What it does

Big Sleep is a small open-source command-line tool for generating images from written descriptions. It combines OpenAI's CLIP with the generator from a BigGAN, and the repository wraps that technique so anyone with a GPU can run it. After installing it with pip, you run a one-line command such as dream followed by a phrase, and the images are saved to the folder where you ran it. It can also be called from Python code, where you can set the learning rate and how often progress images are saved. You can train on several phrases at once by separating them with a bar, and list phrases to steer away from. Options include saving progress images, saving the best-scoring image, using a larger vision model if you have enough memory, and limiting the number of BigGAN classes for extra stability. The repository is released under the MIT license. It notes that the GAN often drifts into noise, which is why the best-image option exists.

Capabilities

Text-to-image from a one-line terminal command
Python interface for use inside your own code
Combine several prompts, or penalize unwanted ones
Save progress images or only the best-scoring result
Optional larger vision model for better output

When to use it

You have a GPU and want to experiment with CLIP-guided image generation locally
You want a simple scriptable text-to-image tool rather than a hosted service

Alternatives to explore

Deep Daze