We can build:
A service similar to stablediffusionweb.com for generating images enough fast. Because open source!
Or a service for storing images like postimages.org.
Or we can make a game based on generating pictures.

For example, a service for generating images: all free images are licensed under a public domain. HERE is our data set for building deduplication.
A service for storing images is used for publication By default. . Also a data set!
The game can also generate a large amount of images. There are sketches. The rules should be quite simple and unusual. Longest but most interesting way i think!
And what do you think?

Another idea what to create on the top of deduplication service.
Let's imagine we have an audio podcast service. It recognized sentences from a recorded audio and draws image for each one.
User may find this speaker by rating images and this only single way to describe yourself and find people with similar interests.
And we can add easy-tu-use audio processing and beautifying features. Audio compressor, noice reduction by single click.

We can effectively deduplicate audio and images. From the technical point of view, this is a 100% matching project.