🧑‍🔬 Mehdi Cherti

Mehdi Cherti works on large-scale machine learning for Vision & language, especially models that learn jointly from images and their accompanying text. The papers here share a common question: how open, web-scale data and openly released code can make research on contrastive image-text models possible outside a small number of industry labs. The methods are empirical and computational. They involve collecting and filtering very large image-text corpora, training CLIP-style models with the open-source OpenCLIP codebase, and measuring how performance changes as models and data grow.

The most cited of Cherti's papers here is LAION-5B (2022). It describes a public collection of billions of image-text pairs gathered from Common Crawl using images and their HTML alt-text. Pairs were kept only when an existing CLIP model judged image and caption to be sufficiently similar, and OpenCLIP models trained on a subset showed that the data was usable. The companion paper, Reproducible scaling laws for contrastive language-image learning (2022), builds on that resource. It trains CLIP models on public LAION subsets while varying the size of the vision encoder, the amount of data and the number of samples seen, which yields a scaling study anyone can reproduce. Both works date from 2022 and together represent the most recent direction in the papers here: open data paired with transparent measurement of scaling behaviour.

Bylines from 2021 to 2025 place Cherti at Forschungszentrum Jülich and its Jülich Supercomputing Centre. Earlier papers came from the Laboratoire de l'Accélérateur Linéaire in Orsay and the Laboratoire de Recherche en Informatique in Gif-sur-Yvette. Frequent collaborators include Christoph Schuhmann, Romain Beaumont, Ross Wightman, Mitchell Wortsman and Cade Gordon. Other co-authors include Jenia Jitsev and Ludwig Schmidt.

Mehdi Cherti has 2 papers in our library, in 2022, most of them on Vision & language; often with Cade Gordon, Christoph Schuhmann, Mitchell Wortsman.