Big Sleep is an open-source command-line tool that generates images from text by combining OpenAI’s CLIP with a BigGAN — one of the playful experiments that kicked off the text-to-image wave. It has around 2,600 GitHub stars.
Created by Phil Wang (lucidrains) and based on a technique by Ryan Murdock, Big Sleep optimises a BigGAN’s latent input until CLIP scores the resulting image as a good match for your prompt. You type a phrase — “a lantern dangling from a tree in a foggy graveyard” — and watch the image slowly dream itself into shape over many iterations. There is no diffusion model and no web UI; it is a single-purpose CLI from the 2021 era of generative art. Despite its age, the repo is a gentle introduction to CLIP-guided generation, and its sample gallery shows the dreamy, painterly look the technique is known for. The repository was last updated in 2022, so today it is best enjoyed as a piece of AI history and a hackable toy rather than a state-of-the-art generator.
Big Sleep is completely free and open source under the MIT license. No accounts, API keys or fees are involved; it runs on your own machine, and a CUDA GPU makes it much faster.
Generative-art hobbyists, students and the curious who want to see how text-to-image worked just before diffusion models took over.