Home

Awesome

laion-prepro

Get billions of image+url from the laion datasets and preprocess them.

<img src="./assets/french_cats.png" width="500">

This repository can be run on

What is laion ?

The laion project has for objective to use commoncrawl to retrieve billions of aligned image+text pairs. It is composed of a central server that track the progress of decentralized (run by anyone) workers that process small chunks of commoncrawl. Currently, 5B such pairs have already been retrieved. Read more about it at the laion 400M release post

What can be done with these dataset ?

Vision and language modeling has been taking off in 2021. Here are some pointers about what this kind of image + text datasets unlocks and why it seems really interesting:

Since then, several efforts have been organized to replicate DALL-E. People organized initially around this awesome dalle replication repository DALLE-pytorch with some nice results that can be seen in the readme. More recently as part of an huggingface events, new results have been achieved (see dalle mini report ) and an online demo is now available dalle-mini demo

The replication effort is still far from achieving the same performance as the original dalle, and it seems it's possible to go even further. Some people also want to make a better CLIP to produce even better generated art.

A large part of the results that can be achieved with such models is thanks to data. Large amount of data. Before laion 400M, the largest open dataset for (image, text) pairs are in the order of 10M (see DALLE-datasets ), which is enough to train okay models, but not enough to reach the best performance. Having a public dataset with hundred of millions of pairs will help a lot to build these image+text models.

Visualization of the dataset

Check the colab and the web demo

laion5B

laion5B and laion400m processing is overall the same, but laion5B being 10x, it required making everything distributed

Read more at laion5B/README.md

laion400m

See laion400m/README.md