Your photos are sent to our servers to generate portraits. We won’t use data from your photos for any other purpose and we'll immediately delete them.
AI PORTRAITS Ars
In the early 1500s Lorenzo Lotto and Giovan Battista Moroni began the psychological
analysis through portraiture. From this moment, the focus on the sitter's identity becomes the leitmotif in the history
of portrait. Portraits interpret the external beauty, social status, and then go beyond our body and face. A portrait
becomes a psychological analysis and a deep reflection on our existence.
AI Portraits Ars uses Artificial Intelligence to reproduce artistic human portraits, with
different styles and levels of abstraction.
For our model training, we adopt a data set of tens of thousands of paintings from the Early Renaissance
to Contemporary Art. This type of portraiture is quite distinctive of the Western artistic tradition.
Training our models on a data set with such strong bias leads us to reflect on the importance of AI
fairness. In the previous work AI Portraits Celebrity, we explored the concept of micro-bias
linked to the training data of only actors, which in some way imposes an actorization of the
user's portrait: “a collection of faces from the society of spectacle that are sedimented in the
neural network, and vaporize my selfie in a cinematographic self."
AI Portraits Ars introduces a very different type of bias with unique themes to explore.
Portraiture has a diverse and ancient history spanning geography and religion. It is discussed in the
works of Aristotle, Plato, Cicero, and Pliny the Elder. We have incredible examples such as the
Fayum portraits from Egypt. In other regions of Africa, there are the masks that
are stylized to represent a character. In China, a portrait tradition can be traced back to the
Han dynasty in 200 BC. In India, we have portrait miniature painting with the Mughal
dynasty of the 17th century. In traditional Jewish and Islamic cultures, a
prohibition on imagery made portraiture a taboo.
In AI Portraits Ars, we focus on the 15th century Europe, which is considered by art historians
like Joanna Woodall, Shearer West, John Berger and many others, as a stylistic inflection point
in the history of portraiture marked by the emergence of realistic depictions of individuals.
Before the 15th century, the practice of commissioned painted portraits of individual sitters was rare.
Perhaps this change in style reflects a shift in societal values toward individualism. The Italian
Renaissance of the late 13th century, with the writings of Dante and the paintings of
Giotto, was a period of increased self-consciousness, in which concepts of unique individual
identity began to be verbalized. This was followed by the Renaissance glorification of the genius of
woman and man, the representation of the unique and extraordinary ability of the human mind.
We encourage you to experiment with the tool as a way of exploring the bias of the model. For
example, try smiling or laughing in your input image. What do you see? Does the model produce an image
without a smile or laugh?
Portrait masters rarely paint smiling people because smiles and laughter were commonly associated with
a more comic aspect of genre painting, and because the display of such an overt expression as smiling
can seem to distort the face of the sitter. This inability of artificial intelligence to reproduce
our smiles is teaching us something about the history of art.
This and other biases that emerge in reproducing our photos with AI Portraits Ars are therefore an
indirect exploration of the history of art and portraiture.
AI PORTRAITS Ars
AI Portraits Ars is able to paint portraits in real
time at 4k resolution. You will find yourself in front of a mirror and feel thousands Rembrandt,
Caravaggio, Titian portraying you moment after moment.
We have trained Generative Adversarial Network (GAN) models to reproduce human portraits, with different styles and levels of
GANs are a very popular class of deep generative models. They are trained to
learn a mapping of a latent vector z∈Z to a generated image y
= G(z) with G being the generator. The latent space Z describes all possible
AI Portraits Ars pushes us towards an intuitive and playful way of interacting with state-of-the-art GAN
models. By showing our face to the neural network, we walk through the Z space and identify the
vector that best describes our face in the multidimensional space of the GAN.
We trained AI Portraits Ars using our GAN on 45,000 portrait images. To allow insertion of own images
into the latent space of a model, we trained an inverter that can approximate the latent vector z
= I(x) from an image x.
This is not a style transfer
With AI Portraits Ars anyone is able to use GAN models to generate a new painting, where facial
lines are completely redesigned. The model decides for itself which style to use for the portrait.
Details of the face and background contribute to direct the model towards a style.
In style transfer, there is usually a strong alteration of colors, but the features of the photo
remain unchanged. AI Portraits Ars creates new forms, beyond altering the style of an existing
Your photos are sent to our servers to generate portraits. We won’t use data from your photos for any
other purpose and we'll immediately delete them.
AI PORTRAITS Ars
MIT-IBM Watson AI Lab
Northeastern University @martino_design