Aspect

A photograph records exactly one viewpoint, and so does a painting. You have looked at these pictures your whole life from the only angle that was ever captured, because there was no other one. There is now. It is inferred rather than recovered: a model called Depth Anything looks at the flat image and guesses the geometry the painter never recorded. Vermeer's room stands up off its plane and you can lean around the pitcher, the Great Wave gets a trough behind it, Madame X steps off the wall. The difference between inferred and recorded is the whole argument, and the site never blurs it.

Depth is inferred once at build time and ships as 8KB greyscale maps, so the hall opens instantly on any device with no model, no download and no WebGPU. One WebGL2 canvas draws a grid displaced by the depth texture, and each painting belongs to its own section, positioned from that section's rectangle every frame, so there is no second clock to lag against. The room behind each picture is reconstructed before you arrive, because pushing the relief pulls every raised thing away from a wall nobody ever recorded. Dragging turns the picture within bounds that stop the camera going under the floor. Only a picture you bring yourself runs the model live, after an explicit click.

Live, and exact about which parts are guesses. The limits page states what was inferred, when, and what was invented behind each canvas. It also took two rebuilds and a deletion to get here: the cross-fade between pictures was removed entirely after four rounds of failing to time it, and every animation turned out to be advancing per frame rather than per second, which had quietly made the whole thing a hardware speed test and hidden three earlier rounds of timing work on exactly the machines that needed them most.