Skip to main content

Discovering the Hidden Vocabulary of DALLE-2

ยท One min read
Zain Hasan

A preview of the paper

TIL that text2image diffusion models learn and use a secret language.

Tested this with the new DALL-E-3 and it works!๐Ÿคฏ

Read a couple of papers and they mentioned that diffusion models when forced to output text generate images of gibberish words.

If you take those words and pass them back in as prompts, the model can draw for you what the word means to it.

For example: "cagama gur gerano" = "a fantasy creature"

I tested this for the newly released DALL-E-3 model and, interestingly, even when told to generate English it still uses this secret learned language instead.

Below is a conversation about fantasy creatures between two farmers in this secret language.

Initial prompt: "Two farmers talking about vegetables, with english subtitles."

After this just prompt the model with individual and word pairs to get images with secret words. I share examples below.

Prompt: "cagama gur gerano"


๐Ÿ”— arXiv Link

๐Ÿ“œ Download paper

Ready to start building?โ€‹

Check out the Quickstart tutorial, and begin building amazing apps with the free trial of Weaviate Cloud (WCD).

Don't want to miss another blog post?

Sign up for our bi-weekly newsletter to stay updated!

By submitting, I agree to the Terms of Service and Privacy Policy.