limits
What this is actually doing.
Written before anyone asked.
The depth is inferred, not measured
No one recorded the distance to Vermeer’s window. A model called Depth Anything V2 Small looks at the flat image and guesses the geometry. It is usually convincing and sometimes wrong. It has no idea what any of these objects are, only what pictures of near and far things tend to look like.
The gallery was inferred in advance
Those eleven depth maps were computed once, on a laptop, and shipped as 8KB images. Nothing runs a model when you arrive. That is why the hall works instantly on any device, with no download and no WebGPU.
Only your own picture runs live
That path loads about 21MB once: a 16MB model and a 5MB runtime, both served from this site, never from Hugging Face or a CDN. It cannot start without your click, and it needs WebGPU: the same model takes roughly 25 seconds on a CPU against 1.5 on a GPU, and a 25-second wait is not worth offering.
Your picture is read into the page and never uploaded. There is no server here to upload it to.
The room behind each picture was invented
Push the relief and every raised thing pulls away from whatever is behind it. Nothing was ever recorded back there. A photograph holds no data about the wall under the pitcher or the sea under the boat.
So a background layer is built for each piece before you arrive. Everything within a few pixels of a depth cliff is erased, and the hole is closed by growing the far side inward. That layer is hung a little further back, and what opens up when you move is it.
It is a reconstruction. Nothing here knows what is actually behind her arm. It knows what the wall either side of it was doing and carries that across, blurred, so it reads as the room continuing rather than as detail somebody recorded. A picture you bring has no such layer, and its cliffs are drawn as shaded sides instead.
The pictures are genuinely free
All eleven are CC0 from The Metropolitan Museum of Art’s Open Access collection: public domain, bundled into this site, credited and linked. The model is Apache 2.0.
Nothing is kept
No backend, no database, no analytics, no third-party script of any kind. A picture you bring lives in this tab and dies with it.