What dataset was Kroma trained on?
I'm mostly curious how many distinct concepts from places like Danbooru/e621 are in there, aside from just photorealistic photos.
Best approach is to lurk on lodestones datasets like https://huggingface.co/datasets/lodestones/booru-essence or https://huggingface.co/datasets/lodestones/e621-captions . Since they exist , they are most likely used to train Krea 2.
For my own usage I compiled the parquets listed in the repo links above into perchance text generators (you can cycle between the sorted categories using prev gen / next gen , 68 generators in total , 300-1000 prompts in each one , totaling ~40K prompts)
https://perchance.org/fusion-t2i-essence-solo-true-multi-false-general-cluster-001-1
⭐️ UPDATE : New version only for Danbooru using the 2026 updated sets:
https://perchance.org/fusion-t2i-danbooru-general-cluster-000-1
https://perchance.org/fusion-t2i-danbooru-questionable-cluster-000-1
https://perchance.org/fusion-t2i-danbooru-sensitive-cluster-000-1
https://perchance.org/fusion-t2i-danbooru-explicit-cluster-000-1
Using https://huggingface.co/datasets/RicemanT/booru-essence-2026
//----//
⭐️ UPDATE 2 : for photorealism NSFW I compiled this
https://perchance.org/fusion-t2i-nsfw-video-still-caption-group-008-1
Compiled from 15K prompts here: https://huggingface.co/datasets/thisnick/nsfw-video-still-caption
Url generator : https://perchance.org/fusion-t2i-essence-lodestone-generator-names
Code for creating the generators https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/sort_parquets_by_category.ipynb
Cheers