Latent Space Interpolation (2019)
Latent space interpolation between images of Mattie Tesfaldet and Xavier Snelgrove. The output was achieved using the novel Fourier Compositional Pattern Producing Networks (F-CPPNs). CPPNs are networks that map (x, y) pixel coordinates to (r, g, b) values via linear and non-linear transformations. F-CPPNs, however, map (x, y) pixel coordinates to Fourier coefficients, which are then fed to an Inverse Discrete Fourier Transform (IDFT) to produce (r, g, b) values. By explicitly modelling frequency information, greater output detail is achieved.
This work was showcased during Element AI’s “Hydrogen One” event back in May 2019. This work is based off of our paper, “Fourier-CPPNs for Image Synthesis.” It uses a novel reparameterization of CPPNs in the frequency domain, using Fourier analysis. This explicit modelling of frequency information allows higher frequency information to be captured in the output when compared to regular CPPNs. F-CPPNs and CPPNs have the benefit of being able to render at arbitrary resolutions, different than the resolutions they were optimized at. The submitted artwork came from a F-CPPN trained on two 224×224 inputs. During inference, multiple 1000×1000 outputs were easily achieved, interpolating between outputting an image of Mattie to an image of Xavier. The final product is a 1000×1000, 60 fps, 27 second video that appears aesthetically interesting.