Making Art Using Algorithms
Making Art Using Algorithms
Making Art Using Algorithms
by ngaskins
More and more artists are using arti cial intelligence or AI in their art making processes. Deep learning is a subset of AI
that includes computer vision and neural (image) style transfer, which refers to algorithms that learn the details of a
source photo, then combine the contents of this image with styles of di erent ones, e ectively transferring the styles to
the original image.
This lesson covers the basic steps for creating unique images using Deep Dream Generator, a computer vision platform
that allows users to input photos into the program and transform them through an AI algorithm. This video by
Computerphile, an educational Youtube channel about videos, explains just how Deep Dream works.
//www.youtube.com/embed/BsSmBPmPeYQ
For this lesson, you can set up a free account at Deep Dream Generator.
Supplies:
Setting up a free account at Deep Dream Generator (DDG) will give you "Newbie" level access. You get more perks based
on the amount of images you generate and share with the DDG community (see photo).
You can become a "Member" after you make 5 Dreams and your account is more than one day old.
Note: Currently, I'm a "Deep Dreamer" and "DDG Evangelist", which means I can generate more higher resolution images
in a given day.
To get started with DDG you will need source images (photos). Unless you are a professional photographer or know
someone with images you can use, you can search for images in the public domain. The public domain consists of all the
creative work to which no exclusive intellectual property rights apply. Those rights may have expired, been forfeited,
expressly waived, or may be inapplicable.
For example, this photo of Billie Holliday is from the William P. Gottlieb collection at the Library of Congress. In
accordance with the wishes of Gottlieb, the photographs in the collection entered into the public domain on February
16, 2010.
Click on the orange "Generate" button to choose an image style. DDG provides free Default and Popular styles but you
can also upload your own images. The algorithm is set up to re-create/process your source image in the style of your
selected image.
Note: You can only use one image style at one time but there are ways around this.
For image styles you can use images of fabrics/textiles, nature, etc. You can also use image editing software such as
Adobe Photoshop or Processing programming language to generate styles.
Examples of some of the styles I use:
Image style #1 is a collage of African fabrics; Image #2 consists of four separate images that are grouped together using
Photoshop; and Image #3 was generated using Processing arrays.
The most time consuming aspect of the DDG process is nding the right settings to generate your new image. Settings
include Resolution, Enhance, Style Weight, and Style Scale. Click the "Settings Info:" link to learn more about each one.
Over time, I found optimal selections for each setting but it really depends on the nal image I want the algorithm to
generate.
Remember: The machine (algorithm) is set up to receive input from you and the more info it gets and the higher the
image resolutions the better the nal output.
Depending on the version (thin vs deep), level, and resolution of the images you may have to wait several minutes for
the generator to display the nal image (output). You have a choice whether or not to make the image public or keep it
private. Remember: more likes result in more perks.
My "Lady Day Redux" image was shared and it's an example of what I call a "one-pass" image. This means I ran the
algorithm once and was satis ed with the result. Usually, I generate several versions of a source photo using di erent
styles. Then, I composite them in Photoshop.
Image compositing is a task of combining di erent images to compose a new image. A common use case is background
replacement of portrait images. For example, in this close up of my Helen Keller portrait, you can see di erent visual
elements I added after DDG nished generating the image. I used this technique to create 10 more "featured futurist"
portraits for the Smithsonian in Washington, DC. Each image was printed on 6' tall totems and displayed throughout the
exhibition.
My digital art process involves a machine generating algorithm but I also rely on my knowledge as a digital artist to
guide and complete the work. I rely on my color theory and foundational art classes to make choices in the source
images and styles I choose. My styles are often custom-made and there are speci c settings I tend to use (ex. Deep Style).
Experimentation has led to better results over time.