Melvin Sevi

I’m a PhD student at Inria Grenoble, advised by Stéphane Lathuilière. I work on RL finetuning for generative image models — mostly autoregressive and flow-matching architectures — trying to get them to follow instructions and preferences more reliably instead of just sampling whatever the pretraining objective happened to reward.

Before the PhD, I did the MVA master’s (Mathématiques, Vision, Apprentissage) at ENS Paris-Saclay, and before that Applied Mathematics and Computer Science at Sorbonne University. In between, I spent a few months at LMU Munich with the Computer Vision & Learning group, working with Björn Ommer’s team on controllable text-to-image diffusion — that turned into a CVPR 2025 paper.

I’m generally drawn to problems at the intersection of generative modeling and control: getting a model to do precisely what you ask, not just what looks plausible.