A floor plan is a spatial object — rooms have positions, dimensions, adjacencies, door placements. A language model processes sequences of tokens. These seem incompatible. But a room is also a discrete entity with attributes: type (bedroom), area (12 m²), position (northwest corner), connections (adjacent to hallway, shares wall with bathroom). Write these as tokens and the floor plan becomes a sentence.
HouseMind (arXiv:2603.11640) introduces discrete room-instance tokens as a unified vocabulary that bridges layouts and symbolic reasoning. Each room instance becomes a token in the model's vocabulary. The floor plan becomes a sequence the model can generate through next-token prediction, understand through instruction following, and edit through targeted token modification.
Note (2026-11-07): the abstract of arXiv:2603.11640 presents HouseMind, a multimodal language model with discrete room-instance tokens that unifies floor plan understanding, generation and editing, and says experiments show superior geometric validity and controllability while remaining efficient and locally deployable. It does not describe the comparison with diffusion models regenerating whole images, how edits are performed by reordering tokens while preserving wall alignments and door positions, the number of room instances per plan, or the claim that the tokenization is lossless; those details and the general lesson about tokenization are my additions. I compared this essay with the abstract only.
The editing capability is what separates this from image-based approaches. Tell a diffusion model to “move the kitchen next to the dining room” and it regenerates the entire image, possibly breaking geometric constraints. Tell a token-based model to do the same and it reorders tokens while preserving the spatial grammar — wall alignments, door positions, room proportions. The edit is local because the representation is local.
The structural insight: the right tokenization scheme is a compression of the domain. A floor plan image has millions of pixels, mostly redundant. The token representation has dozens of room instances, each carrying exactly the information an architect cares about. The compression isn't lossy — it's a projection from pixel space to design space. And design space is where the decisions live.