Manifest Boston Stories

HUB Madness Lightning Interview: AI Portraits

Written by Zoe Dobuler | Apr 2, 2019 1:54:05 PM

HUB Madness 2019 presented by BNY Mellon is in full swing! We got the scoop on AI Portraits, a wild card project that shows us what happens with artificial intelligence collides with our own understanding of ourselves. Read on to hear more from lead researcher Mauro Martino, and what his team has in store for 2019.

Some background on the project:

AI Portraits is a website and research project created by Northeastern researchers Mauro Martino and Luca Stornaiuolo. Trained on millions of photos of actors and actresses, the AI will recreate your photo based on what it understands about faces, producing a "celebrity" version of you — and simultaneously revealing how AI's work. The resulting images examine the concept of identity, and allow us all to share the experience of being portrayed by an algorithm, of discovering how an AI sees us.

Where did the idea for AI Portraits come from? What inspired you to create it?
I have always been fascinated by the communicative power of portraits. There is a historical significance of the human portrait: It outlines the evolution of digital and analogical media. Whenever we have a change in our visual and aesthetic codes, whenever we have a technological advancement, the first repercussions are reflected in the style and technique of portraits. The portrait becomes a barometer that we use to test the aesthetic taste of the times.

AI Portraits was born almost by coincidence, from other research in the field of facial recognition. We started playing with generative models like the pix2pix, and slowly we found ourselves facing a new project.

What has been the biggest challenge of the project? How have you overcome it?
It was the choice of style, without a doubt. We didn't want to be too abstract and deform faces and make them unrecognizable, but on the other hand we didn't want to be too similar to the face portrayed.

We spent many months finding the right balance. My father has always done portraits of friends who came to visit us at home, and I think he is a talented artist. But despite being very good at drawing, I often saw dissatisfaction in the faces of his friends, and he was annoyed and repeated, "It's not a photograph, I designed you the way I see you." In AI Portraits we wanted to feel the same expectation, to find out how we are seen by the artist, and sometimes feel that disillusionment in not being portrayed as we would like. The portraits created by AI Portraits re-interpret our faces, and in the final result there must be something that makes it easy for our friends to recognize us, but that personally leaves us perplexed, as was the case with my father's friends.

What's next for AI Portraits or your research? What are you looking forward to accomplishing in 2019?
AI Portraits currently allows us to download only the low-resolution version of the portrait, and in a few weeks the high resolution will be available. It seems like a little detail, but in reality it required a lot of effort and experimentation. 

We are also considering whether to offer the option of being portrayed by another neural network model that we are developing. It would be like entering AI Portraits and being able to choose between two very different painters. I believe that in a few months we will make this gift to all our followers.

What did you think was the coolest thing to come out of Greater Boston last year (aside from your project, of course)?
I am happy to see Empatica here in Boston, I have known Matteo Lai for many years and I am pleased that he returned to his origins, when we both worked at MIT as researchers.

Cast your vote for AI Portraits before the Epic Eight round ends on April 3!