A private island, a listing video built from photos

Overview

A private island off Madeira Park on the Sunshine Coast. Five bedrooms, three bathrooms, thirteen guests, a house built around a standing tree, and a dock that is the only way in.

The listing has good photographs. What it does not have, and cannot have, is the thing that actually sells this property: nobody drives here. The arrival is the product, and a still photograph cannot show an arrival.

So the cut opens on open water, brings you in to the dock, walks the house from the staircase to the bedrooms, comes back out to the patio at sunset, and ends on the island shrinking below a camera pulling back at dusk. Eighteen shots, ninety five seconds, every frame of it grown from a photograph that was already on the listing.

The problem

A private island is the hardest kind of property to photograph and the easiest to sell on motion. Every still is a postcard, and postcards all look the same after four of them. What a guest is buying is the boat ride, the moment the island resolves out of the water, and the fact that at some point in the evening the only thing moving out there is you.

Filming that properly means a crew on a boat, a drone operator, and a weather window. On a property reachable only by water, the logistics are most of the invoice before anyone has framed a shot.

What I built

A shot list and a render script. The list is a tab separated file: source photo, duration, camera move. The script walks it and produces one clip per line.

photo               seconds   camera move
01-dock-approach.jpg   6      slow dolly forward across the water toward the dock
08-living-sunset.jpg   5      slow push in toward the windows at sunset
17-orca.jpg            6      slow lateral drift right along the dock as the orca passes

The part worth explaining is that not every shot is generated. Four are: the water approach, the island aerial, the orca, and the pull back at dusk. Those have genuine depth for a model to resolve, so a generative camera move produces real parallax and is worth what it costs.

The other fourteen are camera moves cut straight out of the photograph. The source stills are 2560 pixels wide and the frame is 1920, which leaves 640 pixels for a window to travel across at native resolution. Nothing is upscaled and nothing is invented: a push in toward the dining table is a real crop of real pixels, eased in and out so it starts and stops like a camera head rather than a slider.

That split is the whole method. Generative motion earns its price on water, weather and distance. On a flat, evenly lit room it produces a slow zoom, which is a thing ffmpeg will do for free and do more predictably.

What it cannot do

The output is bounded by the input photograph, and the bound is a hard one.

A flat photograph can be moved across or moved into. It cannot be moved through. Where a room was shot square to the wall with even light, the camera has nothing to pass behind, so the shot reads as a considered push rather than a walk. The exteriors here animate beautifully because there is a driveway leading in, trees in the foreground and water with a horizon behind it. The interiors are honest, and they are not the same thing.

The generated shots have their own ceiling. Push a clip past about eight seconds, or ask for two camera moves in one shot, and doorframes start to bend. Every prompt carries the same appended guard for that reason: rigid architecture, no morphing of walls or furniture, no people, no camera shake.

Neither limit is fixed by a better prompt. Both are fixed by a better source photograph.

Where it’s at

A spec build. Nobody briefed me and nobody has paid for it. I made it because I wanted to know whether the method held up on the hardest property I could find, one you cannot drive to and cannot photograph your way out of.

It holds up. The arrival works, the sunset through the living room windows works, and the orca is the shot that makes the whole thing worth watching twice.