AI Design – Generative Adversarial Nets

Today I want to talk to you about the future and no I don’t mean flying cars, not this time anyway. It’s no secret that at the moment, the whole of the tech industry is obsessed with AI. For me, one of the most interesting developments is Generative Adversarial Nets (GAN). Before your eyes glaze over, just glance at the images below, all are AI design a.k.a GAN! GAN was conceived by a man called Ian J. Goodfellow and was dreamt up, as most brilliant ideas are, after a conversation between friends at the pub. You can read more about the story here.

GAN computer generated design
Inside the weird and wonderful mind of a computer. Here are a sample of autogenerated bedrooms, kitchens, churches and dining rooms. These images and the ones in the header can be found here in the original paper. Although far from perfect and still at low resolution, these images are completely computer generated and are quite amazing.

What are Generative Adversarial Nets?

In the published paper, the system is likened to that of money counterfeiters and the police competing against each other. As the counterfeiters get better at what they do, so must the police. The idea behind GAN is pretty much the same: the discriminator (the police) is in competition with the generator (the counterfeiters) and thus data is sent backwards and forwards until the discriminator can no longer detect the fake. Remarkably, GAN is unsupervised learning. It is also possible to achieve better results, with a reduced amount of input data compared to conventional methods.

A quick web search for “Generative Adversarial Nets” will throw up all sorts of weird and wonderful things. From eerily realistic, fake celebrities, to AI design horses and even interior and exterior AI design.

What does this mean for CAD?

More recently, GAN has been used to generate 3D models. For us CAD fanatics, this is where things start to get interesting.

Generative Adversarial Net 3D data recognition the IKEA project AI design
Computer recognizing 3D objects from 2D drawings. Images taken from LearningaProbabilisticLatentSpaceofObject Shapesvia3DGenerative-AdversarialModeling paper.

Some clever folks created 3D-VAE-GA. Don’t worry, computers haven’t yet reached a level of conscious thinking where they take up veganism. It’s simply a combination of GAN with a “variational autoencoder”. This is basically a very clever way of teaching a computer to detect 3D objects from 2D images. This works even when the object is obscured! You can read the full paper here, or If you’re not a fan of technical reading, you can check the 2-minute summary. This builds from a project from some equally clever minds at Massachusetts Institute of Technology’s Ikea Project. The full source code and dataset is available to download.

What is most striking is how well the computer is able to recognize a 3D object on a 2D image. No denying that the results are far from perfect, but just take a moment to imagine how this technology could be used in the future!

Maybe even more interestingly, in the same paper, they demonstrate how they were able to successfully teach a computer how to learn object types and generate its own designs in simplistic 64 x 64 x 64 block shapes.


Generative Adversarial Net 3D modeling AI design
64x64x64 renderings of computer-generated objects for data types, gun, chair, car, sofa, table. To the right, the most similar object from the original source data is shown. It is easy to see that, although similar, the computer-generated objects are not the same as the source. Image also taken from the same paper.

So will a computer take your job?

Maybe not right away, personally, I didn’t think some of those churches looked very stable! A lot of the designs remind me of the creations of the android Commander Data, from the TV show Star Trek. He was only ever able to combine humanity’s existing ideas, but never to generate his own. The designs simply lack the “human touch”. Then again, maybe, given a few more years development, we might see CAD AI design where one has only to specify a few parameters: chair, wooden, 1950s and wait for a computer to provide a fully customized design, every time.

What do you think? Will a computer ever replace a designer, an architect or an engineer? Do you think you can do better? Download the trial now and prove it.

Rose Barfield

Rose is Bricsys' English Content Creator. She has worked in the Automotive, Aerospace and Defense industries as a Technical Illustrator, before coming to BricsCAD. She loves cars, vectors and 3D printing.

Add comment

Try BricsCAD 30 days for free

Subscribe to our newsletter

Want to receive the monthly newsletter in your inbox with the top content on the Bricsys Blog?