生成式介面／世界模型
Runway Solaris Replaces Programmatic Interfaces With Frame-by-Frame Generation, but Text, State, and Accessibility Still Fall Short
Runway’s Solaris treats clicks, drags, and text input directly as conditioning signals for a video model, synthesizing the next 720p interface frame in real time. This research preview demonstrates the possibility of interaction without HTML/CSS, but latency, cost, weights, and reproducible evaluations have not been disclosed.
Runway unveiled Solaris on August 31, positioning it as the first “Interface World Model.” Unlike approaches in which a model first generates HTML, CSS, or JavaScript, Solaris treats the current screen and the user’s clicks, drags, and input as conditions and directly predicts the next frame. The interface is therefore a continuously generated stream of pixels rather than a fixed DOM with prewritten event handlers.
The underlying system is adapted from the Gen-4.5 video model. According to Runway, the team first converted full-video batch diffusion into autoregressive generation that depends only on existing frames. It then distilled dozens of denoising iterations into a small number of steps and continued training the fast model on its own outputs to reduce quality drift during extended interactions. A separate language model interprets intent, determines how the scene should transition, and generates rendering prompts, creating a two-layer architecture in which the “LLM reasons and the world model renders.”
In an official evaluation involving 250 participants and 30 interactions, Runway collected nearly 7,500 pairwise judgments. Compared with programmatic versions generated by Claude Opus 5, Solaris received preference rates of 61% for instruction following and 71% for natural behavior. However, this was a vendor-designed subjective test. Runway did not release the test set, latency distribution, hardware configuration, or failure rate, and the results do not demonstrate that Solaris can correctly execute deterministic business logic such as payments or form validation.
A more immediate limitation is that text remains unreliable, while long sessions may accumulate semantic drift. A pixel-only interface also lacks the structural tree required by screen readers, keyboard navigation, and automated testing. Solaris currently accepts partner applications only, with no API, pricing, or public release date announced. Engineering teams should next watch for support for auditable state, event replay, permission boundaries, and a structured accessibility layer. Until then, Solaris is better suited to dynamic experiences and agent-training environments than as a general-purpose replacement for conventional applications.