Walk and It Appears: Real-Time Unbounded Open-World Scene Generation from a Single Image or Prompt

Turning a prompt or visual reference into a world one can traverse requires new geometry to connect to previous construction and arrive before the user reaches it. We present Walk and It Appears, a system that authors a persistent world program, constructs local geometry through canonical interfaces, and streams completed records under stable identities. These interfaces carry terrain elevations and road crossings between chunks, parcel frames into buildings, and entrances into interiors. The reported evaluation covers 233 runs across 14 task–seed configurations and six language-model backbones. Equal-backbone mean scores favor our system over direct expandable-code generation in physical validity (0.80 versus 0.31), route success (0.80 versus 0.37), and online chunk availability (0.95 versus 0.62) under the reported runtime configurations. A 1.5 km outward-and-return diagnostic probes expansion and revisits; image-conditioned and indoor cases are curated demonstrations. The evaluation lacks per-task score breakdowns and component ablations, so cross-task variability and individual module effects remain unresolved.