Updated: Sep 27, 2019
Following this post, I used Deep Learning to extract the style of my personal art pieces and use it to make a filter for one of my photos. With the weights of a pre-trained neural network (VGG19), I have produced a few personalized filters with my own artwork!
Imagine an app where anyone can choose their favorite artists or styles and let their phone make a new piece of art. You could be in a museum, standing in front of one of your favorite paintings. You take a photo of both the painting and yourself, and through this transformation, you become the main character of this painting!
Taking this idea a bit further, we can also mix artists when sometimes only one artistic style can't express our feelings. Here I try to make a combined style image with Van Gogh's Sunflowers and Picasso's Les Demoiselles d'Avignon. In the below image, I applied each of the styles individually to a photo of me in front of the Fontaines de la Concorde in Paris.
Middle top: Van Gogh's Sunflower; Mid bottom: Picasso's Les Demoiselles d'Avignon; Left: Picasso style Right: Van Gogh Style
To combine the two styles, I use an idea inspired by one of my research projects. In the project, I designed adaptable materials using oscillatory optimization goals. I'm letting the learner, which in this case is the optimizer that mimics the style of a painting, repeatedly change its mind about which style it prefers. The learner oscillates between learning how to paint like Van Gogh and learning how to paint like Picasso. After several cycles, the learner ends up painting in a style of its own that has qualities of both!
The resulting image has the color of the sunflowers in the fountain, while the building next to it has the color of five nude ladies. The colors in the rest of the photo as well as the texture of the water and ledge are a mix between the two. The vortex strokes that Van Gogh is famous for find themselves mixing with the broader structure in the Picasso-styled sky, creating a fluid of colors in the combined image. Finally, the human (me!) in the image is also more abstract and blends into the background.
Each filtered image in this post started off as the original photo with some small random noise added to it. Depending on this noise and how you originally trained the neural network, there is a lot of uncertainty in how the resulting image will look. This uncertainty somehow gives me a greater appreciation for machine-made art since I can relate to the experience of never being able to make two identical pieces of art.
As machine learning tools develop, people will be less and less limited in their ability to express themselves artistically. With just ideas and imagination, everyone can create their own art!
Anyways, this is my first blog post! Thanks for reading!:D
In case you want to check out the code, take look at my GitHub page!