Tentang
Runway's August 2026 research announcement presents an Interface World Model that generates interface frames in response to clicks, drags, and typing. A language model guides scene behavior while a distilled video model renders the interaction, exploring runtime GenUI beyond generated code or component catalogs.
Ringkasan
1. Action-conditioned rendering: Past user inputs condition subsequent frames, with natural-language prompts defining interaction behavior
2. Real-time architecture: Combines autoregressive frame generation, few-step distillation, and training on the model's own outputs
3. Reported evaluation: Runway compares screenshot reconstruction and collects pairwise interaction judgments from 250 participants across 30 examples; these are vendor-reported studies, not evidence of general task reliability
4. Research limits: Stable text, verified grounding, long-session coherence, accessibility, and software integration remain unresolved; the page invites early-access requests