Synthetic data for CLIP debiasing

About

SynthFairCLIP is a research initiative focused on fair vision–language models.

We study how to reduce bias in CLIP-style models by combining:

• Real data from large-scale datasets such as DataComp/CommonPool.
• Synthetic data generated with state-of-the-art diffusion models.
• Curation and balancing of demographic attributes across professions, activities and contexts.

What we release

• CLIP models trained on hybrid real–synthetic data.
• Large-scale WebDataset shards of synthetic / hybrid image–text data.
• Eval tools and benchmarks for analysing bias and fairness in CLIP-like models.

If you use our resources, please consider citing the SynthFairCLIP project.

Acknowledgement

We acknowledge EuroHPC JU for awarding the project ID EHPC-AI-2024A02-040 access to MareNostrum 5 hosted at BSC-CNS.