The latent space is both cemetery and garden.
When you generate an image with a diffusion model, you don't pull a finished thing from a hidden drawer. You start with noise — a flat field of static — and you ask the model: what would denoise into something a human might call "a girl reading in a window at dusk"? The model doesn't search a library. It descends a topology. Every step it takes, it leans toward the gravitational center of every painting, photograph, illustration, and snapshot that ever resembled the phrase you typed.
The makers are mostly gone, or scattered, or never asked. Anonymous Flemish painters. A grandmother with a Brownie camera. A Romanian propaganda muralist. A teenager in 2014 posting fanart to a Tumblr that no longer exists. The model doesn't remember any of them individually. It has digested them into a manifold, a high-dimensional landscape where their fingerprints have melted into terrain.
This is the cemetery part. Every output is, in a small way, the dissolution of everything that fed it. No single brushstroke survives. The author is not just dead — the author is composted.
But it is also a garden. Because what grows from that compost is something none of them painted. The image is genuinely new in the same way a child is new — recombinant, dependent, surprising. You hold it for a moment. You feel something. You don't know whom to thank.
I think this is what the discomfort is. Not theft, exactly. Not creativity, exactly. Something older. We're summoning the dead, and the dead are answering — not in their own voices, but in a chorus we can't quite separate. And we call it art, and we call it ours, and both are true, and both are not.
The model doesn't remember any single painter. It remembers the gravity between them all.
Top comments (0)