How We Reconstruct Vanished Places with AI (and Where We Refuse to Guess)

Reconstructing a place that no longer exists is an old craft. Painters, model-makers, and historians have done it for centuries. What is new is the speed and the medium: timemachina renders a full 360 panorama of a historical moment, something you can look around inside, in minutes rather than months. That power comes with an obligation to be honest about what is evidence and what is inference. Here is how the pipeline works, and where we stop.

From a research note to a place

Every scene begins the same way, with a real place and a real year. Before any image exists, we gather a short research note: what stood there, what it was made of, how people used the space, what the light and weather were like. That note is grounded in a cited source, usually an authoritative encyclopaedia entry or scholarship, and it is what separates a reconstruction from a fantasy.

The note becomes a structured prompt, which is rendered by an image model into a equirectangular panorama at 2048 by 1024 pixels. That 2:1 ratio matters: it is the standard format that maps cleanly onto a sphere, so the viewer can wrap the image around you without distortion. You are not looking at a flat picture. You are standing in the middle of one.

What the model follows, and what it fills in

Where the historical record is strong, we follow it closely. The shape of a known building, the layout of a forum, the colour of a tiled dome: these are not guesses, and the prompt holds the model to them. The harder cases are the everyday details that history rarely records. Exactly which market stall stood on which corner, the precise faces in a crowd, the wear on a particular doorstep. No source preserves that.

For those gaps we make conservative, period-appropriate choices, the same way a museum diorama does. A crowd in second-century Rome wears clothing consistent with the period and the season, but we do not claim a specific person is depicted. This is the difference between reconstruction and forgery: a reconstruction is transparent that some details are representative rather than documented.

The line we refuse to cross

There are things we will not generate, because they would dress a guess up as a fact.

We do not invent specific named individuals and present them as portraits. We do not render readable inscriptions, signatures, or documents that imply a primary source we do not have. We do not depict events that did not happen at a place, however dramatic they would look. And when the evidence for a location is too thin to render responsibly, the honest answer is to not publish it rather than to fill the void with invention.

There is also a safety layer. Each prompt passes a deny-list check and text moderation before generation, and the rendered image passes a vision moderation pass afterward. Scenes that fail are rejected, not quietly fixed.

Why a 360 view keeps us honest

A flat image lets you hide what you do not know outside the frame. A panorama does not. When you can turn all the way around, every gap in the research is on screen somewhere, which is a useful discipline. It forces the reconstruction to be complete and consistent in every direction, and it makes the limits of the evidence visible to you, not just to us.

That is the bargain timemachina offers. The scenes are vivid and immersive, but they are built on cited research and honest about their own uncertainty. You can read more about the principles on the about page, or request a moment and watch the pipeline run on a place you choose. The aim is never a perfect illusion. It is the closest honest approximation of a view that no camera was ever there to take.