AI floor plans that remain editable

A plan can look convincing in a thumbnail and fail when we follow a route through it. Doors may connect the wrong spaces; furniture may not fit. The first question is what the tool represents: pixels, room relationships, boundaries or building elements.

Two research approaches

Graph2Plan (2020) combines a building boundary with a layout graph. Users can influence room counts and connections; the method predicts a raster plan and refines room boxes toward a vector layout.

HouseDiffusion (CVPR 2023) generates room and door polygon coordinates conditioned on a graph. Both approaches make relationships part of generation. We cite them for that design lesson, without treating them as a ranking of current systems.

A conversational brief can make instructions easier to express. An editable representation is still needed for a dimensional change such as widening a corridor.

Compare arrangements under the same brief

For a small apartment trial, we would fix the boundary, entrance and service zone, then ask for two arrangements with the living room near the main frontage. Record which adjacencies are required and which are preferences.

1. Topology: inspect room connections and routes through private spaces.
2. Geometry: check closed boundaries, overlaps, gaps and dimensions.
3. Use: review furniture, privacy, daylight intentions and arrival.

A graph edge or drawn door helps explain a connection; project requirements still need architectural review.

A repeatable study for Fraktal

We would use plans we have permission to share and include irregular boundaries, protected structure and missing information. Compare each method with an unassisted starting layout, recording failures and the reconstruction needed before editing.

Our proposed experiment asks what happens after the first attractive image: can a designer widen a corridor or change a room connection through an understandable revision, without redrawing the plan from scratch?