HouseOfGAN
In May 2018, I had the opportunity to take part in the SmartGeometry "Machine Minds" workshop in Toronto. This is one of the results from the "FreshEyes" cluster.
What if AI could team up with architects as design partners?
This is a question that has been in my mind for quite a long time... This also happened to be one of the clusters theme for the SmartGeometry 2018 workshop in Toronto.
Part I
Teamed up with Gabriel Payant and Thimoty Logan, we decided to train a Machine Learning model with a rather large data set of images and then set up a GAN (Generative Adversial Network) to create new computer-made houses.
So we started by using classified 3D house models from students at University of Berkley, and a Grasshopper script we recieved for the workshop, starting from the 3D to the generated black and white images. Each image represents a view plan, a front elevation and two side views of the houses.
The images have then been used to train a ML model with the help of Lobe, a web-based interface to train and test ML models.
Check out their website at https://lobe.ai/, this brings Machine Learning to everyone !
Once the ML model was trained, we then set up a GAN with Kyle Steinfe (UC Berkley). After a while, the results were pretty impressive. The next images have all been created by the computer on its own.
Part II
At the same time, we also wanted to automate 3D model reconstruction out of these GAN-made images of houses. We started by using the first set of black and white images shown earlier, and came with different results depending on different methods.
You can see in the next GIF a 3D model being created from a black and white image : a chocolate alike house.
After getting some promising results, we finally applied these 3D recreation methods to some of the GAN generated black and white tiles and the results were quite fascinating to watch.
If this work intrigues you, I definitely recommend you to check the website where all the work of every participant has been uploaded and better explained.
A huge thanks to the cluster's champions : Adam Menges (Lobe.ai), Kat Park (SOM), Kyle Steinfe (UC Berkley) and Samantha Parker (SOM). This work wouldn't have been possible without their research work and their help all along the week. This has been a really inspiring workshop, and it certainly opens up new ways of designing Architecture.
Also thank you to all the participants, especially Gabriel Payant and Thimoty Logan, for the good collaboration and all the work done during the week, even with bad poutines…
Antoine -