andropar/relaion2b-natural

relaion2B-en-research-safe natural scores

This dataset contains naturalness scores for images in RELAION2B-en-safe. The scores predict how "natural" or "photographic" an image looks (vs artificial/rendered content). Scores range from 0-1, where higher values = more natural.

Quick stats:

• ~2M+ image URLs from RELAION2B
• Only URLs that were also in LAION-2B-en have scores
• File format: Snappy-compressed Parquet

Dataset Structure

Column	Description
url	Image URL from RELAION2B
natural_score	Naturalness prediction (0-1), null if no match

Files are named relaion2b_natural_part-*.snappy.parquet and match the original relaion2b-en-safe dataset.

Dataset creation

We first obtained a small set of natural and non-natural images, by manually labeling 200k images from LAION-2B-en in an active learning loop. Selection criteria for natural images were:

• No watermarks, logos, or banners in the image
• No heavy editing (black-and-white filters, high contrast or saturation, etc., photoshopped images)
• Real-world scene or object

We trained a logistic regression model on CLIP ViT-L/14 features (768-dim) of the images and applied the model to image embeddings of the original LAION-2B-en dataset. As this dataset can't be shared any longer, we finally matched URLs between it and the new relaion2b-en-safe dataset.

Usage

Load with pandas:

import pandas as pd
df = pd.read_parquet("relaion2b_natural_part-000.snappy.parquet")

# Get natural images only
natural = df[df['natural_score'] > 0.7]

Load everything:

import glob
files = glob.glob("relaion2b_natural_part-*.snappy.parquet") 
df_all = pd.concat([pd.read_parquet(f) for f in files])

With Hugging Face datasets:

from datasets import load_dataset
dataset = load_dataset("your-username/relaion2b-natural")

Use cases

• Filtering image datasets for natural/photographic content
• Quality assessment for computer vision training data
• Research on image naturalness
• Preprocessing step before training vision models

Limitations

• "Naturalness" is based on our specific training data - might not match your definition
• These are ML predictions, not ground truth
• Some URLs might be broken or point to different images now, but we can't check for that

Citation

TBD.

---

Questions? Issues? Let me know in the discussions!