Mathurin Massias

New paper on the generalization of Flow Matching arxiv.org/abs/2506.03719 🤯 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 👇👇👇

🚨 🔬 PhD positions at Google DeepMind in France 🇫🇷 We are advertising Master Level Intern positions at Google DeepMind within our Frontier AI Unit. These could lead to co-advised PhD positions with Google DeepMind and French academic institutions. job-boards.greenhouse.io/deepmind/jobs/…

With @Qu3ntinB we have one offer for a Marie Sklodowska-Curie postdoctoral fellowships at Inria, to work on generative models : inria.fr/en/marie-sklod… contact me if interested! RT appreciated <3

There is an Associate Professor position in CS at ENS Lyon, with potential integration in my team, starting in sept 2026 : DM me in interested! Details at ens-lyon.fr/LIP/images/Pro…

To understand these phenomena, we study the spatial regularity of the velocity/denoiser over time: we observe a gap between the closed-form and trained model. Applying Jacobian spectral regularization, we recover effects seen previously on perturbed denoisers (drift vs noise).

Different loss weightings favor different times: which temporal regime drives the generation quality? Controlled perturbations reveal: drift type effects at early times (& good FID) and noise type at late times (& bad FID)

🌀New paper on the generation phases of Flow Matching arxiv.org/abs/2510.24830 Are FM & diffusion models nothing else than denoisers trained at every noise level? In theory yes, *if trained optimally*. But in practice, do all noise level matter equally?

Strong afternoon session! Ségolène Martin on how to go from flow matching to denoisers (and hopefully come back?) and Claire Boyer on how learning rate and working in latent spaces affect diffusion model

Followed by Scott Pesme on how to use diffusion/flow matching based MMSE to compute a MAP (and nice examples!), and Thibaut Issenhuth on new ways to learn consistency models

Kickstarting our workshop on Flow matching and Diffusion with a talk by Eric Vanden Eijnden on how to optimize learning and sampling in Stochastic Interpolants ! Broadcast available at gdr-iasis.cnrs.fr/reunions/model…

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 gdr-iasis.cnrs.fr/reunions/model…

This is happening today, 6pm Paris time!

Some 📸🤳from the ongoing #MlssSenegal2025 🙌🏿

@max_tensor @Qu3ntinB You don't even need to train a model to see that the target is, on real data, not stochastic as soon as t is not very small See the animation in second tweet & figures 1b and 2 in arxiv.org/pdf/2506.03719, which do not involve training a model

@mathusmassias @Qu3ntinB Does CFM really bury stochastic targets or did it just test a tiny U-Net on CIFAR-10 & CelebA-64? Run it on ImageNet-256 with augment ON & non-Gaussian noise before declaring noise “dead.”

New paper on the generalization of Flow Matching arxiv.org/abs/2506.03719 🤯 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 👇👇👇

@danijarh @Qu3ntinB The closed-form minimizer u* is singular at t=1 if x is not a training point, so this reasoning does not apply: following u*, you exactly generate training points (up to a finite set). See e.g. Thm 4.6 in arxiv.org/abs/2410.23594

Interesting! Since the flow is an ODE (cont. in time and space), paths can't branch or merge. Doesn't that imply that for every two generated datapoints, it's possible to generate one in between. And so the model can't purely memorize but needs to interpolate at least a bit? Curious to understand your intuition better

@DaniloJRezende See eg this potential explanation which was all the rage a few weeks ago x.com/johnjvastola/s…

@DaniloJRezende absolutely! Our point is more 1) that the velocity has a closed-form (already observed by others) 2) that out of several hypotheses explaining why models don't learn it we rule out "target stochasticity" (see second tweet)

@giladturok @Qu3ntinB As you can see in the first gif, following the closed-form minimizer, different x_0 get mapped to the same x_1 (which is one of the training point)

@mathusmassias @Qu3ntinB Naive q: In CFM we regress a neural velocity field to a known, true velocity field. The true velocity field at time t is computed with sample x_0 from a base dist and x_1 from a target dist. At inference time, if we sample a diff x_0, shouldn’t we automatically get a diff x_1?