A Ideas
This week I’ve been experimenting with Style transfer, “which is the technique of recomposing images in the style of other images. These were mostly created using Justin Johnson’s code based on the paper by Gatys, Ecker, and Bethge demonstrating a method for restyling images using convolutional neural networks.”
The results from these networks are particularly impressive for their ability to apply the texture and colors that make an artwork unique. Something I love about style transfer is how in creating an image you end up discovering and looking at art from throughout history. However, interesting questions arise concerning who the artist of the final piece is. Is it me, the AI, the style reference Artist, or the writer of the algorithm?
There are many potential applications and uses for this technique today and for the future. In the near future an AI could be trained to choose a style it thinks goes well with an image that it is sent as input. The AI could even be the one that choose the input image based on what it perceives your preference to be, at which point it would have full control over the final images.
Some exciting applications that are possible now is using real-time style transfer to transform video and images. A public installation with this effect would allow people to experience an artist’s style interactively from a new perspective. Soon it could be possible to run these networks on mobile devices which would open up a new way of creating compelling stylized photography on the go.
Nvidia has developed an impressive algorithm that can generate hyper-realistic faces as well as other images it has been trained on. The model takes in 3 reference images and generates an entirely new face from them. These networks have a lot of potential applications but also can be misused in many ways.
A compelling use for this technology when it is possible to run the network in real-time would be an installation art piece where viewers see a mirror of their transformed face. There could be three stations where each person sees a unique face that is the composite of all the participants. The resulting face would be familiar and follow your facial expressions. The age, race, and gender could all be changed, allowing participants to reflect on their identity and how all people are actually very similar and connected.
On a less altruistic note, this same technology can be used to create convincing but fake images and videos that can be used to manipulate perceptions. A few examples include, fraudulent ID’s, fake news and impersonation. As these technologies advance it becomes important for new ways of detecting truth from fabrication to also be developed.
