Generating Large Images from Latent Vectors

In some domains of digital generative art, an artist would typically not work with an image editor directly to create an artwork. Typically, the artist would program a set of routines that would generate the actual images. These routines compose of instructions to tell the machine to draw lines and shapes at certain coordinates, and manipulate colours in some mathematically defined way. The final artwork, which may be presented as a pixellated image, or printed out on physical medium, can be entirely captured and defined by a set of mathematical routines. Many natural images have interesting mathematical properties. Simple math functions have been written to generate natural fractal-like patterns such as tree branches and snowflakes. Like fractals, a simple set of mathematical rules can sometimes generate a highly complicated image that can be zoomed-in or zoomed-out indefinitely. Once such a function is found, then the image can be automatically scaled up and down, or stretched around, by just scaling the inputs. If this function has some fun properties or exhibit some internal structure, it will be interesting to see what the image looks like if we blow up the image to a very high resolution much bigger than the original image. This function can also be defined as a neural network, with arbitrary architectures.