Melvin Sevi

I’m a PhD student at Inria Grenoble, advised by Stéphane Lathuilière. My research explores how to make image generation more controllable, reliable, and computationally efficient. I focus on post-training generative models with reinforcement learning, with an interest in autoregressive, flow-matching, and hybrid architectures.

My broader goal is to help models follow compositional instructions and human preferences while using computation effectively. I’m interested in how visual representations and generation processes shape learning, decision-making, and control.

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.

Research interests

• Control and reliability: helping generative models follow instructions, represent compositional relationships, and reflect human preferences.
• Efficient learning and generation: understanding where learning signals and computational resources are most useful.
• Structured visual generation: exploring representations that support planning, refinement, and controllable synthesis.

Publications · Earlier projects · Email