Skip to content

Dataset · zai-org

137k human image generation quality comparison pairs

zai-org/ImageRewardDB

English expert preference dataset for training image generation reward models.

What's inside

The dataset contains 137,000 comparison pairs annotated for images generated from text prompts based on DiffusionDB. The creators developed a special pipeline with annotator training and quality control to capture human preferences in the text-to-image task.

What you can do with it

  • Training reward models to evaluate aesthetics and prompt-following accuracy
  • Aligning (RLHF) diffusion models with human preferences
Tasks
Text → image
Languages
English
Size
100K–1M rows
Downloads this month
12.4K
Likes
48

How to load it

from datasets import load_dataset

ds = load_dataset("zai-org/ImageRewardDB")