We're training a text-to-image model from scratch and open-sourcing it (photoroom.com)