Runway published its Solaris research note on 31 August. The claim inside it is not that the model makes prettier screens. It is that the screen no longer needs code underneath it at all.
Solaris is what Runway calls an Interface World Model — a real-time system that generates an interactive interface frame by frame as you use it. Click a button and no event handler fires. The model simply renders the next frame showing what that click did. Drag a shirt onto a photograph of yourself and the model paints the result. There is no page, no component tree, no state machine. There is a stream of images that responds to your mouse.
The argument Runway makes for why this matters is the interesting part, and it is an argument about loss. Every piece of software shipped today starts as a visual idea and gets translated into an intermediate representation — code — before it can do anything. That translation is lossy in two directions. Behaviour has to be defined ahead of time, so the product ships as a frozen subset of everything a user might want to do. And visual fidelity gets sacrificed to make the thing run fast. Runway tested the first half of that claim directly: it asked state-of-the-art multimodal models, including Claude Fable 5, GPT-4o and Gemini 2.5 Pro, to reconstruct 30 web interfaces from a single screenshot. Reconstruction quality degraded consistently as visual complexity rose. Natural images fared worst, because rich visual detail does not survive a round trip through language.
The comparison Runway ran
Runway put Solaris against a coded interface generated by Claude Opus 5 — same starting image, same interaction requests — and ran a user study: 250 participants, 30 interaction examples, nearly 7,500 pairwise judgements. On “which result better follows the instruction,” Solaris won 61% to 24%, with 13% called equivalent. On “which behaves more naturally within the scene,” the gap widened to 71% against 21%.
That second number is the one to hold onto. A coded interface can usually execute the requested change; it just treats the change as an isolated update to a DOM node. A world model already carries some notion of how objects, materials and light behave, so the change lands inside a coherent scene. Runway is not claiming better engineering. It is claiming that physics comes free when you skip the abstraction.
Our take: The demos are the least important thing here. The real target is agent training. Runway says the reason today's models fumble booking a hotel or ordering groceries is that they learn one specific coded layout and cannot generalise to a slightly different one. An engine that generates novel, never-before-existing interfaces on demand is an infinite supply of training environments for computer-use agents — which is exactly the bottleneck everyone else is currently solving by buying racks of physical Mac minis. Solaris is a UI story on the surface and a data-generation story underneath.
What it cannot do
Runway is unusually direct about the limits, and they are not small ones. Legible text remains among the hardest problems in video generation, and interfaces depend on text more than almost any visual domain; Runway floats a hybrid fix where image models render text-heavy views during pauses. Coherence degrades over long sessions. And there is the trust problem: for anything instructional or commercial, a convincing wrong answer is worse than no answer, so today the model stays anchored to whatever real product imagery you seed the starting frame with. Accessibility is listed as an open question — a generated interface still has to work with screen readers and accessibility APIs, and nobody has shown that it does.
Solaris runs at 720p. It is built on Gen-4.5 and follows GWM-1, Runway's general world model. It is not a public product: Runway says it is working with key partners and is collecting early-access requests through a form.
What to watch
- Who the “key partners” turn out to be. A retail brand would signal storefronts. A model lab would confirm the agent-training thesis.
- Whether text gets solved or routed around. The hybrid image-model workaround is a tell that the pure approach does not yet cover real interfaces.
- Cost per session. Runway concedes generating every frame is still dearer than serving a page built once. That gap decides whether this is a product or a research artefact.
- Accessibility. An interface with no DOM has nothing for a screen reader to read. That is a regulatory problem in several jurisdictions, not just a design one.
The honest read: this is a research preview with a genuinely new idea and a long list of unsolved engineering. Worth understanding now, not deploying now.