Youssef Saied

PhD student with @francoisfleuret.

[Computer Vision and Pattern Recognition Papers @CSVisionPapers]

Velocity Scaling in Flow Matching

Youssef Saied, François Fleuret
https://t.co/toTnuLiCyV [𝚌𝚜.𝙲𝚅]

[Shumpei Takezaki @shumpeiMaxwell]

Velocity Scaling in Flow Matching

TL;DR: Challenges the idea that MSE training causes a velocity deficit in Flow Matching. Instead, scaling corrects a population-level time lag in sampling. A time-lag-based scale improves ImageNet-256 FID from 28.0 to 12.2.
https://t.co/eRuGOD5Rxa

[Ramón Calvo @noctrog]

Your RAE does not need all of those tokens.

Meet PoolDINO: 4x-16x fewer tokens, similar generation quality, and both cheaper training and sampling.

Here's RAE vs. 8x compression on an M1 Pro CPU (no accelerator).

🧵

[Mariia Drozdova @MariiaD_ML]

1/7) What happens when we add a persistent memory state and remove timestep conditioning from a denoising model?

We obtain an anytime iterative solver!

With less than 250k-params, we obtain a 99.9% accuracy on Sudoku-Extreme and 98.3% on Maze-Unique.

https://t.co/FL4Mhi7DuZ

[General Intuition @gen_intuition]

Introducing MIRA.

A playable, multiplayer world model.
A dream of Rocket League.

Trained on 10k hours of data collected with publicly available bots, MIRA learns the dynamics of a four-player game. The model runs in real time at 20 fps, based on the keys you and the other players press.

Built by General Intuition and @kyutai_labs, in collaboration with Epic Games. Not used to develop Rocket League.

▶️ Play the demo, read the technical report, and explore the open-source code at https://t.co/JjlsamGE1D

At ICML? Find us at Booth 111 to try it yourself and dig into the results with the team.

[François Fleuret @francoisfleuret]

TL;DR: To denoise with levels of noise unseen during training simply normalize / denoise / denormalize. The resulting equivariance shines for sampling!

@YoussefMMSaied (ICML 2026)

@sciences_UNIGE @UNIGEnews

[Youssef Saied @YoussefMMSaied]

10/

📄 https://t.co/VZkcKTtPEs

With François Fleuret.

ICML 2026, Seoul, July 6–11.

Happy to chat about NE, structural priors, or denoising — at the conference or here.

[Youssef Saied @YoussefMMSaied]

1/

What happens when a denoiser trained at one noise level is reused inside an iterative sampler?

We trained SwinIR with Noise2Noise at σ=10, then dropped it unchanged into the constrained sampler of @ZKadkhodaie & @EeroSimoncelli (NeurIPS ’21) for 10% random inpainting.

Baseline SwinIR: 6.08 dB on Set12.

SwinIR-WNE: 23.87 dB.

Same backbone. Same N2N pairs. Same sampler.

With François Fleuret @francoisfleuret. ICML 2026. 🧵

[Youssef Saied @YoussefMMSaied]

9/

Bonus: WNE turns Jacobian rows into actual adaptive filters.

By Euler’s theorem for homogeneous functions, at differentiable inputs, f(y)=J_f(y)y.

This is not a Taylor approximation: each output feature is exactly the dot product between the input and one Jacobian row.

NE additionally forces each filter to sum to 1.

Under mismatch, WNE filters stay edge-aligned; baseline filters smear.

[Bálint Máté @balintmt]

new preprint on solvation free energies:

tl;dr: We define an interpolating density by its sampling process, and learn the corresponding equilibrium potential with score matching. https://t.co/sO0EI4sy5r

with @francoisfleuret and @tristanbereau
(1/n)

[Eloi Alonso @EloiAlonso1]

Ever wanted to play Counter-Strike in a neural network?

These videos show people playing (with keyboard & mouse) in 💎 DIAMOND's diffusion world model, trained to simulate the game Counter-Strike: Global Offensive.

💻 Download and play it yourself → https://t.co/vLmGsPlaJp

🧵

[Daniele Paliotta @DanielePaliotta]

neurips!