Michaël Soumm

I build generative models you can steer.

Founding AI Researcher at Davis AI in Paris, where I work on discrete diffusion models for floor plan generation. Before that, I did a PhD at CEA List (Université Paris-Saclay) and postdocs at Télécom Paris, Sorbonne Université and Inria, on diffusion and flow-matching models, controllable generation and disentangled representations.

Portrait of Michaël Soumm

Selected work

NeurIPS 2026Spotlight

Diff-CA

Separates what two image datasets share from what only one of them has, so that factor can be swapped or interpolated without losing image quality.

Project page Paper Code

AAAI 2026Oral

SAGA

A training-free way to make text-to-image models follow the prompt, for both diffusion and flow-matching models.

Project page Paper

Experience

Publications

News