Paper Dreams

Guillermo Bernal, Lily Zhou

Paper Dreams explores how human creativity can be supported by artificial intelligence.

Prior research on AI and creativity has primarily focused on using machine intelligence to learn “creativity” from humans, from transfer learning of artistic styles to generation of original paintings, poems, and music by the machine. However, the application of current machine learning algorithms to the augmentation of human creativity is a relatively unexplored area. By creating a dynamic human-machine back-and-forth and working with representations inside machine learning models, we can offer people new "smart" tools for brainstorming and creative expression.

Our system in its current form provides users with a canvas where they draw parts of a sketch that the machine tries to recognize and offers to complete. We do this by building a neural network which takes a small number of input variables, called latent variables, and produces the entire sketch as output.

In addition, the system also helps users move forward with their creation. To date, Paper Dreams augments the drawing experience in three different ways: by adding textures/colors, suggesting other elements/drawings for the scene, and introducing serendipity. To adjust the level of serendipity, the user has control of a dial that determines how "predictable" vs "unpredictable/out there" they want these machine additions to be.

While the current version of Paper Dreams works with 2D sketches, the same human machine collaborative creativity approach could be applied to other domains such as 3D models and more.

Some of the driving questions for this project are:

  • To what extent are these new tools enabling creativity?
  • Can they be used to generate ideas which are truly surprising and new, or are the ideas clichés, based on trivial recombinations of existing ideas?
  • Can such systems be used to develop fundamental new interface primitives?
  • How will those new primitives change and expand the way humans think?