← Krenels Lab

From a listing photo to an interactive 3D building

Photo-to-3D generation for a property catalog — and where it stands.

Spin one of the finished 3D mansions in the browser and, at a certain angle, a fir tree grows out of the wall. In the input photo a tree stood in front of the facade, and the generator never told it apart from the building — it fused the two into one piece of geometry. That is the shape of the whole experiment: the model captures the side it can see and fills in the rest by guessing.

We were checking whether photo-to-3D could go on a production line — running a whole property catalog through it so every listing gets an interactive 3D view, without sending out a photographer with a scanner. We took three mansions from a Moscow brokerage's catalog and built each one four ways, from a single flat photo to a hand-finished model. Every mansion and method is in the lab below — spin the models and read the scores.

Three mansions, four build methods each, live. Spin a model and read the scores. Open full screen →

Not production-ready yet

For a production line the technology is not there. The results are too inconsistent from lot to lot and from frame to frame: one house comes out clean, the next welds a tree or a neighboring building onto the wall, and someone has to check every model by hand. We use 3D case by case, for a specific lot that already has warm interest behind it, rather than across the whole catalog. The basic hygiene of the format — which inputs give what, how to put a 3D viewer on the web — we went through and kept.

What's next: preprocessing the input

The result rides on the input photo more than on the generator. The model reads light as shape: a shadow on a wall becomes a dent, glare on the glass becomes a bump, so a realistic photo full of sun and reflections feeds it false geometry.

The direction we want to try next is preprocessing the photo before it ever reaches the generator: turning a realistic shot into a flat, toy-like render with simple, unshaded textures and no shadows. With the light stripped out, the input holds shape alone, and there is nothing left for the generator to mistake for relief.

Start a conversation