticipate. What's more, the translation process itself sacrifices visual fidelity, since a design reduced to code loses part of the richness of the original. By generating each frame and each response to user interaction directly through a single model, Solaris eliminates that intermediate step: there is no conversion and therefore, according to Runway, no loss of information.

The company credits Solaris with three new capabilities. The first is that it is entirely visual: the image is the application itself, with no implementation layer hidden behind what the user sees. Runway's example is a virtual clothing fitting room in which, from a single reference image of the user, they can take a garment off a rack and drag it onto themselves to try it on, as naturally as in a physical store. The second is that it is "alive": because it renders continuously, the scene evolves constantly instead of waiting for the user's next action (reflections change with the light, objects respond naturally when manipulated), and it is enough to ask in natural language for things like "move the table to see how it looks" or "change the color of the sofa". The third is that it is "open": whereas traditional interfaces only support the interactions their developers anticipated, Solaris can sustain very different behaviors within the same scene, without being tied to predefined workflows. Runway gives as examples building a salad by dragging ingredients into a bowl, or generating an interactive combustion demonstration in which the user can experiment with different materials and observe physically plausible reactions.

According to the article, this kind of system did not exist until now because it combines two worlds that have traditionally been separate: systems that "know things" (search engines, AI assistants) respond with static content (text, an image, perhaps an embedded video), while systems that respond in real time (JavaScript/CSS, video game engines and, more recently, interactive world models) generate rich experiences but understand nothing about the product, the task or the user's intent. Building an "interface world model" requires solving three technical challenges at once: speed (interaction stops feeling fluid beyond half a second of delay, whereas video diffusion models take seconds or minutes per clip), coherence (the interface must remain consistent —text, layout, object identity— throughout an entire session, not just in an isolated clip, and small errors accumulate the longer the generation runs), and cost (generating every frame is still more expensive than serving an already-built page once, although Runway claims the same techniques that made the system work in real time also cut its cost by orders of magnitude compared with a standard video diffusion model).

As for the architecture, Solaris starts from Gen-4.5, Runway's video generation model, adapted to understand interaction and respond in real time, and continues the path opened by GWM-1, its general world model. The system treats user input (clicks, drags) as a conditioning signal for the next frame, in the same way it would treat text or images, and because it only observes interactions that have already happened (never future ones), it learns the cause-and-effect relationship between visual action and outcome without those behaviors having to be explicitly programmed. To achieve interactive speed, Runway describes a three-stage process: first they taught the model to generate frames autoregressively (each one depends only on the previous ones); then they distilled the denoising process, which normally requires dozens of steps, down to just a few; and finally they trained that fast model on its own outputs so that visual quality remains stable across long interactions. They also separate "reasoning" from "rendering": a language model interprets the user's requests, decides when an interaction should modify the current scene or transition to a new one, and generates the prompts that guide Solaris in each state, while the world model handles generating how that state looks and behaves in real time. Runway also highlights what it calls "redefining the mouse": because interactions are described in natural language rather than programmed, any object in the scene can become a new tool (for example, clicking on a cat so that subsequent clicks apply its fur texture to whatever is touched).

On evaluation, Runway presents two experiments. The first measures the "cost of translation": they evaluated several state-of-the-art multimodal language models (among them, according to the article itself, Claude Fable 5) on the task of reconstructing web interfaces from a single screenshot, across a collection of 30 interfaces of increasing visual complexity, measuring fidelity with structural similarity (SSIM) and with DINOv3 features (to detect whether the visual content survives even if elements shift position). The result, according to Runway, is that every model loses information when reconstructing, and that loss grows with visual complexity, being greatest in natural images, which cannot be represented precisely through language. The second experiment directly compares Solaris with a coded interface generated by a state-of-the-art language model (the article mentions Claude Opus 5), starting from the same initial image and the same interaction requests. Runway ran a study with 250 participants across 30 interaction examples, collecting close to 7,500 pairwise comparative judgments, asking which result better followed the given instruction and which behaved more naturally within the scene. According to the reported data, Solaris was preferred in 61% of comparisons versus 24% for the coded version (13% rated equivalent) on following the instruction, and in 71% versus 21% (6% equivalent) on natural behavior.

The article itself is transparent about the system's current limitations. Legible, stable text remains one of the hardest problems in video generation, even though interfaces depend on it more than almost any other visual element; Runway suggests as a possible solution a hybrid system in which image models generate text-heavy views during brief pauses, while video handles continuous interaction. They also acknowledge a trust problem: in instructional or commercial experiences, a convincingly generated but incorrect response is worse than no response at all, so grounding in real product images and verified reference material is a focus of active research. Maintaining visual and semantic coherence across long, open-ended sessions remains an ongoing research area, and it also remains to be ensured that a generated interface works correctly with the rest of the software stack, including accessibility technologies such as screen readers.

Beyond the technical demonstration, Runway lays out a vision of a future in which the "app" as a unit of interaction could cease to make sense: if the operating system can generate the interface needed for any task, there is no longer any need to classify software into a fixed catalog of standalone applications (one for shopping, another for news, another for restaurant reservations), and what the user needs would simply appear, personalized. On the commercial side, a storefront would stop being a fixed design identical for every visitor and become a generated environment that preserves brand identity while adapting to each person; tutorials, for their part, would stop repeating the same sequence for everyone and instead render the next step in each user's own context, adapting to their progress.

Solaris is not yet publicly available: Runway says it is working with "key partners" for its launch and has set up an early access request form for interested companies, without offering for now a general availability date or pricing details.