Runway's Solaris Is an Interface World Model That Generates Software Screens Frame by Frame Without Code

Runway describes Solaris, released on August 31, 2026, as the first model in a new family called Interface World Models, generating software interfaces frame by frame in real time as a person interacts with them. Solaris builds on Runway's Gen-4.5 video generation model, adapted to understand interaction and to respond in real time, and it takes clicks, drags, and text input as conditioning signals. Runway reports a user study with 250 participants across 30 interaction examples and nearly 7,500 pairwise judgments, in which Solaris was preferred over coded results 61% of the time on instruction-following and 71% on natural behavior within the scene. No public launch date is given, and only an early access request form is available.

What Removing the Intermediate Representation Actually Means

The central claim behind Solaris is that it eliminates the intermediate representation between visual design and functionality. In the conventional path a screen must be translated into HTML, CSS, and JavaScript before it does anything, while Solaris draws the next frame directly without that translation step. The screen showing a pressed button is not the render of a state change in code but an image the model generated.

Interaction enters this structure as a conditioning signal. Runway states that clicks, drags, and text input are processed as conditioning, and describes a language model handling reasoning and decision-making paired with the world model that handles rendering. What to show is decided by the language model, and how it looks is produced by the world model.

Two technical choices make this feasible, and Runway names both. The first is autoregressive generation: each frame depends only on what came before, so no future frame is needed and generation can wait for the next user action and continue from there. The second is distillation, where Runway states the many-step denoising process was distilled into just a few steps, making the system "orders of magnitude cheaper to run than a standard video diffusion model." Visual quality is maintained at 720p.

The Half-Second Figure Governs Every Design Choice Here

The most practical baseline in Runway's write-up is the observation that interactions stop feeling interactive somewhere around half a second of delay. Video generation models are usually judged on visual quality and temporal consistency, but for a model that generates interfaces, latency itself becomes the product requirement. A screen that takes a full second to respond is a video, however good it looks.

Read against that constraint, the two technical choices stop looking independent. Autoregressive generation enables frame-by-frame streaming, and distillation cuts per-frame compute. The 720p figure belongs to the same logic: raising resolution raises per-frame cost and breaks the latency budget, so 720p reads more accurately as a point chosen inside that budget than as a ceiling on quality. Runway states only that visual quality is maintained at 720p and publishes neither a frame rate nor a measured latency.

The cost claim warrants the same reading. Orders of magnitude cheaper than a standard video diffusion model is a comparative figure rather than an absolute one, and the write-up does not name the model being compared against or the measurement conditions. What it costs per user-hour to hold an interface live is not established in this announcement.

The 61% and 71% Did Not Measure Software Quality

The user study used 250 participants, 30 interaction examples, and nearly 7,500 pairwise judgments, asking two questions: how well instructions were followed, and how natural behavior was within the scene. On instruction-following, Solaris took 61% against 24% for the coded result, with 13% rated equivalent. On natural behavior, Solaris took 71% against 21% for the coded website, with 6% rated equivalent.

The comparison target is the part to read carefully. What was evaluated is not a shipped product but a coded artifact built from the same instruction, and the questions asked about instruction-following and naturalness rather than correctness or reliability. The study therefore measures expressive fidelity at the mockup stage, not completeness as software. A plausible screen appearing on click and the numbers on that screen matching a real database are different problems, and only the first was in scope.

Runway's reconstruction experiment approaches the same boundary from the other side. Recreating 30 diverse interfaces from a single screenshot each, Claude Fable 5, GPT-4o, and Gemini 2.5 Pro were scored with structural similarity (SSIM) and DINOv3 feature comparison, and reconstruction quality consistently degraded as visual complexity increased. That failure of the screenshot-to-code path on complex screens is the problem setting Solaris targets.

Four Limits Runway States Itself

Four limitations appear in the announcement, and the list defines where this model currently sits. The first is text: Runway writes that stable, legible text remains one of the hardest problems in video generation. Given that most of an interface is characters, this is not a peripheral constraint.

The second is trust. Solaris cannot guarantee accuracy and requires grounding in verified context, which means the balance or inventory count on a generated screen carries no guarantee from the model that it matches a real value. The third is long sessions, where Runway states that maintaining visual and semantic coherence over extended interactions remains active research. The fourth is accessibility, which needs integration with screen readers and accessibility APIs.

Put together, the four narrow the model's current use. A screen whose text wavers, whose values are unguaranteed, that drifts over time, and that a screen reader cannot parse is a mockup rather than production software. Runway saying only that it is working with key partners toward a public launch, with no date, fits that position.

Where the Line Between Prototype and Software Falls

Set beside existing AI coding tools, Solaris is clearly not a replacement for them. Tools that generate code produce artifacts that can be verified, edited, and deployed. A model that generates frames produces an image sequence, so there is nothing for version control, tests, or code review to operate on. Leaving no artifact for the development pipeline is the structural constraint of this approach.

The model's place is therefore upstream of code: checking an idea as a screen immediately, comparing several interaction directions quickly, building a mockup for user testing. That stage has been served by tools like Figma prototypes, which wire together predefined screen transitions, and the difference with Solaris is that it also produces a screen for actions nobody defined.

That difference is also the risk. Producing a screen for undefined actions means producing plausible screens for behavior that cannot be implemented or that contradicts real data. Stakeholders reading a mockup as an implementation commitment is a problem prototyping tools have always carried, and it grows as the screens become more responsive.

What a Product Team Can Take From This Today

There is nothing a product team can do with Solaris right now. No public launch date exists, only an early access request form, and the announcement says nothing about non-English interface support or regional availability. Non-Latin script rendering deserves particular scrutiny, since Runway itself names text stability as one of the hardest open problems and composed scripts such as Korean Hangul are harder to generate than Latin characters.

What is worth borrowing immediately is the evaluation design. Running 7,500 pairwise judgments across 250 participants and 30 examples, separating instruction-following from naturalness as two distinct questions, and pairing the reconstruction test with two metrics (SSIM and DINOv3) is a template any team evaluating generative UI can reuse. Most generative-UI experiments stop at qualitative review, which makes this structure the more transferable part of the release.

What is verified as of August 31, 2026 is that Solaris is an Interface World Model built on Gen-4.5 and adapted for interaction understanding and real-time response, that it generates frames autoregressively at 720p, and that it reported 61% and 71% preference rates in a user study. Frame rate, measured latency, the baseline model behind the cost comparison, the public launch date, and language support are not published.

Source: Runway research announcement, "Introducing Solaris," August 31, 2026. Cited facts: Solaris framed as the first model in the Interface World Models family; the Gen-4.5 video generation model adapted to understand interaction and respond in real time; clicks, drags, and text input processed as conditioning signals with a language model for reasoning paired with the world model for rendering; autoregressive frame generation where each frame depends only on what came before; many-step denoising distilled into a few steps with the claim of being orders of magnitude cheaper to run than a standard video diffusion model; visual quality maintained at 720p; the statement that interactions stop feeling interactive around half a second of delay; a user study of 250 participants, 30 interaction examples, and nearly 7,500 pairwise judgments with 61% versus 24% (13% equivalent) on instruction-following and 71% versus 21% (6% equivalent) on natural behavior within scenes; a reconstruction-fidelity test recreating 30 diverse interfaces from single screenshots with Claude Fable 5, GPT-4o, and Gemini 2.5 Pro measured by SSIM and DINOv3, where quality consistently degraded as visual complexity increased; the stated limitations covering legible text, the absence of accuracy guarantees and the need for grounding in verified context, visual and semantic coherence over long sessions, and integration with screen readers and accessibility APIs; and the note that Runway is working with key partners toward a public launch with only an early access request form available.