LAION-Natural

This is an alias repository for discoverability. The actual data lives in the repositories below.

LAION-Natural is a large-scale naturalness scoring and filtering dataset for ReLAION-2B, introduced in How to sample the world for understanding the visual system (Roth & Hebart, CCN 2025).

Also known as: LAION-Natural · LAION natural · ReLAION-Natural · ReLAION-2B-Natural · LAION-2B-Natural

Datasets

Dataset	Description	Size
LAION-Natural (scores)	Naturalness scores for 2.1B images from ReLAION-2B, plus the trained classifier	~2.1B rows
LAION-Natural Embeddings	CLIP ViT-H/14 embeddings for ~500M natural photographs (score > 0.7)	~514M embeddings, 711 GB

What is LAION-Natural?

• Naturalness scores for 2.1 billion web images, predicting how "natural" or "photographic" each image looks versus artificial/rendered content
• ~500M natural photographs identified at the recommended threshold of 0.7
• 26k human labels collected via active learning to train the classifier
• Classifier included (logistic regression on CLIP ViT-L/14 features, ROC AUC 0.89)

Use Cases

• Filter non-photographic content from web-scraped image datasets
• Select natural image stimuli for vision research and cognitive science
• Train vision models on curated natural photographs
• Image similarity search over 500M natural photographs using pre-computed CLIP embeddings

Citation

@inproceedings{
  roth2025how,
  title={How to sample the world for understanding the visual system},
  author={Johannes Roth and Martin N Hebart},
  booktitle={8th Annual Conference on Cognitive Computational Neuroscience},
  year={2025},
  url={https://openreview.net/forum?id=T9k6KkZoca}
}