A floor plan is a constraint problem, not an image

Most AI floor plan tools generate a picture of a plan and then trace it back into geometry. That works until you look closely — and the place you look closely is the CAD file, which is the one a builder or drafter actually opens.

What a plan actually has to satisfy

• Room areas have to sum to the target.
• Rooms have to tile the footprint — no gaps, no overlaps.
• Walls have to close.
• Every room has to be reachable without crossing a bedroom.
• The kitchen has to actually touch the dining room.
• A door has to fit in the wall it is drawn on.

A generative model satisfies these statistically. It gets them right most of the time and fails quietly the rest, and the failure lands downstream in the export.

What solving changes

PlanForge partitions the footprint recursively and allocates each cut in proportion to the room areas you asked for. Area error is not optimised away — it is structurally impossible, and what is left is dimension rounding. Because the geometry is exact integers in millimetres, the DXF is a serialisation rather than a guess.

It tells you what is wrong with it

The objective function is written down, so every plan comes with a critique: "Pantry is not off the kitchen −2.5", "Bath 2 is under minimum width". A trained image model cannot produce that screen, because it has no articulable reason for anything it draws.

What to do before you build

Take the concept to a licensed architect or engineer. These plans are for exploring and communicating a direction — code compliance, structure, services, site conditions and permitting all still need professional judgment. In most jurisdictions only a licensed practitioner may submit plans for approval at all.