Laion Aesthetics 12M UMAP
Dataset containing UMAP embeddings of 12 million images from the Laion-aesthetics corpus, filtered for high aesthetic quality. Each point represents a 2D projection of the CLIP embeddings according to various parameters.
Approximately 12 million images with UMAP embeddings, 2.74 GB in Parquet format
MIT
Description
Laion Aesthetics 12M UMAP is a dataset of 2D embeddings generated by UMAP using CLIP vectors from an aesthetic subset of the LAION corpus. It includes 3 projections with different values of the n_neighbors parameter, offering multiple perspectives for analysis and visualization.
What is this dataset for?
- Visualization and clustering of images according to their aesthetic quality
- Exploring large image sets by similarity
- Image research and creation of recommendation systems based on aesthetics
Can it be enriched or improved?
This dataset can be enriched by integrating additional metadata or by combining with other embeddings to improve the granularity of analyses. Human or automatic annotations on aesthetics could also complete the corpus.
🔎 In summary
🧠 Recommended for
- Computer vision researchers
- Data scientists
- Recommendation system developers
🔧 Compatible tools
- UMAP
- Scikit-learn
- Pandas
- Hugging Face Datasets
💡 Tip
Test various n_neighbors settings to optimize the visualization according to your needs.
Frequently Asked Questions
What is the UMAP in this dataset?
UMAP is a dimensional reduction technique that projects CLIP embeddings in 2D to facilitate visualization and clustering.
How many different projections are available?
Three projections with n_neighbors equal to 10, 30 and 60, offering several perspectives for analysis.
Does this dataset contain the original images?
No, it only contains the UMAP embeddings of the images, not the images themselves.
