Here's the tale of how @jder.bsky.social and I scaled Samudra, a neural ocean emulator capable of predicting 8 years of the ocean on a single GPU, to operate at a full 1/4° resolution (16x the size in bytes). It was quite a humbling process.

Open Athena@openathena.ai · 6d ago

Simulating ocean climate takes a supercomputer 4,600+ CPU cores to produce 12 simulated years per day (SYPD). Samudra 2 produces 4,800 SYPD on 1 GPU at the same resolution. In a new blog, @al.merose.com reports on Samudra, a neural ocean emulator built in collaboration with NYU & MIT: bit.ly/oa-ss

NeurIPS submissions confirmed to be a heat-loving species. Warmer year, bigger bloom. Every degree we add, the deadline gets denser 🌻 Good news for the field, we're having a really good growing season 👨‍🌾

You can train your image-to-image flow matching model on badly aligned data, you just have to tell it how bad it is. 🫣 Great work from @lebellig.bsky.social and Aimi Okabayashi with cool applications to remote sensing. It's FlowEO 2.0!

lebellig@lebellig.bsky.social · 2w ago

The last project of my PhD is finally out! 🪴 It was a pleasure collaborating with Aimi on this work! We introduce A²BM: Alignment-Aware Bridge Matching, a new framework for image-to-image translation with weakly aligned image pairs. Paper 📄: arxiv.org/pdf/2607.16294

The last project of my PhD is finally out! 🪴 It was a pleasure collaborating with Aimi on this work! We introduce A²BM: Alignment-Aware Bridge Matching, a new framework for image-to-image translation with weakly aligned image pairs. Paper 📄: arxiv.org/pdf/2607.16294

I organize a 1 day workshop on Generative modelling @ENS Lyon, October 9th Call for oral/poster contributions is open; details at gdr-iasis.cnrs.fr/reunions/mod...

I'd like to announce that at @openathena.ai, @jder.bsky.social and I helped @m2lines.bsky.social release Samudra 2. We scaled this neural ocean emulator to train on 16x the size of data in bytes on the same hardware budget. We can now skillfully predict 8 years of the ocean on a single GPU at a 1/4°

🌊 Samudra 2: A Fast, Cheap AI Ocean Model, Now at the Scale That Matters

M²LInES’ neural ocean emulator now runs multi-year simulations at eddy-permitting resolution on a single GPU, turning a supercomputer-scale…

medium.com

Slides for our ICML tutorial on Memorization and Generalization of Diffusion and Flow Matching Models are now available ! 🌀 memorization-generalization.github.io @quentinbertrand.bsky.social

Heading to #ICML2026 🇰🇷 and interested in diffusion models, flow matching, and their generalization capabilities? Don't miss the excellent tutorial by @mathurinmassias.bsky.social and @quentinbertrand.bsky.social on Monday! 📍 Hall D1 🗓️ Monday, July 6 🕘 9:00–11:30 AM Details: icml.cc/virtual/2026...

Heading to #ICML2026 🇰🇷 and interested in diffusion models, flow matching, and their generalization capabilities? Don't miss the excellent tutorial by @mathurinmassias.bsky.social and @quentinbertrand.bsky.social on Monday! 📍 Hall D1 🗓️ Monday, July 6 🕘 9:00–11:30 AM Details: icml.cc/virtual/2026...

🛰️ Introducing UniverSat: one transformer backbone for Earth Observation that handles ANY sensor, ANY spatial, spectral & temporal resolution, ANY scale — with a single set of weights. 🌍

I took Lyon's rainy days in early May as a warm welcome, now with the heatwave I think my acclimatisation is complete, so it's time to make it official! I've started a postdoc on generative models in Inria's Ockham team working with @mathurinmassias.bsky.social and @quentinbertrand.bsky.social ☀️

We explored the impact of variability sources in generative modeling. Turns out, we've been neglecting the error bars associated with training variability all along! We should aim to report results that we are sure of their scientific validity, instead of seed engineering!

Kyutai@kyutai-labs.bsky.social · 2mo ago

🎰 Welcome to the FID Lottery. We pulled the lever 25 times on the same machine. Identical diffusion model, identical ImageNet class-cond recipe, only the seed changed. The house paid out anywhere from 33.59 to 35.69 FID. A 2.1-point spread, pure luck. Step onto the floor 👇🧵

one anecdote: I was searching for a link to one of my old papers on Google, and the automated Gemini summary attributed all of my lab's work to my husband

Mel Andrews@bayesianboy.bsky.social · 2mo ago

When utilized in literature review, LLMs consistently 1. fail to mention female authors in female-led literatures, 2. insist that men are more influential or more heavily cited when this is contradicted by objective citation counts, and 3. attribute women’s work to hallucinated male scholars.

The alpha version of my new book "Optimal Transport for Machine Learners" is out, with in particular an online version with interactive figures www.gpeyre.com/ot4ml/

🎆 New paper! "Random Process Flow Matching: Generative Implicit Representations of Multivariate Random Fields", by Julien Lalanne, accepted to ICML'26 🥳 We're proposing flow-matching for inpainting in ultra-sparse setup, with applications to seismic interpolation. 📜 arxiv.org/abs/2605.28625 1/

Random Process Flow Matching: Generative Implicit Representations of Multivariate Random Fields

Generative modeling provides a powerful framework for learning data distributions. These models initially relied on probabilistic methods such as Gaussian Processes (GP) for uncertainty-aware predicti...

arxiv.org

"Accept (spotlight)" at ICML'26 😎 Our paper brings particle filters back to life: autoregressive diffusion models + posterior sampling yield optimal proposals for Bayesian filtering, scaling up to GenCast-sized systems. arxiv.org/abs/2605.20028 w/ Thomas Savary and @francois-rozet.bsky.social

Training-Free Bayesian Filtering with Generative Emulators

Bayesian filtering is a well-known problem that aims to estimate plausible states of a dynamical system from observations. Among existing approaches to solve this problem, particle filters are theoret...

arxiv.org

ICML Conference@icmlconf.bsky.social · 3mo ago

Congrats again to authors of accepted #ICML2026 papers! The camera-ready deadline is 5/28. Drawing your attention to two specific features: 1. As last year, to help communicate research to a broad audience, papers will have lay summaries. Tips & details in blog 1/3

Thrilled to share that MIRO is accepted to ICML 2026 @icmlconf.bsky.social ! 🎉 By training on the reward scores, we can simply condition the model on high rewards at inference time to guarantee top-tier, aligned outputs. We’ve updated our paper with some additional results!

MIRO: MultI-Reward cOnditioned pretraining improves T2I quality and efficiency

The default paradigm of post-training text-to-image generators includes post-hoc selection of generated images, and subsequent training with one reward model to align the generator to the reward, typi...

arxiv.org

Nicolas Dufour@nicolasdufour.bsky.social · 9mo ago

We introduce MIRO: a new paradigm for T2I model alignment integrating reward conditioning into pretraining, eliminating the need for separate fine-tuning/RL stages. This single-stage approach offers unprecedented efficiency and control. - 19x faster convergence ⚡ - 370x less FLOPS than FLUX-dev 📉

📢 The TerraBytes workshop is returning for a 2nd edition - this time at ECCV 2026. Submit your paper before June 18th and join in Malmö, Sweden! 🔗 terrabytes-workshop.github.io

👏 Folks! If you are curious about the Generative Modeling via Drifting paper, but you find it difficult to understand → I wrote a different interpretation of it. It's called: "An Expectation-Maximization interpretation of Generative Modeling via Drifting" davidpicard.github.io/pdf/An_Expec...

My PhD thesis manuscript will be available in the coming months, but I’ve written two blog posts based on the related work chapter: 1. Generative modeling with flow-based models 🪚 2. Data-translation with flow and diffusion bridges 🔨 Open to feedback and discussions! lebellig.github.io/blog/

Want better local models? Want models that are made completely transparently? OpenAthena.ai is hiring on our Marin.community team. We’re especially looking for SWEs and research scientists with experience in data and post training. DM me.

My first blog post in over a year is a deep dive on flow maps🗺️, or how to learn the integral of a diffusion model to enable faster sampling and several other cool tricks. It's the longest one yet👀 Let me know what you think! sander.ai/2026/05/06/f...

Learning the integral of a diffusion model

A deep dive on flow maps.

sander.ai

My PhD thesis manuscript will be available in the coming months, but I’ve written two blog posts based on the related work chapter: 1. Generative modeling with flow-based models 🪚 2. Data-translation with flow and diffusion bridges 🔨 Open to feedback and discussions! lebellig.github.io/blog/

Delighted to have successfully defended my PhD thesis on "Generative models for Earth Observation, from denoising to domain adaptation" 🪴 This wouldn’t have been possible without the support of my colleagues, my PhD advisor, and my family and friends. Thank you all!

While it is intuitively clear that straighter trajectories should reduce discretization error when integrating an ODE (for instance, in flow matching), I could not find a precise bound. I therefore rewrote the proof of Cauchy-Lipschitz to make this explicit. github.com/gpeyre/Discr...

I wrote a short mathematical companion tutorial to my notebook on discrete diffusion models. It gives an informal derivation of the connection between maximum likelihood estimation of the backward transition kernel and denoising score matching. github.com/gpeyre/Discr...

What if AI could invent enzymes that nature hasn’t seen? 👩‍🔬🧑‍🔬 Introducing 🪩 DISCO: Diffusion for Sequence-structure CO-design 📝 Blog: disco-design.github.io 📄 Paper: arxiv.org/abs/2604.05181 💻 Code: github.com/DISCO-design...

Two research blog posts that look really interesting 👀 "Teaching AI to Invent Enzymes Nature Never Imagined", DISCO: Diffusion for Sequence-structure CO-design disco-design.github.io "How to Generate Text in One Step", Flow Map Language Models one-step-lm.github.io/blog/

🚨 arxiv.org/abs/2604.06129 PoM: A Linear-Time Replacement for Attention with the Polynomial Mixer This paper is the result of doing a lab-wide hackathon on an idea I've had for some time. Probably the paper with the highest number of authors I've ever done. It's a CVPR Findings 26. Thread 🧵👇

PoM: A Linear-Time Replacement for Attention with the Polynomial Mixer

This paper introduces the Polynomial Mixer (PoM), a novel token mixing mechanism with linear complexity that serves as a drop-in replacement for self-attention. PoM aggregates input tokens into a comp...

arxiv.org

For those interested in normalized gradient methods and optimal transport: I introduce a new class of "spectral" Wasserstein distances for which spectrally normalized gradient descent (Muon but without momentum and small step size ...) is a spectral-W gradient flow: arxiv.org/abs/2604.04891

Muon Dynamics as a Spectral Wasserstein Flow

Gradient normalization is central in deep-learning optimization because it stabilizes training and reduces sensitivity to scale. For deep architectures, parameters are naturally grouped into matrices ...

arxiv.org