Runway's Project Continuum Turns Your Desktop Into Live AI-Generated Video

Runway’s research arm has demonstrated Project Continuum, an experimental operating system shell that generates selected interface regions as live video. Those regions respond to clicks, drags and resizing while the host environment continues to run conventional software. Continuum remains an unreleased concept demo, but it shows how Runway plans to apply its Solaris interface world model to a desktop-like environment.

Four surfaces, four jobs

• Portals are animated overlays that appear while a user browses. A real-time video model generates a visual explanation, place or simulation related to the page, and the user can drag it beside the source material.
• Responsive Video Interfaces recompose generated scenes when their display area changes. In the demo, narrowing a window causes a staircase to become a spiral staircase, allowing the scene’s geometry to fit the available space.
• Interactive Worlds expand an image, clip or thumbnail into a navigable environment. People and software agents can then move through the generated simulation.
• Visual Thinking gives coding agents a live visual trace. The model depicts each execution step and tool call, providing another way to monitor long autonomous workflows.

Solaris renders the response

Project Continuum supplies the shell, while Solaris generates the interactive surfaces frame by frame. Runway describes Solaris as an Interface World Model: it receives mouse actions and other events as conditioning signals, then predicts the visual state that should follow. In plain terms, each action helps determine the next generated frame.

Runway says Solaris combines autoregressive generation, few-step distillation and training on its own outputs. Autoregressive generation builds each frame from the preceding state and latest action. Few-step distillation compresses generation into fewer inference operations, reducing latency. Training on generated output is intended to improve coherence during longer interactions.

Solaris draws on Runway’s Gen-4.5 video model and follows its GWM-1 general world model research. Within the generated regions shown in Continuum, developers do not define a DOM, component tree or individual event handler for every visual response. The model produces the interface pixels directly, leaving open how production applications would expose controls, state and semantics to other software.

Why generated pixels matter

Conventional interfaces translate a design into components, layout rules and application logic before rendering it. Runway’s approach lets a model carry visual intent directly into interactive output. A generated surface could adapt its composition to the available space, current task or user context while retaining a chosen visual style.

For developers, that approach would shift some interface behavior from deterministic code into probabilistic inference. It could support layouts and simulations that are expensive to specify by hand, while introducing new requirements for testing, state management, reproducibility and accessibility.

Visual Thinking maps most directly to an existing workflow problem: supervising coding agents during long tasks. A visual execution trace could expose state changes and tool activity that are difficult to follow in a fast-moving text log.

Research limits show through

Solaris remains a research preview with unresolved performance and consistency constraints. Technical reporting cites a 720p output target and per-frame latency below 500 milliseconds, although independent testing has not verified those figures. Related Runway world models have shown visual and geometric drift during rapid camera movement, inconsistent details over long sequences and limited image referencing after the initial frame.

Accessibility remains unresolved because generated pixels do not inherently provide the semantic structure that screen readers and other assistive technologies use. A production system would need a reliable accessibility layer covering controls, labels, focus order, state changes and keyboard navigation.

The developer contract is missing

• Integration: Runway has published no API, SDK or embedding documentation for Continuum.
• State and replay: The demo does not explain persistence, deterministic reproduction, undo behavior or recovery after a model error.
• Performance: Hardware requirements, network demands, concurrency limits and inference costs remain undisclosed.
• Testing: Runway has not described how teams would inspect, automate or validate generated interactions across model versions.
• Security: The available material does not define authorization boundaries for model-generated controls or safeguards for data passed into generated surfaces.

Access remains gated

Project Continuum is currently available only as a Runway Labs video demonstration, and Solaris access requires a request. Runway has announced no pricing, public API or downloadable release. The work outlines a possible interface architecture in which generated surfaces coexist with HTML, CSS and component frameworks as another UI runtime.