Make a "prompt" view of Google's Imagen's Drawbench

Welcome to the wormhole of fun that is text-to-image generators! Just today, Google announced the release of Imagen - “a text-to-image diffusion model with an unprecedented degree of photorealism and a deep level of language understanding.”

A curated set of quirky but realistic AI-generated images involving cute animals like so – :panda_face: :dog: :cat: – greet you on the landing page. They are based on a list of 200 prompts provided by Drawbench: a comprehensive benchmark for text-to-image generators. The list of prompts consists of a handful of categories, from which one could ask “what is the most common category across this list?” … followed by, “can this be visualized?” :woman_shrugging:

Quick answer: “yes.”

With a csv download of this google sheet and a quick drag-and-drop into GraphXR, it is possible to extract on categories and see - walla! Reddit is in the lead with the most prompts. No surprise there!

And this is not the first time we fell into the wormhole of graphing AI-generated images. Last time it was all about dreams! :sleeping:

With a neural network image generator (Aleph2Image Delta model built from OpenAI’s Clip model and DALL-E decoder developed by tech artist @advadnoun), we transformed dream journal entries cataloged in UCSC’s dreambank into a series of AI-generated “dream-like” images, layering graph technology and interactive views to navigate the dreamscape. You can revisit this Neo4j talk to see how it all came together.

If curious about exploratory data visualization, get started with a free GraphXR account today. You can follow these learning resources or, simply, get in contact with us. We want to hear from you! :hugs: