2021, Images generated by guiding a FFT-based generator with CLIP

Vertebrata is work-in-progress series that experiments with image generation using the Contrastive Language-Image Pre-training (CLIP) model.

This series depicts different animal species from each of the five subgroups of vertebrate animals: bird, fish, reptiles, amphibians and mammals. The images are generated using a FFT-based generator, guided using CLIP to match the prompts provided in natural language. The matching process is performed using gradient descent, optimizing the FFT parameters to obtain the final images. For each of the images in the series, the generation was guided using prompts including the desired species and habitat.

Aesthetically, the aim of the experiments was to depart from the overly familiar aesthetics of the Deep Dream-based images into a more painterly result. The animals on each of the images are not perfectly defined, but hinted, resulting in images in which the animals appear at a glance, but disappear the moment one focuses on each of them, leaving in its place an abstract shape.