Skip to content

Dataset · KBlueLeaf

7.4M anime art captions from Florence-2

KBlueLeaf/danbooru2023-florence2-caption

Large parquet-format corpus with detailed text captions of Danbooru2023 images for training image generators.

What's inside

The dataset contains 7,438,449 parquet records linking Danbooru2023 image IDs to detailed English captions generated by the microsoft/Florence-2-large model with the MORE_DETAILED_CAPTION tag. Average caption length is around 108–120 tokens (depending on the Flan T5, DFN CLIP, or Qwen2 tokenizer). Image files themselves are not included—only IDs and text.

What you can do with it

  • Training and fine-tuning text-to-image models and LoRA on long anime illustration descriptions
  • Training image-to-text models for automatic art tagging and captioning
Tasks
Text → image, Image → text, Text generation
Languages
English
Size
10M–100M rows
Downloads this month
144
Likes
66

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("KBlueLeaf/danbooru2023-florence2-caption")