Valentin De Bortoli

prof_pic.jpg

I do research at Google DeepMind in London. I work on diffusion models — how to build them, how to sample from them efficiently, and why they work — and more broadly on sampling, optimal transport and Markov chain theory.

Short CV

  • 2023 – now · Research Scientist, Google DeepMind
  • 2022 – now · CNRS Researcher (on leave), ENS Ulm
  • 2020 – 2022 · Postdoctoral Researcher, University of Oxford
  • 2017 – 2020 · PhD, ENS Paris-Saclay

news

Aug 01, 2026 The DiffusionGemma technical report is out.
Sep 20, 2025 Three papers accepted at NeurIPS 2025, on inference-time annealing, memorization in diffusion models, and proximal MCMC for non-log-concave sampling.
May 01, 2025 Distributional Diffusion Models with Scoring Rules accepted at ICML 2025.

selected publications

  1. NeurIPS·2025
    On the Edge of Memorization in Diffusion Models
    Sam Buchanan, Druv Pai, Yi-Ting Ma, Valentin De Bortoli

    Studies where diffusion models stop generalising and start reproducing their training data, and what controls the boundary between the two.

  2. NeurIPS·2023
    Diffusion Schrödinger Bridge Matching
    Yuyang Shi, Valentin De Bortoli, Andrew Campbell, Arnaud Doucet

    Reformulates the Schrodinger bridge as iterative Markovian fitting, removing the error that accumulated across outer iterations in earlier bridge solvers.

  3. ICML·2023
    SE(3) Diffusion Model with Application to Protein Backbone Generation
    Jason Yim, Brian L. Trippe, Valentin De Bortoli, Emile Mathieu, Arnaud Doucet, Regina Barzilay, Tommi Jaakkola

    A diffusion model on SE(3) that generates protein backbones by denoising the rigid-body frame of each residue.

  4. TMLR·2022
    Convergence of Denoising Diffusion Models under the Manifold Hypothesis
    Valentin De Bortoli

    Convergence guarantees for denoising diffusions when the data lies on a low-dimensional manifold - the regime where the score is unbounded and earlier analyses break down.

  5. NeurIPS·2022
    Riemannian Score-Based Generative Modelling
    Valentin De Bortoli, Emile Mathieu, Michael Hutchinson, James Thornton, Yee Whye Teh, Arnaud Doucet

    Extends score-based generative models from Euclidean space to Riemannian manifolds, so data living on spheres, tori and other geometries is modelled in its own coordinates.

  6. NeurIPS·2021
    Diffusion Schrödinger Bridge with Applications to Score-Based Generative Modeling
    Valentin De Bortoli, James Thornton, Jeremy Heng, Arnaud Doucet

    Casts generative modelling as a Schrodinger bridge - an entropic optimal transport problem - and solves it by iterative proportional fitting on learned scores.

All publications →