Real depth, extracted on your device: a control input, not a guess.

A depth pass (Depth-Anything / MiDaS-class) runs in the browser at $0 cost to us (§V) and hands the whole system a real spatial map: parallax, depth-of-field, and foreground / mid / background blocking that doubles as a generation control. The point that matters: a measured depth map beats an LLM's idea of one, same honesty line the V/D/C receipts draw, pointed at pixels.

1 · get a depth map

Drop a photo extractor wires next The on-device Depth-Anything pass is a gen-pipeline slot (engine lane). Until it lands, preview a built-in sample scene below: no fake “extracting…”, ever (§XXXVII).

or preview a built-in scene

Each scene has a different depth structure. Watch the parallax spread, where focus lands, and what each band isolates change as you switch.

sample scene depth map

depth nearfar

3 · blocking as gen control tap a band to isolate

what real depth unlocks

Honest parallax on the §LVI glassOur layered glass fakes its depth today. A real map makes cursor / scroll parallax automatic from any photo.

Depth-of-field & relief lightFocus a plane, fall light across real geometry: the lit-not-emissive principle finally has something to light.

Depth-map-as-storyboardBlock a scene by plane, and that blocking IS both the storyboard and the gen control (the ControlNet path).

The bridge to “Your Captures” / splatsPhoto-to-depth-to-relief is the honest first step toward a 3D-feeling scene, shippable long before full 4DGS.

what's real here

Live now: the parallax, depth-of-field, depth-map view, and band-blocking: all running on three self-contained sample scenes (flat silhouettes at declared depths: landscape, portrait, still life), so this demo is real and needs no model. Wires next: the on-device photo-to-extract step (Depth-Anything ONNX, a gen-pipeline slot on the engine lane) and the ControlNet hand-off. Nothing fabricates an extraction it didn't run (§XXXVII).