ReLAION-2B Natural: Naturalness Scores for 2.1 Billion Images

ReLAION-2B Natural contains naturalness scores for 2.1 billion images from the ReLAION-2B-en-research-safe collection, released by andropar in 2026. These scores differentiate between photographic content and artificial or rendered images to facilitate large-scale image filtering.

Use Cases

• Filtering synthetic or rendered content from training pipelines using the natural_score column
• Statistical analysis of photographic versus artificial content distribution in the LAION-2B-en collection
• Curating high-fidelity photographic subsets for generative model fine-tuning

Strengths

• 2.1 billion records
• Parquet format for efficient querying
• Direct compatibility with ReLAION-2B-en-research-safe identifiers

Limitations

• Scores are model-predicted rather than human-annotated
• Restricted to the research-safe English subset of LAION
• Requires joining with external image data to be actionable

Provenance

Source

andropar via Hugging Face

Collection Method

model-based prediction

Freshness

Last updated March 2026.

Geography

US region