Generative Search Engines: Initial Experiments

Text-to-image generation involves producing an image which somehow reflects a given text prompt. We investigate the creative potential of a novel approach to this task. This employs three neural models working in concert: a generative adversarial network producing images with input latent vectors chosen by a search guided by a pair of models able to assess the appropriateness of a generated image for a text prompt. For evaluation purposes, we re-frame the task to be analogous to Google-like image search and introduce notions of efficiency, fidelity, variety, sophistication and coherence in the generated images. We have found the approach remarkably successful and explore here its potential for various creative tasks. We propose two approaches to increase efficiency in the generative process, and we evaluate the approach in an experiment simulating commercial design usage. We further suggest ways in which a generative search engine could be used in videogame design via standard usages and via an always-on modality for continuous creativity.