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

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 😅

#KostasThoughts: Please stop using AI tools to write reviews. I want your authentic thoughts, even if they contain spelling, grammar, & factual mistakes. An LLM should not influence your judgment. The purpose of multiple reviewers is to provide diverse views and error checking.

There has to be a decoupling from research for the sake of increasing knowledge, and research for the sake of pleasing VCs. As long as these conference are the key to a career making millions, they will be overflown by people just taking the gamble and thus not taking the reviewing part seriously.

All the slop review drama unfolding at ICLR is painful to watch. But having so many eyes on it puts us in a better point to find solutions. This would have taken longer to be exposed on a closed platform where we may have had a few screenshots and a few author stories at best.

Update on #CVPR2026 full paper & compute form submission issue: We would like to inform authors that the OR submission system requires the submission of a compute reporting form along with any updates to the full paper. We have identified this as a system-related issue. 1/2

This Nature retrospective is quite interesting. To me, the only solution to the credit assignment problem is obvious: stop believing a single person is responsible for every big discovery. It's an artifact of our monkey brain requiring a face for storage, not the reality of how knowledge progresses.

Image generation becomes much more energy efficient. 👍

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/

Privileged to the diffusion master @nicolasdufour.bsky.social give to our team (full house) a tour of his excellent works in data and compute efficient diffusion models and a sneak preview of his latest MIRO work. Check it out 👌

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.

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 📉

👀 arxiv.org/abs/2510.25897 Thread with all details coming soon!

"The Principles of Diffusion Models" by Chieh-Hsin Lai, Yang Song, Dongjun Kim, Yuki Mitsufuji, Stefano Ermon. arxiv.org/abs/2510.21890 It might not be the easiest intro to diffusion models, but this monograph is an amazing deep dive into the math behind them and all the nuances

"familiar to most members of society"? Hell, if I know 20% of the things in that table, that's a maximum. Who are those guys and in what society do they live? This is exactly why I think the concept of AGI is meaningless.

Probably the best paper title of my career. To read with the Indiana jones soundtrack. And yes we solved and differentiated quite well 80 millions (small) Fused Gromov Wasserstein problems per epoch using a neural network on GPU.

Did you open-source your #ICCV2025 works? As a PyTorch Ambassador, I would like to write an article to introduce open-sourced ICCV 2025 works (including workshops and demos) for promoting open-source/science + PyTorch If interested, share your work via the form in my reply 👇

Over the past year, my lab has been working on fleshing out theory + applications of the Platonic Representation Hypothesis. Today I want to share two new works on this topic: Eliciting higher alignment: arxiv.org/abs/2510.02425 Unpaired learning of unified reps: arxiv.org/abs/2510.08492 1/9

Happening now at VRG Colloquium by @vickykalogeiton.bsky.social

Very proud of our recent work, kudos to the team! Read @davidpicard.bsky.social’s excellent post for more details or the paper arxiv.org/pdf/2502.21318

Final note: I'm (we're) tempted to organize a challenge on that topic as a workshop at a CV conf. ImageNet is the only source of images allowed and then you compete to get the bold numbers. Do you think there would be people in for that? Do you think it would make for a nice competition?

🚨Updated: "How far can we go with ImageNet for Text-to-Image generation?" TL;DR: train a text2image model from scratch on ImageNet only and beat SDXL. Paper, code, data available! Reproducible science FTW! 🧵👇 📜 arxiv.org/abs/2502.21318 💻 github.com/lucasdegeorg... 💽 huggingface.co/arijitghosh/...

Some conclusions and food for thought for the future research.

Another great event for @valeoai.bsky.social team: a PhD defense of Corentin Sautier. His thesis «Learning Actionable LiDAR Representations w/o Annotations» covers the papers BEVContrast (learning self-sup LiDAR features), SLidR, ScaLR (distillation), UNIT and Alpine (solving tasks w/o labels).

Familiar names among #ICCV2025 Outstanding Reviewers from our team 😇 Antoine Guédon @antoine-guedon.bsky.social Sinisa Stekovic Renaud Marlet 👏 @iccv.bsky.social iccv.thecvf.com/Conferences/...

Getting close to the beautiful fall colors right outside the lab 🍁🍂

I coined a new verb on my slack channel. Suttonize: verb. Removing dedicated processing that require careful tuning and replacing it with simpler but more intensive computation. Ex: "We should suttonize this: just remove the encoder and concat the sources together with a few self-attention layers".

Today is Antoine Guedon's PhD! Already pretty cool visuals right at the start.

Did not know that all authors that submit 3 ICLR of more must be reviewers with 6 (or more?) papers. Seems fair to review 2 papers per submission but after the full NeurIPS summer (and rejection) I would have appreciated a reduced load to do them better (I say that as an ICML 2025 top reviewer).

Finally the dates for @cvprconference.bsky.social are public! cvpr.thecvf.com/Conferences/... And there's an additional week for supp videos 🥰

As an SAC for @neuripsconf.bsky.social, I don't agree with PCs approach to reject papers based on ranking. I ranked my accepted papers as requested and explicitly stated that I support the acceptance of all papers. I was not given an explanation of why papers at the end of the rank were rejected.

Annnnnd it's a reject! Scale is a religion and if you go against it, you're a heretic and you should burn, "despite [the reviewers] final ratings". But scale is still not necessary! Side note: First time swinging reviews up (from 2,2,4,4 to 2,4,4,5) does not get the paper accepted. Strange days.

Unfortunately, our submission to #NeurIPS didn’t go through with (5,4,4,3). But because I think it’s an excellent paper, I decided to share it anyway. We show how to efficiently apply Bayesian learning in VLMs, improve calibration, and do active learning. Cool stuff! 📝 arxiv.org/abs/2412.06014

If you're interested in human pose estimation and mesh recovery from LiDAR data, we have this massive survey: arxiv.org/abs/2509.12197 Salma and Nermin put a tremendous amount of work in it, there's everything: the tasks, all the methods organized, datasets, numbers, challenges and opportunities.

We will have GPT6 before GTA6

In related news: the NeurIPS decisions are next week.

1/n🚀Gaussians > Differentiable function > Mesh? Check out our new work: MILo: Mesh-In-the-Loop Gaussian Splatting! 🎉Accepted to SIGGRAPH Asia 2025 (TOG) MILo is a novel differentiable framework that extracts meshes directly from Gaussian parameters during training. 🧵👇

Let's push for the obvious solution: Dear @neuripsconf.bsky.social ! Allow authors to present accepted papers at EurIPS instead of NeurIPS rather than just additionally. Likely, at least 500 papers would move to Copenhagen, problem solved.

NeurIPS has decided to do what ICLR did: As a SAC I received the message 👇 This is wrong! If the review process cannot handle so many papers, the conference needs yo split instead of arbitrarily rejecting 400 papers.

Why would you take the plane when a train goes that fast? You arrive directly in the city center, and there's no hour long security check. Please @ec.europa.eu tax kerosene and use the money to lower the price of rail.

Going to Strasbourg for #GRETSI2025 but maybe pushing to Berlin?🤔 In the 90s there were trains called "eurocity" that would go across countries. Mine was the Heinrich Heine from Paris Est to Prague. I would take it in the evening and arrive in Leipzig Hbf the next morning. It was very convenient.

Ah! Smart guy didn't even put the title of the paper in the thread 🙄 "Markov Chain Estimation with In-Context Learning" Simon Lepage, Jeremie Mary, David Picard arxiv.org/abs/2508.03934

Next week, I'll be in Strasbourg for the GRETSI (@gretsi-info.bsky.social) to present a small discovery on transformers generalization we made with Simon and Jérémie while working on generative recommender systems. I love these "phase transition" plots. 📜: arxiv.org/abs/2508.03934 Short summary 👇

We're already at +1.5°C, now imagine what this curve will look like at +4°C. There's no adaptation to that. As long as we're in a democracy, everyone of us has a voice to prevent further warming through public policies. As long as we have a free market, everyone of us can choose what not to fund.

🚀 DinoV3 just became the new go-to backbone for geoloc! It outperforms CLIP-like models (SigLip2, finetuned StreetCLIP)… and that’s shocking 🤯 Why? CLIP models have an innate advantage — they literally learn place names + images. DinoV3 doesn’t.

My take on GPT-5: The best achievement is that the LLM wrote its own tech report. Or at least it really looks like it. (cdn.openai.com/pdf/8124a3ce...)

Dear bsky friends, I have a question: Do you really think that the visual quality of these images is so bad that the research that produced them is deeply flawed? And if I told you that the model was mostly trained on ImageNet with a bit of artistic fine-tuning at 1024 resolution, still really bad?