Nice work on residential layout generation — how do you handle room adjacency constraints?
Really cool project — the idea of using a GPT-J variant to generate architectural layouts from natural language prompts is a great approach, and the geometric representation you're using seems flexible enough for downstream CAD/rendering pipelines.
Curious how you're handling room adjacency constraints (e.g. kitchen near dining, bathroom not adjacent to bedroom entrance) during generation — is that baked into the training data distribution, or is there a post-processing/validation step?
For context, we've been working on a similar problem from a different angle — going from a floor plan (sketch or rough layout) to a fully rendered, style-consistent visualization, plus some auto-layout tooling. If anyone here is exploring the generation → rendering pipeline, might be a useful reference point: AI Floor Plan Generator
Would love to hear more about how the model performs on non-rectangular boundaries or multi-story constraints — any plans to extend beyond single-floor apartments?
Hello! Thanks for the kind words, also didn't expect people around here after so many years :)
Indeed,adjacencies were learned from data distribution. That's powerful but also as you can imagine limiting given data was not that diverse necessarily. That said, I found models (even the ones used here from almost 4 years ago) could generalise quite well (especially when they were producing outputs not fitting to user's prompt!).
And because models learned to create valid designs (check paper, 99% validity for most models), it made them interesting data generation engines.
I can imagine models today can (1) do a lot of this in context and (2) could be finetuned to incorporate a lot more detail and nuance. I haven't had time to work on this, but still a cool area of research.