Runway's Solaris Renders UIs Frame by Frame, But Is Still a Demo With No Pricing or API
shashib · x · 2026-09-05
Runway introduced Solaris on August 31, a new release it places in a category it calls Interface World Models, about nine months after its GWM-1 world model.
- How it works: Built on the Gen-4.5 video model, Solaris extends GWM-1's approach (which modeled 3D environments and avatar training data) to software. It takes clicks, drags, and typing as inputs and predicts the screen frame by frame — no DOM, no CSS, no event handlers; a button is just pixels, and the model paints what happens next.
- Claimed performance: Runway says Solaris beats frontier LLMs on two metrics — structural similarity (coherent layouts) and information retention (interfaces remembering state across interactions), the two hardest problems in generative UIs.
- Caveats: Runway pitches it as the end of the design-to-code pipeline, but for now it's a research demo: 720p target resolution, no published pricing, and no API — far from a usable frontend replacement.
Related event: Runway Unveils Solaris, an Interface World Model(2 posts)→
More from Multimodal
- One-shotting a 3D scene with Astra in 20 minutes — dr_After · 2026-09-05
- imagebench: 50 prompts to benchmark open-source image models on DGX Spark — Aware_Web9715 · 2026-09-05
- A found-footage experiment with a custom FLUX-2 LoRA, Python and post-fx — uisato · 2026-09-05
- Dev builds 2D-video-to-3D animation translation layer in an hour to test Astra's spatial reasoning — Relevant-Magic-Card · 2026-09-05
- Invideo Launches AI-Powered Video Editor to Automate Footage Organization and First Cuts — azed_ai · 2026-09-05
- Full House Reimagined as an MCU Trailer: AI Video Test Goes Viral — Sad_Coach_1433 · 2026-09-05