Every piece of software interface anyone has ever used, no matter how it was designed, ultimately runs as code — buttons, drags, and animations are all instructions executing against a program. Runway’s Solaris, announced August 31, skips that step entirely. It’s what the company calls an “interface world model”: an AI system that generates the interface itself, frame by frame, in direct response to what a person does — no HTML, no app logic, no intermediate code representation at all.
Treating a click like a prompt
Solaris works autoregressively, generating each frame based on what came before it, which is what makes real-time response possible without the multi-step denoising delay typical of diffusion image generation. Runway compresses that denoising process down to a handful of steps, then trains the accelerated model on its own outputs — a technique meant to keep the visuals stable across long sessions rather than drifting or degrading. User actions — a drag, a click, a typed word, a spoken command — function exactly like a text or image prompt normally would: conditioning signals that steer what the next frame looks like. A separate language model handles the reasoning layer, deciding whether an interaction should just tweak the current scene or trigger a full transition to something new; Solaris itself only handles turning that decision into pixels.
Runway’s own framing is that today’s alternative — a multimodal model trying to reconstruct a usable interface from a static screenshot — loses real information in the translation, and loses more of it as the interface gets more visually complex. Solaris’s bet is that skipping the code-as-intermediate-step entirely avoids that lossy translation altogether.
What it can actually do right now
The demos Runway has shown are notably concrete rather than abstract: a virtual clothing store where dragging an item onto yourself previews the fit; a scene where saying “move the table” produces a continuous, physically plausible response instead of an instant jump-cut; an x-ray image that responds differently depending on what kind of drag gesture is applied to it; assembling a salad by dragging ingredients into a bowl; clicking on an object to “pick up” its texture or drawing style for use elsewhere in the scene. In a study comparing Solaris against coded interfaces across 30 interaction examples with 250 participants, Solaris was preferred 61% of the time for following instructions correctly and 71% of the time for producing natural-feeling behavior.
The gaps are real, and Runway says so directly
Runway’s own limitations list is unusually candid for a product announcement: stable, legible text rendering in a real-time generated interface is still unresolved (the company is exploring hybrids that pair static image models with video models for text specifically); there’s no mechanism yet to verify that a generated interface is factually accurate to whatever it’s representing, so current systems anchor to reference images or starting frames rather than reasoning freely; long, open-ended sessions still risk losing coherence over time; and integration with screen readers or other accessibility infrastructure hasn’t been addressed at all. None of that undercuts how different the underlying idea is from anything else currently shipping in AI-assisted interface work — it’s a genuinely new category, not an incremental improvement on existing no-code or AI-coding tools, even with the rough edges Runway is upfront about.