Researchers turn floor plan generation into a next-token problem

Generative AI can paint a stunning fantasy castle, but ask it to draw a functional two-bedroom apartment, and it usually hallucinates doors into solid walls or outputs blurry pixels that no CAD software can parse.

Researchers Kaede Shiohara and Toshihiko Yamasaki published a paper set for CVPR 2026 titled "Unified Vector Floorplan Generation via Markup Representation." Instead of treating floor plans as images, they took a completely different route: they decided to treat houses like web pages.

HTML for architecture

The secret sauce is something they call Floorplan Markup Language, or FML. Just like HTML describes text, buttons, and layouts for a browser, FML encodes rooms, walls, doors, and geometric coordinates into clean, sequential markup tags.

Once a blueprint is structured this way, the messy spatial generation problem suddenly becomes what AI does best: next-token prediction.

The researchers built a transformer model called FMLM. Feed it conditions, and it simply generates the markup tokens one by one. Because the output is native markup rather than raw pixels, it natively produces clean vector lines and easily adapts to variable numbers of rooms and doors.

One model for every constraint

Traditionally, floor plan tools have been rigid. Old rule-based systems made structurally sound houses that looked completely boring. Recent diffusion or GAN models offered variety, but they required separate, bespoke setups depending on the prompt—whether you started with site boundaries, adjacency graphs, or partial sketches.

FML unifies both the input constraints and the final layout into the exact same grammar. In benchmarks on the standard RPLAN dataset, FMLM beat previous task-specific state-of-the-art models across the board—using just a single generalist model.

For now, the project is purely academic. The team has shared the CVPR paper and a project page, but no GitHub repo or live playground has been released yet.

Why it matters

For architects and interior designers, native vector output is the holy grail. An image generator only gives you a mood board, forcing a human draftsman to manually redraw every wall in CAD. A model that spits out structured markup can plug directly into professional design software.

For the broader industry, it is a brilliant showcase of paradigm shifting. Instead of over-engineering specialized graph neural networks or complex geometric solvers, the researchers simply remapped architectural topology into text. If you can express a house as a sentence, you get to ride the coattails of all the massive scaling improvements happening in transformer architectures.

We think this markup approach is remarkably clever, though its limits will be tested on complex, multi-story commercial spaces where tag syntax might explode in complexity. Still, once open code lands, this could easily become the default blueprint for CAD-native generative design.