The model generates the interface autoregressively, one frame at a time, where each frame depends only on what came before. User input (clicks, drags) is treated as conditioning for the next frame, the same way it treats text or images. In Solaris a hybrid approach is used: a language model (LLM) interprets user requests and decides how the interface should evolve, while the world model handles the visual generation. Rendering must be cheap enough to keep up with the user — interactions stop feeling interactive at around half a second of delay.
Traditional interfaces require converting a visual design into intermediate code and treat interactions as isolated state updates. An IWM removes the code layer and renders the interface directly, preserving scene coherence because the model already understands how objects, materials and environments behave.
Generates the interface layer frame by frame; each frame depends only on the preceding frames.
An LLM interprets user requests and determines how the interface evolves.
Clicks, drags and other interactions are used as a conditioning signal for the next frame.
Stable, legible text remains one of the hardest problems in video generation.
Generated answers risk being convincing yet incorrect.
Coherence degrades during extended, open-ended interactions.
Integration with screen readers and accessibility APIs remains unresolved.
On 31 August 2026 Runway announces Solaris as the first model in the new Interface World Models family.
Solaris targets visual quality that holds at 720p.
Interactions stop feeling interactive around half a second of delay.
Each frame is generated conditionally on prior frames and on user input.
Generation is autoregressive — each frame depends on previous ones, forcing sequential inference.
