AI AND GANS FOR FACADE GENERATION

 METHODOLOGY

TESTING VQGAN + CLIP FOR FACADE GENERATION 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

We realized that to achieve a clearer outcome we should avoid general terms and be specific about the text we write, moreover adding artistic styles and references allowed us to achieve a clearer outcome.

TESTING STYLEGAN 2 + LATENT WALKS FOR FACADE GENERATION

 

 

 

 

 

 

 

 

 

We thought that it would be interesting to see what the computer would make of the relationship between nature and architecture. To accomplish this, we had to first gather images of both facades and organic patterns in nature. In total we had around 1000 images of both. Next we trained the Style Gan to be able to recreate images of facades and organic patterns. We trained for around 48 hours. And then from this training we were able to do latent walks between the two styles of images.

LATENT WALKS IMAGES

AI GENERATED FACADE IMAGES

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

It was interesting to see, even though we trained the model with all different façade types and building uses, the model tended to generate buildings with similar window to wall ratios in the style of a standard mixed used, modern building Not only could the model make buildings, but also surrounding scenery and roads. It wasn’t always perfect as you can see on the right. But we believe that if we had let the model train for longer it would have gotten there.

AI GENERATED NATURAL PATTERNS

 

 

 

 

 

 

 

 

Since we trained the model also with organic images, we were also able to generate unique organic images. Sometimes these images came out fairly clearly, on tho other hand some of them got blended with the façade images.

LATENT WALKS ANIMATIONS

 

CONCLUSION

In conclusion, vqgan was very quick and efficient, but it had troubles recognizing certain facade related words. We had more success with style gan, but had we more time,  we would have had more we would have obtained more images to train the model from and would have trained the model for longer. We are happy with the results we got. The generated images, while not perfect, are definitely original. We believe that this process could be helpful to designers coming up with initial designs and concepts for a building facade

CREDITS

STYLEGANS-LATENT WALKS – VQGAN FOR FACADE GENERATION is a project of IaaC, Institute for Advanced Architecture of Catalonia
developed at digital tools foe generative and building design  in 2021/2022 by:
Students: Sophie Moore | Charbel Baliss
Faculty: Oana Taut | Aleksander Mastalski