Here is Gwern dataset https://www.gwern.net/Danbooru2018 with **2.536.329** Danbooru images
**till 01.01.2019 rating:safe resized to 512x512 px** with some meta-information
used for image recognition training (deep danbooru resnet model etc.)
**in zipped format, acceptible to all torrent clients.**
Meta information included in "initial" JSON format and "normalized" 3-tables CSV
(posts with some additional stats, taglist with some additional info, tags occurrences in posts).
Here - https://nyaa.si/view/1206322 - is a next step: 1.227.622 thumbnails from several sources
selected (only ~360.000 taken from this release) with much higher initial images quality
and enriched with much more calculated metadata (including face detected).