Spatial AI Research
Space-LLM 1.1
Generate structured spaces
August 8, 2024
Space LLM 1.1 is a transformer model, trained to design floor plans.
This model is able receive boundaries and entrances as input, and outputs layouts of indoor spaces as bedrooms, bathrooms, kitchens, etc.
Beyond Parametric Design
Parametric design algorithms and models have been used in Architecture, Engineering, and Construction (ACE) software for a few years now.
Some of these parametric programs are able to create incredibly great results. Rhino Grasshopper is a great example of a tool that changed the industry with innovative designs.
However, parametric tools rely either on rigid input parameters or complicated visual programming interfaces.
With parametric designs you can’t just ask “Why did you make that decision for the design?”. You can, of course, debug the parametric model, but then you would either need proficiency in programming or a lot of patience to look to the “spaghetti” interface behind it...
Space-llm 1.1 is now accessible from the Vitrus Platform
Source - Example of Visual Programming Interface (a.k.a. “spaghetti” interface)
Using Descriptive Design
Using Large Language Models to design spaces
Beyond simplicity of usage and capability of asking the model to clarify it’s decisions, descriptive is more generalizable.
Vitrus’ exploration with LLMs and design began in March 2024, with the first models (“fast-0.1” and “architect-0.5”) being able to create very crude examples of designs. These were based on both closed and open-source foundational models, with extensive prompt engineering and other soft controls to achieve these results.
Since then, the foundation models dropped 25x in cost and have significantly improved in efficiency and size (a proxy for “reasoning” capabilities).
Improvements over version 1.0
-
The new model is much better at respecting the boundaries of the property (red lined shape on demos).
-
This model accepts entrance points (doors), and will design with them as a factor.
-
With a new method for output validation, we got a 100% accuracy on model output structure.
The first version used to error often as the output sometimes had the wrong data typing.
Video with detailed demo of version 1.0
Using Vitrus on your
your Developments
You can run Vitrus design engine on your property! Vitrus AI is able to create designs based on the zoning and typologies to build. All this in seconds, no CAD needed!
Our Enterprise platform will help you analyze the building potential of a property with a clear preliminary design, that references the building codes and site context.
