Runway's Solaris renders an interactive interface as a live visual stream, generating each frame in response to clicks and drags instead of converting a design into code first. The company introduced it as its first Interface World Model, built for real-time interaction rather than video clips.
- Rendering replaces the code step. A single world model synthesizes every frame and every response to input, so the visual design becomes the interface with no intermediate representation to convert.
- User actions condition the next frames. Mouse clicks, drags, and other interactions feed the model as signals for what to render next, without those behaviors being programmed ahead of time.
- A language model handles intent while the world model handles pixels. Runway separates reasoning from rendering: the LLM interprets requests and decides how a scene should change, while Solaris generates how that change looks and responds.
- Built on Gen-4.5, tuned for speed. Runway adapted its Gen-4.5 video model for interaction and real-time response, targeting 720p output and latency low enough to stay interactive.
- Early-access research system. Runway is opening Solaris through a request form with select partners and names open limitations around stable text, trust, and long sessions.
Reasoning and rendering run as two coordinated models
Solaris generates the interface one frame at a time while a language model steers what happens next. The LLM interprets a user's request, decides whether an interaction should modify the current scene or transition to a new one, and produces the prompts that guide each rendered state.
Runway describes three steps that make generation fast enough to feel interactive. It trained the model to generate frames autoregressively, distilled the many-step denoising process down to a few steps, then trained the fast model on its own outputs so quality holds over longer sessions.
You start Solaris from a state such as a product scene or brand environment, and it streams frames as you click, drag, or type. There are no predefined screens or templates; text prompts define what each interaction means inside a given scene. Runway notes that interactions stop feeling interactive beyond roughly half a second of delay, so the model generates each frame from only what came before to keep up.
Runway's user study preferred Solaris over a coded baseline
Runway ran a comparison against a coded interface generated by Claude Opus 5. Both systems started from the same image and received the same interaction requests across 30 examples, and 250 participants supplied nearly 7,500 pairwise judgments.
By Runway's account, participants preferred Solaris in 61 percent of comparisons for following the requested interaction, against 24 percent for the coded result, with 13 percent rated equivalent. For natural behavior within the scene, Solaris was preferred in 71 percent of comparisons against 21 percent, with 6 percent equivalent. Runway attributes the gap to the model already understanding how objects and materials behave, rather than treating each interaction as an isolated update. All methodology and results come from Runway.
What Solaris can't do yet, and where it could reach creative work
Runway says Solaris is strongest at ambient motion, click-and-drag interactions, and scene transitions, and it lists the parts that are not solved. Stable, legible text is one of the hardest problems in video generation and central to interfaces; trust and grounding matter because a convincing wrong answer is worse than none; long-session coherence, accessibility, and integration with the existing software stack all remain active research. Runway is clear that this is an early research system, not a replacement for production software, and points to demos of the behavior.
Analysis: for creative teams, the near-term value is in what generated surfaces could do for tool interfaces, visual prototyping, adaptive tutorials, and interactive brand or product experiences. An interface that renders directly from reference imagery could let a team try a layout or a product configuration without building it first. Those uses stay speculative until the text, grounding, and session-length limits close.
