How depth-AI turns a flat photo into cinematic motion

A monocular depth model estimates how far every pixel in a photo sat from the camera. Moving a virtual camera through that depth map makes near objects travel further across the frame than far ones — real parallax. The result reads as a drone push-in or arc, from one still and with no GPU.

A single photo is flat, but the scene it captured wasn't. Monocular depth estimation is a class of AI model that predicts how far each pixel sat from the camera — near foreground, mid-room, far window — turning one image into a layered 2.5D model of the space.

The same technique gives exterior photos a drone-like move; see drone-style video from photos for what it can and can't replace.

With that depth map in hand, the engine can move the 'camera' through the scene. Near pixels are pushed and panned more than far ones, exactly as parallax works in the real world: hold up a finger, move your head, and it shifts more than the wall behind it. The result is a push-in or drift the eye reads as a genuine camera move, not a flat zoom.

The craft is in the easing and the variety. Linear motion looks robotic, so each move uses smooth acceleration curves and a rotating set of presets — a slow dolly-in for a living room, a gentle rise for an exterior, a calm push for a bedroom — so a six-photo tour doesn't feel like the same effect applied six times. Edge handling matters too: without care, the borders of a pushed photo smear or tear, so the engine insets and replicates edges to keep them clean.

Why does this approach matter commercially? Because it runs on ordinary hardware. Depth estimation is a one-time inference per photo, so the marginal cost of turning a still into motion is cents, not a videographer's day rate — which is what makes cinematic video viable for every listing instead of just the flagship ones.

It's the engine behind every PropReel render: drop in a flat photo and get a film-grade move out the other side — no drone, no gimbal, no editor — and the first video is free.

Why straight lines are the hard part

Almost every failure mode in 2.5D motion shows up on a straight line. Real estate photography is nothing but straight lines — door frames, counter edges, skirting, window mullions — and a human eye that has spent its life in rectangular rooms notices a millimetre of bend instantly. A depth map that is slightly wrong in a smooth outdoor scene is invisible; the same error across a kitchen counter is the thing that makes a viewer say the video looks fake without being able to say why.

The practical fix is not a better model, it is a smaller move. PropReel measures the maximum bend a proposed camera path would introduce and falls back to a pure affine move when it crosses a threshold — a slightly less dramatic shot that is structurally perfect, instead of a dramatic one that warps a doorway.

What depth motion cannot do

It cannot reveal what the photo did not capture. Push in far enough and you reach the edge of the information the camera recorded; past that, any model is inventing, and invented architecture is exactly the artefact that reads as AI. It also cannot fix a bad photo: a dark, crooked or cluttered still animated beautifully is still a dark, crooked, cluttered still. Motion is a multiplier on the source, not a substitute for it.