Navigating Neural Fields with Vision-Language Models

Generative art is an enduring discipline in the field of computer science that has traditionally taken on a wide variety of creative implementations. But if we view the current landscape of generative art without a discerning eye, the scope of techniques and methods may look quite flat – only diffusion models, LLMs, and their LoRAs to be seen. In this work we aim to showcase a variation of an older technique for image generation that can create striking visual art without relying on training data, exhaustive computation, or narrowly defined priors. Specifically, we revisit the CPPN-NEAT algorithm, and retool it to be more amenable to current generative model workflows. Instead of evolutionary augmentation, we generate random Watts-Strogatz graphs, convert them to neural fields, and generate the resulting image at an arbitrary resolution. We obtain high-quality samples by using an off-the-shelf VLM to make pairwise selections between generated examples. Images that survive multiple rounds are selected for final human review. This automated procedure is simple, and allows us to quickly and easily generate 12000px x 12000px images on a consumer desktop machine, in a style that is distinct from publicly-available image generation models.