Nicolas Dufour

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!

Instead of relying on pretrained video latent space we study which tradeoff of spatial and temporal resolution is optimal for real time WM. We open source all our models and training recipes so you can easily adapt it for your needs!

🚀 New paper: Balancing Frequencies and Pixels in Flow Matching We tackle the low-frequency bias in pixel-space flow matching and train JiT up to 40% faster without any architectural changes. 📄 Read it here: arxiv.org/abs/2609.02748

It's a cool paper, both old school and new school. There's an Easter egg: I designed a game for the teaser. Did you get it right? I tried it in my ML course and almost nobody got it.

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

Highly recommend! A great place to meet people passioned about generative models

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...

Excited to share that our paper on sprite-based image decomposition is accepted at TMLR! 🎉 Sprite-based models are highly interpretable but struggle to scale to complex, multi-object images. 
 1/3

There's a demo of the model btw: huggingface.co/spaces/nicol...

1/ MIRO: MultI-Reward cOnditioned pretraining improves T2I quality and efficiency. Wed, Jul 8, 2026 • 10:30 AM – 12:15 PM KST 📜 arxiv.org/abs/2510.25897 🖥️ nicolas-dufour.github.io/miro/ 🧬 huggingface.co/nicolas-dufo... With @arrijitghosh.bsky.social and @lucasdegeorge.bsky.social being on site

🛰️ 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. 🌍

Dufour et al., "The FID Lottery: Quantifying Hidden Randomness in Generative Model Evaluation" We all know it's expensive to train multiple times, but we are now at a point where it is inevitable. Statistical significance should not be ignored. Don't bold over 1~2% differences.

I'll need to track how many time I refer to this paper. It's probably going to be my new language filler.

Babe, stop everything! New favorite paper of the year is out! kyutai.org/fid-lottery/ arxiv.org/abs/2606.20536

Hypnotizing to watch. Great work, co-authored by our very own @nicolasdufour.bsky.social

𝗦𝘂𝗿𝗳𝗹𝗼: 𝗖𝗼𝗻𝘀𝗶𝘀𝘁𝗲𝗻𝘁 𝟯𝗗 𝗦𝘂𝗿𝗳𝗮𝗰𝗲 𝗙𝗹𝗼𝘄 𝗠𝗼𝗱𝗲𝗹 𝘄𝗶𝘁𝗵 𝗚𝗹𝗼𝗯𝗮𝗹 𝗦𝘁𝗮𝘁𝗲 Antoine Guédon, Shu Nakamura, Nicolas Dufour ... Angjoo Kanazawa arxiv.org/abs/2606.13644 Trending on scholar-inbox.com

What if you could turn any number of photos (3, 8, 15, or even 60) into one clean 3D surface (pts & mesh) with Flow Matching? Check out our new work, Surflo: Consistent 3D Surface Flow Model with Global State. 🧵 1/N 🔗https://anttwo.github.io/surflo/

We will release our code+data asap, please stay tuned! Thanks to my amazing coauthors: @nicolasdufour.bsky.social Shu Nakamura Jiahui Lei @kyotovision.bsky.social @akanazawa.bsky.social 📜arXiv: arxiv.org/abs/2606.13644 🔗Project: anttwo.github.io/surflo/ 💻Code (soon): github.com/Anttwo/Surflo 8/8

Check out our latest work! 🚀 We learn a global state and decode the point cloud pointwise, allowing to decode as many points as you want. Plus, we introduce some clever guidance tricks to ensure global consistency, yielding high-quality meshes from just a few views! 👇

Surflo: Consistent 3D Surface Flow Model with Global State @antoine-guedon.bsky.social, Shu Nakamura, @nicolasdufour.bsky.social, Jiahui Lei, Ko Nishino, @akanazawa.bsky.social arxiv.org/abs/2606.13644

Playing with some post-processing on a small in-house-but-soon-to-be-released model.

The model is so fast and easy to use that I vibe-coded a small game with it in 1h 😅 Runs flawlessly on a consumer GPU if you're looking for a small local model to tinker with.

Open-source fueled the LLM revolution, but Physical AI hasn't fully benefited from this flywheel yet. Today, we're launching kesai.eu, our mission to democratize robotics research! First milestone: training a frontier-level self-driving policy using significantly less data than typically required.

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 📉

👏 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...

Excited to share my work as a Student Researcher at Google Zurich: UniGeoCLIP! 🌍🚀 W/ Eduard Trulls, Jan Hosang, @loicland.bsky.social & @pesarlin.bsky.social , we built a framework aligning 5 geospatial modalities in one space. Presented at EarthVision @ #CVPR2026. 🧵👇

New paper: Back into Plato’s Cave Are vision and language models converging to the same representation of reality? The Platonic Representation Hypothesis says yes. BUT we find the evidence for this is more fragile than it looks. Project page: akoepke.github.io/cave_umwelten/ 1/9

Checkout our recent work, where we only need web images to learn a novel view generation model! We can navigate inside any image, without any video/multi view data or prior models! Congrats to Adrien for this great first PhD paper! (with @davidpicard.eurosky.social and @ptrkprz.bsky.social)

I thought I would do a thread, but honestly the post is so good: kyutai.org/blog/2026-04... It explains "One View Is Enough! Monocular Training for In-the-Wild Novel View Generation" arxiv.org/abs/2603.23488 done in colab with the smart people at kyutai

🚨 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 🧵👇

🚨 Happy to announce CVPR@Paris'26 which will take place on June 1st in Paris. The goal of the event is to share a little bit of the conference before it happens. We will have poster sessions as well as several plenary talks by world-class speakers. info: cvprinparis.github.io/CVPR2026InPa...

arxiv.org/abs/2603.23488 👀 I'll make a detailed thread later

I'm commenting that number on slack with @nicolasdufour.bsky.social and I just realized that if you add the 16k active submissions at CVPR, even considering a sizeable overlap between the 2, there are currently well over 30k active papers in review. That's nuts

It was a big pleasure to be in Nicolas's committee. Congratulations to Nicolas for the great work, and congratulations to the advisors too!

Yesterday, @nicolasdufour.bsky.social defended is PhD. I really enjoyed the years of collaboration w/ @vickykalogeiton.bsky.social (& @loicland.bsky.social) Video: youtube.com/live/DXQ7FZA... Big thanks to the jury @dlarlus.bsky.social @ptrkprz.bsky.social @gtolias.bsky.social A. Efros & T. Karras

Congrats Nicolas ! On the PhD and on those beautifully crafted slides 🤩

Nicolas ( @nicolasdufour.bsky.social ) is defending his PhD right now. I was so in awe of the presentation that I even forgot to take pictures 😅

Check out our new work: MIRO No more post-training alignment! We integrate human alignment right from the start, during pretraining! Results: ✨ 19x faster convergence ⚡ ✨ 370x less compute 💻 🔗 Explore the project: nicolas-dufour.github.io/miro/

Image generation becomes much more energy efficient. 👍

I'm super happy about Nicolas' latest work, probably the magnum opus of his PhD. Read the thread for all the great details. The main conclusion I draw from this work is that better pretraining, in particular by conditioning on better data, allows us to train SOTA models at a fraction of the cost.