David Bert

We have been exploring JITs (arxiv.org/abs/2511.13720) a lot recently, and we are quite enthusiastic about them. Here are some exciting results obtained in just 3.5 days on 4 nodes. Because JITs are trained to predict x₀ directly in pixel space, we can layer standard computer-vision losses on top of the usual flow-matching objective. That’s exactly what PixelGen (arxiv.org/pdf/2602.02493…) does, using LPIPS and a DINO-based perceptual loss. We tried replicating this with a small PRX model trained from scratch, and the results are really promising. 🧵

We ran a large training ablation logbook for PRX (our text-to-image model trained from scratch). A few changes had surprisingly large impact. Part 2 of the PRX series is out — summary + link in the thread 🧵

Great write up from @photoroom_ML on from scratch diffusion models training, testing and reflecting upon a lot of the recent papers huggingface.co/blog/Photoroom… great catch up with the field if you've missed a few papers (I did)

@AdilZtn @UMA_Robots Congrats! This is super exiting and I can’t wait to see what you are working on!

🦾 I’m excited to share that I’ve joined @UMA_Robots as a Founding Research Scientist. UMA is building mobile and humanoid robots with human-level dexterity and a deep understanding of the world, universal mechanical assistants designed to give people more time, more safety, and more capability. I’m grateful to join an exceptional team working at the frontier of robotics and embodied AI. The journey ahead is big. Join us at uma.bot

The LLM space is blessed to have awesome tech reports that help push the field further exponentially. The diffusion community kinda lacks that a bit. Good folks at @photoroom_app decided to change that by releasing PRX under Apache 2.0 with solid reporting.

The photoroom team published an open end-to-end Diffusion Training Playbook! Have a look

The next blogs will talk about how they accelerated training and post-training! Excited for it, stay tunned: huggingface.co/blog/Photoroom…

@Birchlabs If you backup per iterations and not epoch It sometimes means that you are restarting from data recently seen in the previous run

when your best loss improvement comes from resuming a run

@AdilZtn @EugeneVinitsky @ncklashansen I agree!

@EugeneVinitsky Definitely, @ncklashansen consistently produces publications with fresh ideas. The open-source code is clean, and we can build on top of it. Finally, TDMPC rocks!

Q1: Who is your favorite RL researcher that you think should be more widely known? Why? 1/2

👋👨‍🍳🍵 After a year of cooking up a secret project, I'm thrilled to officially reveal: The 𝐋𝐄𝐍𝐒 𝐏𝐫𝐨𝐣𝐞𝐜𝐭. By combining modern tools of Explainable AI, how much can we explain a ResNet50? 🧶

Interested in results at the intersection between explainability🔍 and optimal transport 🚚? Come check out "On the explainable properties of 1-Lipschitz Neural Networks: An Optimal Transport Perspective" on Tuesday at 5:15pm, panel #1508.

Happy to share that our work "Look where you look! Saliency-guided Q-networks for RL visual tasks" has been accepted at #NeurIPS2022 🎉🥳 Authors: @DavidBert7 @MehdiZouitine1 @AdilZtn @EmRachelson Paper link: arxiv.org/abs/2209.09203 1/n