Skip to content

Dataset · zzliang

GRIT: region-grounded image-text pairs

zzliang/GRIT

Large-scale corpus of image-text pairs where caption phrases are grounded to corresponding image regions.

What's inside

GRIT is a large-scale Grounded Image-Text pairs corpus built on image-text pairs from COYO-700M and LAION-2B. It extracts text fragments (noun phrases and referring expressions) and links them to image regions. Size is 10M to 100M examples; modalities are images, tabular data, and text, language is English.

What you can do with it

  • Train a text-to-image model with phrase-to-region grounding
  • Fine-tune an object detector using text descriptions
  • Evaluate visual grounding and zero-shot classification
Tasks
Text → image, Image → text, Object detection, Zero shot classification
Languages
English
Size
10M–100M rows
Downloads this month
8.4K
Likes
170

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("zzliang/GRIT")