A landing page rendered live, on demand, with no HTML behind it. That’s what Runway previewed on August 31 with Solaris, which the company calls the first “Interface World Model.” There is no code in this thing. When you click, drag, or type, a video model generates the next frame of the interface in real time. Close the tab and the page stops existing.
If you build websites for a living, or spend serious budgets turning approved mockups into shippable pages, that last sentence deserves a slow read.
What Solaris actually is
Most generative UI tools still end in code. You describe a page, a model writes React, a developer cleans it up, and you ship. Solaris removes that final stretch entirely. The interface you interact with is video, rendered frame by frame while you use it. Nothing is stored anywhere as HTML or CSS. The render is the product.
Runway built it on Gen-4.5, their video generation model, retrained to treat each input as the conditioning signal for the next frame. A separate language model handles the reasoning side: it decides what should happen in the scene next. The world model draws it, at 720p.
How it runs in real time
Video generation is normally slow. Dozens of denoising steps per frame does not add up to an interface that responds like an interface. Runway’s writeup describes three moves to get there. Autoregressive frame generation, so each frame builds on the previous one instead of being drawn from scratch. Distillation, squeezing those dozens of denoising steps down to a few. Then training the fast model on its own outputs, which keeps quality from sliding during a long session.
None of this is magic. It is the same recipe that made chat streaming feel instant, applied to pixels.
The screenshot problem
The most interesting part of the announcement isn’t the demo. It’s the evaluation Runway ran alongside it.
They tested how faithfully frontier multimodal models, including GPT-4o, Gemini 2.5 Pro, and Claude Fable 5, can recreate an interface from a single screenshot. Then they measured the information loss using SSIM and DINOv3 feature comparison. The loss was real in every case. Which is Runway’s actual argument, stated plainly: if translating a picture of an interface into code always loses something, stop translating. Render the picture.
Outside coverage points to a 250-person study in which users preferred Solaris-generated interfaces over coded ones in 61% of comparisons. Sixty-one percent is not a landslide. It is, however, a strange result for the first version of a category that did not exist a week ago.
Why this lands on a marketer’s desk
It would be easy to file Solaris under “video company does a weird research demo” and move on. I think that would be a mistake.
Design-to-code is one of the quiet money pits of digital marketing. The approval cycles, the dev sprints, the QA passes, the change requests that turn a two-week build into six. If the mockup can be the product, that entire translation layer becomes optional. The pitch stops being a promise of a page and starts being the page. Agencies that bill by that translation layer should be the first to pay attention here, and probably the last to say so out loud.
Then there is the less comfortable part. If interfaces are generated per person, per session, in real time, the foundations of conversion optimization get shaky. A/B testing assumes a stable artifact with controlled variants. Web analytics assumes the page you measured is roughly the page the next visitor sees. An interface that adapts to each viewer is either the personalization dream marketers have chased for a decade or the end of comparable measurement. Possibly both.
And the use cases Runway lists are less abstract than they sound. Virtual showrooms you rearrange by dragging. Tutorials that adapt to what you have already done on screen. Training environments for agents, with layouts that never existed before the agent walked in. The showroom one is a commerce play, not a research toy.
The catches, and they are big
Early access only. No public weights, no API, no pricing. 720p is fine for a demo and not fine for a brand page.
Then the deeper problems. Cost per session is unknown. Accessibility support is unknown. And an interface made of pixels has no DOM, which means a search crawler has nothing to read. If your business depends on organic search, that is not a detail. That is the whole ballgame until someone solves it.
Treat Solaris as a preview of where interfaces are heading. Adopting it today is not an option; the gaps above are real.
What to do now
- Watch, don’t rebuild. Nothing here ships to production today, and the gaps above are real. Put it on the research list, not the roadmap.
- Pull one number this week. How many hours, sprints, or invoices did your team spend this quarter turning approved designs into live pages? That is the budget line this category of technology, from Runway or someone else, eventually shrinks. You cannot plan for the erosion of a number you have never measured.
- Rehearse adaptive content with tools you already have. Dynamic landing pages and personalization platforms exist today. If per-visitor interfaces are where things go, your measurement practice needs the rehearsal more than your design practice does.
- Ask your analytics owner one question. If you could no longer assume every visitor sees the same page, what would you actually be measuring? Worth answering before the question stops being hypothetical.
The mockup has always been an expensive promise of a product. Solaris is an early, imperfect attempt to make the promise itself shippable. I do not know whether Runway’s specific approach wins; 720p with no API is a long way from a brand site. But the question it raises does not depend on Runway winning. Once you have seen an interface with no code in it, it is hard to unsee the translation tax you are paying every day.
