---
license: cc-by-4.0
pretty_name: "LAION-Natural Embeddings"
task_categories:
  - feature-extraction
  - image-classification
  - zero-shot-classification
  - image-to-image
language:
  - en
tags:
  - laion
  - laion-natural
  - laion-2b
  - relaion
  - embeddings
  - clip
  - image-embeddings
  - natural-images
  - image-retrieval
  - photograph-detection
  - ccn2025
  - visual-neuroscience
size_categories:
  - 100M<n<1B
---

<p align="center">
  <img src="photorealistic_earth.jpg" alt="LAION-Natural Embeddings" width="150"/>
</p>

# LAION-Natural Embeddings: CLIP ViT-H/14 Features for ~500M Natural Photographs (CCN 2025, Roth & Hebart)

**LAION-Natural Embeddings** provides pre-computed **CLIP ViT-H/14 embeddings** for ~500 million natural photographs from ReLAION-2B, filtered using the [LAION-Natural](https://huggingface.co/datasets/andropar/relaion2b-natural) naturalness classifier (score > 0.7).

> **Also known as:** LAION-Natural Embeddings · ReLAION-Natural Embeddings · LAION-2B-Natural Embeddings
>
> Part of the LAION-Natural dataset family, introduced in: [*How to sample the world for understanding the visual system*](https://openreview.net/forum?id=T9k6KkZoca) (Roth & Hebart, CCN 2025)

## Quick Start

```python
from datasets import load_dataset
import numpy as np

# Load with streaming (recommended for this size)
ds = load_dataset("andropar/relaion2b-natural-embeddings", streaming=True)

for row in ds['train']:
    url = row['url']
    embedding = np.array(row['embedding'])  # 768-dim CLIP embedding
    score = row['natural_score']  # 0.7 - 1.0
    break
```

## Overview

| | |
|---|---|
| **Total embeddings** | ~514 million |
| **Embedding model** | [CLIP ViT-H/14](https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K) |
| **Embedding dimensions** | 768 |
| **Natural score range** | 0.7 - 1.0 (filtered for natural photographs) |
| **Format** | Parquet (Snappy compressed) |
| **Total size** | ~711 GB |

## Example Images

All images in this dataset have natural scores > 0.7. Examples from different score ranges:

**Natural score 0.7 - 0.8:**
![Examples with scores 0.7-0.8](examples/natural_0.7-0.8.png)

**Natural score 0.8 - 0.9:**
![Examples with scores 0.8-0.9](examples/natural_0.8-0.9.png)

**Natural score 0.9 - 1.0:**
![Examples with scores 0.9-1.0](examples/natural_0.9-1.0.png)

*Thumbnails shown solely to illustrate dataset characteristics. Source: ReLAION-2B-en-research-safe (Apache 2.0). Underlying images remain under the copyright of their original creators.*

## Dataset Structure

| Column | Type | Description |
|--------|------|-------------|
| `url` | string | Image URL from ReLAION-2B |
| `natural_score` | float32 | Naturalness prediction (0.7 - 1.0) |
| `feature_row_id` | int64 | Row index in original LAION-2B embeddings |
| `embedding` | float32[768] | CLIP ViT-H/14 image embedding |

Files are named `relaion2b_features_*.parquet` (2,298 files total).

## Usage Examples

**Load embeddings with PyArrow:**
```python
import pyarrow.parquet as pq
import numpy as np

# Load a single file
table = pq.read_table("data/relaion2b_features_00000.parquet")

# Convert to numpy
embeddings = np.array([e.as_py() for e in table['embedding']])
urls = table['url'].to_pylist()
scores = table['natural_score'].to_pylist()

print(f"Loaded {len(embeddings)} embeddings, shape: {embeddings.shape}")
```

**Image similarity search:**
```python
import numpy as np
from sklearn.metrics.pairwise import cosine_similarity

# Assuming you have a query embedding from CLIP
query_embedding = ...  # Your 768-dim CLIP embedding

# Find most similar images
similarities = cosine_similarity([query_embedding], embeddings)[0]
top_k = np.argsort(similarities)[-10:][::-1]

for idx in top_k:
    print(f"Score: {similarities[idx]:.3f}, URL: {urls[idx]}")
```

**Build a FAISS index for fast search:**
```python
import faiss
import numpy as np

# Load embeddings from multiple files
all_embeddings = []
for i in range(10):  # First 10 files
    table = pq.read_table(f"data/relaion2b_features_{i:05d}.parquet")
    embs = np.array([e.as_py() for e in table['embedding']], dtype=np.float32)
    all_embeddings.append(embs)

embeddings = np.vstack(all_embeddings)

# Build index
index = faiss.IndexFlatIP(768)  # Inner product (for normalized vectors)
faiss.normalize_L2(embeddings)
index.add(embeddings)

# Search
query = np.random.randn(1, 768).astype(np.float32)
faiss.normalize_L2(query)
distances, indices = index.search(query, k=10)
```

**Stream with HuggingFace datasets:**
```python
from datasets import load_dataset

ds = load_dataset("andropar/relaion2b-natural-embeddings", streaming=True)

# Process in batches
batch = []
for i, row in enumerate(ds['train']):
    batch.append(row['embedding'])
    if len(batch) >= 1000:
        # Process batch
        embeddings = np.array(batch)
        # ... your processing ...
        batch = []
    if i >= 10000:
        break
```

## Use Cases

- **Image similarity search:** Find visually similar images in a large corpus
- **Zero-shot classification:** Use CLIP embeddings for classification without training
- **Clustering:** Discover visual themes and patterns in natural photographs
- **Training data:** Use as features for downstream models
- **Research:** Study visual representations at scale

## Related Datasets

- **[LAION-Natural (scores)](https://huggingface.co/datasets/andropar/relaion2b-natural)** — Natural scores for all 2.1B ReLAION-2B images, plus the trained naturalness classifier with portable weights and evaluation diagnostics. Use this to get URLs for images with any score threshold.
- **[LAION-Natural (alias)](https://huggingface.co/datasets/andropar/laion-natural)** — Alias repository for discoverability

## Attribution

- **Source dataset:** [ReLAION-2B-en-research-safe](https://huggingface.co/datasets/laion/relaion2B-en-research-safe) by LAION
- **Embedding model:** [CLIP ViT-H/14](https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K) trained on LAION-2B

## Licensing / Content

This repository contains only **metadata and derived features** (URLs, natural scores, embeddings). No images are distributed.

- The underlying images are hosted by third-party websites and remain under their original copyrights and terms of use.
- Our additions (embeddings, documentation) are released under **CC-BY 4.0**.
- This dataset is based on [ReLAION-2B-en-research-safe](https://huggingface.co/datasets/laion/relaion2B-en-research-safe), which is licensed under **Apache 2.0**.
- Please check license compatibility for any commercial usage.

## Citation

```bibtex
@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}
}
```

## Limitations

- Embeddings are from CLIP ViT-H/14, which was trained on web data and may reflect its biases
- "Naturalness" is based on a learned classifier - see [relaion2b-natural](https://huggingface.co/datasets/andropar/relaion2b-natural) for methodology
- Some URLs may be broken or point to different/removed images over time
- Intended for research use

---

Questions or issues? Open a discussion!