New paper on the generalization of Flow Matching https://t.co/BJMHUnY6xJ 🤯 Why does flow matching generalize? Did you know that the flow matching target you're trying to learn **can only generate training points**? with @Qu3ntinB, Anne Gagneux & Rémi Emonet 👇👇👇

I am thrilled to announce that our work on the generalization of flow matching has been accepted to NeurIPS as an oral!! See you in San Diego 😎

One-day workshop on Diffusion models and Flow matching, October 24th at ENS Lyon Registration and call for contributions (short talk and poster) are open at

This is happening today, 6pm Paris time!

Some 📸🤳from the ongoing #MlssSenegal2025 🙌🏿

Then why does flow matching generalize?? Because it fails! The inductive bias of the neural network prevents from perfectly learning u* and overfitting In particular networks fail to learn the velocity field for two particular time values See the paper for a finer analysis 😀

We propose to regress directly against the optimal (deterministic) u* and show that it never degrades the performance On the opposite, removing target stochasticity helps generalizing faster.

Yet FM generates new samples! An hypothesis to explain this paradox is target stochasticity: FM targets the conditional velocity field i.e. only a stochastic approximation of the full velocity field u* *We refute this hypothesis*: very early, the approximation almost equals u*

Same here. My solution is I gave "very low" grade to 99% of the papers

On Saturday Anne will also present some very, very cool work on how to leverage Flow Matching models to obtain sota Plug and Play methods: PnP-Flow: Plug-and-Play Image Restoration with Flow Matching, poster #150 in poster session 6, Saturday at 3 pm https://t.co/0GRNZd3l8O

It was received quite enthusiastically here so time to share it again!!! Our #ICLR2025 blog post on Flow M atching was published yesterday : https://t.co/2V5BLl6T2p My PhD student Anne Gagneux will present it tomorrow in ICLR, 👉poster session 4, 3 pm, #549 in Hall 3/2B 👈

MLSS is coming to Senegal 🇸🇳 in 2025! 🌍 📍 AIMS Mbour, Senegal 📅 June 23 - July 4, 2025 An international summer school to explore, collaborate, and deepen your understanding of machine learning in a unique and welcoming environment. Details:

For people who could not attend the #NeurIPS2024 tutorial on flow-matching, here is a friendly introduction! https://t.co/IJIO2Bzo8X

The optimization loss in FM is easy to evaluate, and does not require integration like in CNF. The whole process is smooth! The illustrations are much nicer in the blog post, go read it ! 👉👉 https://t.co/TSkg1VZ5cn 👈👈

FM learns a vector field u pushing the base distrib to the target through an ODE To learn it, introduce a conditioning random variable, breaking the pb into smaller ones that have closed form solutions Here's the magic: the small problems can be used to solve the original one!

FM is a technique to train continuous normalizing flows (CNF) that progressively transform a simple base distrib to the target one 2 benefits -no need for likelihoods nor ODE in training -makes the pb better posed by defining a *unique sequence of densities* from base to target

Anne Gagneux, Ségolène Martin, @qu3ntinb, Remi Emonet and I wrote a tutorial blog post on flow matching: https://t.co/TSkg1VZ5cn with lots of illustrations and intuition! We got this idea after their cool work on improving Plug and Play with FM: https://t.co/0GRNZd3l8O

Looks like the last reason to stay here won't remain valid for long !

btw this is also my first post on 🦋 with the handle mathurinmassias 😀

New blog post: the Hutchinson trace estimator, or how to evaluate divergence/Jacobian trace cheaply. Fundamental for Continuous Normalizing Flows https://t.co/W94Q08QIRx