Skip to content

how it works

How it works

No download. No dedicated GPU — it runs on the integrated graphics you already have, in a WebGPU browser (Chrome 113+, Edge 113+, Firefox 141+, or Safari 18+). About eight seconds the first time. Two seconds every time after.

The goal: a first-time viewer shouldn't be able to tell, at a glance, whether they're looking at PAXAL or at a screenshot from a current-generation console title. This page is the demo. The platform behind it hosts more.

THREE WORKERS, ONE FRAME

The main thread renders. Everything else — fetching, decoding, persisting — lives in its own worker so the canvas never waits.

BROWSERtab opens, HTML paints instantlyMANIFESTchunk graph fetched from edgePERSIST WORKEROPFS read/write, warm cache 8s→2sSTREAMING HOSTpriority queue, backpressureDECODE WORKERgaussian-splat decompress, off-mainPLAYCANVASWebGPU raster, motion-aware LODGPUintegrated or discrete — same path

WHAT MAKES IT FAST

warm cache

OPFS persists every decoded chunk. Second visit boots in under two seconds because the GPU buffer is the only thing that hasn't been built yet.

motion-aware LOD

Driving fast? Lowest-detail splats. Parked? Highest. The controller swaps in under 200 ms when frame budget tightens — never as a jolt, always between motion frames.

three-worker pipeline

Manifest, decode and persist all run off-thread. The main thread is free for the canvas; the fetch path is free for whatever order the camera asks for next.

PATCHING THE PHOTOREAL

Patching makes the photoreal interactive — collision, physics, and behavior layered onto the capture. The full brief is at /platform/patching.

IN MOTION

first visit ≈ 8s to first pixel · return ≈ 2s — same capture, served warm from the OPFS cache.

The captured spaces you can drive today are catalogued at cities.

Words don't do this justice. Open it in a tab and drive.

▶ Try it