opendiffusionai/cc12m-2mp-realistic

Overview

A subset of the "CC12m" dataset. Size varying between 2mp <= x < 4mp. Anything actually 4mp or larger will be in one of our "4mp" datasets. This dataset is created for if you basically need more images and dont mind a little less quality.

Quality

I have filtered out as many watermarks, etc. as possible using AI models. I have also thrown out stupid black-and-white photos, because they poison normal image prompting.

This is NOT HAND CURATED, unlike some of our other datasets, so do understand there are some considerably less quality images hiding in here. That being said, I did take advantage of the WD14 tag, "head out of frame" to remove THOSE shots.

Captions

Trying something a little new. This dataset has FOUR choices for captions:

• —LLaVA, long format
• —LLaVA, short format
• —Moondream2b (which is shortish, but different content than above)
• —WD14 v3 (tag style)

Using multiple captions

If you want to use the images with multiple caption styles, you can download the set ONCE.... then copy the image tree, and use the util at https://github.com/ppbrown/vlm-utils/blob/main/dataset_scripts/extracttxtfromjsonl.py to extract one of the other caption sets from the .jsonl file here

Downloading

Grab the "crawl.sh" script (which is just a wrapper around img2dataset) to grab the actual images mentioned in this one. Make sure to edit the script to set which caption type you want.