Runway highlights three characteristics of the system: visual generation, continuous rendering, and open-ended interaction. A generated retail environment, for example, could let a user drag clothing onto an image of themselves, reposition products, or alter the scene directly. In another example, users could move furniture or change colors through natural-language instructions while the environment updates continuously. That approach also changes how interaction is defined. Solaris can assign new behaviors to objects through prompts rather than hard-coded controls. Clicking an object can alter what subsequent clicks or drags do, giving the same scene different interaction patterns without rebuilding the application.
Solaris is based on Runway’s Gen-4.5 video model and follows the company’s earlier work on GWM-1, its general world model. Runway adapted the underlying system to interpret user actions and generate frames quickly enough for interactive use.
The model treats clicks, drags, and other inputs as signals that influence the next generated frame. Because each frame depends on prior frames and previous interactions, Solaris learns how a scene should change in response to user behavior without requiring every reaction to be manually specified.
Runway says it converted the model into a real-time system through several changes to its video-generation process. Solaris generates frames sequentially, uses a reduced number of denoising steps, and is trained on its own outputs to help maintain image quality during longer interactions.
A language model works alongside the visual model. The LLM interprets what the user wants, decides whether the current scene should change or transition, and produces instructions that guide Solaris. The world model then handles the visual rendering and immediate response. Runway argues that this separation allows one system to determine what should happen while another determines how it should appear. Users provide an initial scene, and Solaris continues generating from that state as new actions arrive. Rather than selecting from preset screens or templates, the interface evolves from the generated sequence.
The company also sees Solaris as a potential environment for training AI agents. Runway argues that agents trained primarily on conventional software can become dependent on familiar layouts, making them less adaptable when interface structures change. Solaris can instead generate constantly changing interfaces, giving agents a broader range of visual environments in which to practice computer-use tasks.
Runway tested its approach against multimodal language models by measuring how accurately they could reconstruct interfaces from screenshots. The evaluation covered 30 examples ranging from simple webpages to visually complex scenes and used structural similarity and DINOv3-based comparisons to measure how much visual information survived reconstruction.
According to Runway, reconstruction quality declined as visual complexity increased. The company argues that Solaris avoids some of that loss because it does not first convert an interface into a language or code representation before reproducing it.
Runway also compared Solaris with a coded interface generated using Claude Opus 5. Both systems received the same starting image and interaction requests. In a study involving 250 participants and nearly 7,500 pairwise judgments, users preferred Solaris for following the requested interaction in 61% of comparisons, compared with 24% for the coded result. Thirteen percent were judged equivalent. The difference was larger when participants rated which result behaved more naturally within the scene. Solaris was preferred in 71% of comparisons, while the coded interface was chosen in 21% and 6% were rated as equivalent.
Runway says Solaris currently performs best with ambient motion, click-and-drag interactions, and transitions between scenes. Several limitations remain, including maintaining readable text, preserving consistency over long sessions, grounding outputs in reliable information, and integrating generated interfaces with accessibility tools and the broader software stack.
Text remains a particular challenge because real-time video generation can struggle to keep lettering stable across frames. Runway says one possible approach is to combine image models for text-heavy views with video models for continuous interaction rather than relying entirely on one generation method. The company is also still working on longer-running sessions, richer grounding from documents and product data, and compatibility with systems such as screen readers and accessibility APIs.
Runway ultimately envisions Solaris as an interface layer in which software does not have to be organized around fixed applications or predefined page structures. Instead, an interface could be generated for a specific task and adapt as the user works through it.
Solaris is Runway’s first model in its Interface World Model family. The company is working with partners ahead of a public launch and has opened requests for early access.
This analysis is based on reporting from Runway.
Image courtesy of Runway.
This article was generated with AI assistance and reviewed for accuracy and quality.
