Andrei Bursuc

We're releasing Franca: a new fully open-sourced (data, code, weights, logs) vision foundation model that finally matches & sometimes outperforms DINOv2, SigLIPv2 & CLIP on ViT-G. This is the fruit of a fun collaboration btwn @valeoai & @FunAILab spearheaded by @shawshank_v 👇

You know you're on vacation when you post more on Strava than you do on X

Home

Franca Nested Matryoshka Clustering for Scalable Visual Representation Learning

Can open-data models beat DINOv2? Today we release Franca, a fully open-sourced vision foundation model. Franca with ViT-G backbone matches (and often beats) proprietary models like SigLIPv2, CLIP, DINOv2 on various benchmarks setting a new standard for open-source research🧵

If you would like to work on open foundation models and datasets necessary for their creation and cooperate with LAION @laion_ai - here opportunity to have your own lab and work on largest supercomputers in EU together with LAION, doing open-source ML/AI fz-juelich.de/en/careers/job…

Cleaning up em dashes is the new formatting references nitpicking activity of senior authors before submission

Video recordings from our workshop on Embodied Intelligence and tutorial on Robotics 101 @CVPR are now up, just in time to catch up with things over the summer. Enjoy! #CVPR2025

We propose new scaling laws that predict the optimal data mixture, for pretraining LLMs, native multimodal models and large vision encoders ! Only running small-scale experiments is needed, and we can then extrapolate to large-scale ones. These laws allow 1/n 🧵

While scaling laws typically predict the final loss, we show in our ICML oral paper that good scaling rules enable accurate predictions of entire loss curves of larger models from smaller ones! w/@Locchiu, @andrewgwils, J. Pennington, A. Agarwala: arxiv.org/abs/2507.02119 1/10

Have you all managed to watch all the #CVPR2025 workshop videos in your list so far? Asking for a friend 🙃

Vive la France! 🇫🇷 This incredible country always found ways to give great gifts to the world across generations: science, culture, diplomacy, etc., even AI nowadays. Historically, it's been mostly a great host for foreigners (myself included). Ever thrilled to be part of France!

Nice! They take the SigLIP encoder out of Gemma3 and further fine tune it on large amount of medical images to get MedSigLIP. This should be a solid image encoder for medical purposes anyone dealing with such images should give a try! (cc @iScienceLuvr) Why take SigLIP out of…

🚨 Did you know that small-batch vanilla SGD without momentum (i.e. the first optimizer you learn about in intro ML) is virtually as fast as AdamW for LLM pretraining on a per-FLOP basis? 📜 1/n