Human Face Image-Text Pairs from LAION-400M

LAION-Face is a subset of the LAION-400M dataset containing 50 million image-text pairs identified as containing human faces. It was first used to train the FaRL model for face analysis tasks. A 20 million image subset is also provided for faster evaluation.

Use Cases

• Train vision-language models like FaRL on 50 million image-text pairs for face analysis tasks.
• Fine-tune face detection models using the image-text pairs identified as containing faces.
• Use the 20 million image subset for rapid evaluation and prototyping of face-related models.

Strengths

• Contains 50 million image-text pairs, providing a large-scale resource.
• Derived from the LAION-400M dataset, a known large-scale web crawl.
• Includes a curated 20 million pair subset for faster evaluation.

Limitations

• Dataset size and specific image characteristics are unknown.
• Potential for web-crawled data biases and label noise is present.
• No column or feature-level metadata is provided for detailed analysis.

Provenance

Source

LAION-400M dataset

Collection Method

Face detection applied to a web-crawled image-text dataset.

Freshness

2022-11-18