The BigSleep uses AI to generate pictures that don’t exist

BigSleep uses two AI models to generate images that don't exist.
Image: BigSleep

Artificial intelligence (AI) can be used in a wide variety of ways. One area that has seen several advancements in recent years is image generation. Today’s neural networks are able to come up with impressive pictures from scratch, often generating new works that hadn’t existed up until that point.

A perfect example of this is The BigSleep. It combines two AI models to generate pictures based on a provided text prompt, Digital Trends reports. Of course, those pictures didn’t exist until the AI generated them. It’s a bizarre and thought-provoking experiment that pushes the limits of what AI can do.

Twice the Fun

BigSleep was created by Ryan Murdock, a 23-year-old cognitive neuroscience student from the University of Utah. In an interview with Digital Trends, he discussed how the program came to be and revealed a bit about how it works.

Murdock says, “At a high level, BigSleep works by combining two neural networks: BigGAN and CLIP.”

The first is well-known in the AI community. BigGAN was created by Google and is a general adversarial network. In essence, refines its output results over time by playing a sort of tug-of-war with the images, or noise, it generates.

Digital Trends notes that CLIP is a neural network designed to match images to descriptions. It was created by OpenAI, the company known for its video game-playing algorithms. CLIP has plenty of real-world uses, but its role in BigSleep is to help match the user’s text input to an appropriate picture.

BigGAN is responsible for generating noise output (to ultimately create an image) while CLIP determines if that output matches the user’s original input. Ultimately, BigSleep can generate images that match each prompt in about three minutes. That’s an impressive feat considering the amount of computing power that goes into each attempt.

Murdock tells Digital Trends, “BigSleep is significant because it can generate a wide variety of concepts and objects fairly well at 512 x 512 pixel resolution. Previous work has produced impressive results, but, by my knowledge, much of it has been restricted to lower-resolution images and more everyday objects.”

By contrast, BigSleep is able to generate images of more abstract ideas, such as “an intricate drawing of eternity” (pictured above).

AI Gets Creative

Although most of today’s real-world applications for AI are some form of data analysis or real-time monitoring, the technology is capable of so much more. It’s likely that the popularity of AI-generated media will continue to grow in the coming years.

So far, AI has already demonstrated proficiency at creating human-like text, original music, and various forms of art. In fact, an AI painting sold at an art auction for $432,500 back in 2018.

Programs like BigSleep demonstrate that the technology is not only getting better but that it is also getting more accessible. Murdock’s project can be found for free online via a Google Colab notebook.

Using it is a bit confusing for those that aren’t familiar with Google Colab projects, but exploring it is worth the time. A group of Reddit users has also posted a helpful step-by-step guide for those that want to try out BigSleep for themselves.


Please enter your comment!
Please enter your name here