A 1.4B parameter text2im model from CompVis, finetuned on CLIP text embeds and curated data.
Want to make some of these yourself?