erlich is the text2image latent diffusion model from CompVis (with additions from glid-3-xl) finetuned on a dataset collected from LAION-5B named Large Logo Dataset. It consists of roughly 100K images of logos with captions generated via BLIP using aggressive re-ranking.