Architecture Trends Tech Trends

AI Architecture Generative Design Housing

AI Architecture Generative Design Housing

Will architects one day be replaced by computers? That's the idea being put forward by Stanislas Chaillou. But, is AI architecture really practical. If so, how far are we away from achieving fully computer-designed buildings? Should architects start retraining?


The AI designed architecture is created using Generative Adversarial Nets (GANs)...
The layouts are generated in 3 stages; the footprint, the layout and the furniture. A human need to manually input details such as the door and the main windows. They also have to ability to tweak the layout at different stages of the design process. The computer generates the designs by calculating the most probable layout based on a series of existing layouts.

Generative Adversarial Nets, is there anything they can't do?

Chaillou's concept is to design apartment block layouts using Generative Adversarial Nets (GANs).

Currently, in the CAD design world, one can not go far without hearing about the seemingly limitless possibilities of GAN. In the past we have explored GAN generated photographs and even, computer designed 3-dimensional models. There have even been generatively designed automobiles parts. In BricsCAD you're currently able to use smart tools such as BIMify to automatically detect and label areas, so why shouldn't you be able to automatically generate whole layouts?

The design process is broken down into 3 stages:

  1. The footprint
  2. The layout
  3. The furniture

Currently, this design process still requires some level of human input and the architect is able to tweak the design at each stage. You can try it yourself.

AI Architecture Generative Design Housing- 1OUuWZdrCfChTtxRl6CKB-g

Human interaction can manipulate the final result. Images courtesy of Stanislas Chaillou

But how is AI architecture design possible?

It might seem a little mind-boggling that a computer can generate such intricate designs, however, what it's actually doing is generating a series of colored pixels that are then converted into a design, based on input data. Each colored square becomes an area of the house; orange the bedroom, green the living room, etc. The walls are marked in black.

The technique uses a pix2pix GAN-model, a clever piece of, open-source code that he uses to convert simplistic blocks of color into a more sophisticated render. He then combines this with Google deep learning tools.

AI Architecture Generative Design Housing- 18GdmnQA1DX1xtjQAO5mAAQ

The three stages of an AI-generated architectural floor plan. Images courtesy of Stanislas Chaillou

The design process

To begin, the computer was "trained" to design the shape of a building footprint. Images were fed into it from real layouts of the city of Boston which taught the computer typical footprint shapes for a given plot of land and it used these to generate new footprints.

AI Architecture Generative Design Housing- Figure 4-1024x543

Image demonstrating AI-generated layouts based on input data. Images courtesy of Stanislas Chaillou

For the second step, the architect needs to manually place the front door (green square) and the main windows (red lines). This is then sent back to the computer and a layout is autogenerated based on input data from over 800 real-life apartment layouts.

AI Architecture Generative Design Housing- Figure 5-1024x543

Images show how the computer fills in possible layouts from a footprint and simplifies the shape of colors to form rooms. Images courtesy of Stanislas Chaillou

One completed the computer calculates the layout of furniture. The final results are pretty darn impressive!

AI Architecture Generative Design Housing- Figure 6-1024x543

Image illustrating how the colored spots (left) are converted to a fully functional layout (right). Images courtesy of Stanislas Chaillou

Multiple stories

Chaillou intends for his layouts to be used as apartments. In an apartment block, not all layouts are the same across floors. An option was added to give the architect the option to specify the individual footprint of any given floor. However load-bearing walls are not currently specifiable. This could be fixed using the same technique as door and window placement.

What's next?

In future research, Chaillou hopes to be able to push his technique to far more complex and a-typical layouts.

The generated designs are not vector, which means that they can not be imported into CAD software. More work needs to be done to make this a possibility.

AI Architecture Generative Design Housing- 14ydCrLudDzgM7MGxiizUVA AI Architecture Generative Design Housing- 1y1RfhQ7Mto2GFYzq8luKIA AI Architecture Generative Design Housing- 1K0EjfUpGae HAicHxscefw
AI Architecture Generative Design Housing- Figure 11 h AI Architecture Generative Design Housing- 1P5XR0dzXeuEUab4 YvQuoA AI Architecture Generative Design Housing- 1uV -l88Fd8g60UOhdKVe7g
AI Architecture Generative Design Housing- 13caHYViQ446K8vdGOsZQyw AI Architecture Generative Design Housing- Figure 11 g AI Architecture Generative Design Housing- Figure 11 i
Images Courtesy of Stanislas Chaillou

Should Architects start retraining?

The designs are not without limitations and the current process does not work without human input. I certainly wouldn't approve a building design without human review first! There's every chance you could end up with no bathroom or a completely inaccessible living area.

The computer isn't really designing anything new it's simply calculating similar results from existing designs. If computers were to design all the buildings of the future, the world would be a very boring place to live in. The current design is also limited to the input data from building layouts of one region and doesn't reflect the design styles and requirements of a global audience. A building in DC Washington might not be suitable for an apartment block in South Africa or Paris, for example.

However, this technique may have some merit in game design. In large-scale, free-roaming games, I can see a generative design solution like this being huge. It would save the game designer from having to repeat the same layout endlessly or manually input different layouts, whilst giving players a constantly evolving game.

Further reading:

Follow us on social media


This is Bricsys

Bricsys, part of Hexagon, is the global technology company that creates the BricsCAD® family of computer aided design products and the Bricsys 24/7 project collaboration platform.

© 2022 Bricsys NV. All rights reserved. 
Part of hexagon