[DuskStar]

I wish I could add Gwern's Danbooru dataset [0] here - 2.7TB of labeled anime images. But they only support torrent files up to 10MB, and that's over 20MB for the full dataset or 12MB for the SFW low-rez set...

Incidentally, when the torrent file for your anime image collection passes 20MB, something has obviously gone very w̵r̵o̵n̵g̵ right.

0: https://www.gwern.net/Danbooru2019

[DuskStar]

I should probably point out that this dataset has been used for some machine learning tech demos in the past, for example This Waifu Does Not Exist [0], a StyleGAN-based automatic anime portrait generation tool. So it's not completely outside of what the site already hosts...

0: https://www.thiswaifudoesnotexist.net/

[gwern]

More than demos, papers too: https://www.gwern.net/Danbooru2019#applications