Henry Li

prof_pic.jpg

About Me: I study the theory, guidance and fine-tuning of diffusion models. I have previously interned with the Seed Foundation Team at TikTok / ByteDance, the Center for Artificial Intelligence at Bosch, and the Center for Computational Mathematics at the Flatiron Institute. I have published both theoretical and applied work on diffusion models.

Theoretical

  • ICML 2023 Workshop 1) Exponential weight averaging as damped harmonic motion 2) Non-normal Diffusion Models : Non-normal diffusion processes and connecting EMA training of diffusion models to two-body Hookean systems.
  • ICLR 2024 Spotlight Likelihood Training of Cascaded Diffusion Models via Hierarchical Volume-preserving Maps
    : Retaining likelihood estimation capabilities in multi-stage diffusion models (in collaboration with Meta AI).

Applied

  • NeurIPS 2024a Solving Inverse Problems via Diffusion Optimal Control
    : Guided diffusion modeling through the lens of optimal control (in collaboration with Bosch AI).
  • NeurIPS 2024b Boosting Alignment for Post-Unlearning Text-to-Image Generative Models
    : Model alignment and unlearning in diffusion models.
  • In Submission Measurement Consistent Tweedie’s Solving Inverse Problems with the Conditional Posterior Mean (see "publications" for pre-print) : Fast and efficient guided diffusion modeling for image reconstruction in the presence of measurement noise.
  • In Submission Dual Diffusion for Unified Image Generation and Understanding
    : Large-scale diffusion models for simultaneous text and image generation (in collaboration with Seed Team at TikTok / ByteDance).

news

Feb 01, 2025 On the market for research positions in generative modeling. If you know of a suitable role, please reach out!
Dec 01, 2024 Two papers accepted to NeurIPS 2024!
Jul 01, 2024 One spotlight paper accepted to ICLR 2024!

selected publications

  1. arXiv
    dualdiff.mp4
    Dual Diffusion for Unified Image Generation and Understanding
    Zijie Li*Henry Li*, Yichun Shi, and 4 more authors
    In arXiv, 2024
  2. NeurIPS
    unlearn.png
    Boosting Alignment for Post-Unlearning Text-to-Image Generative Models
    Myeongseob Ko*Henry Li*, Zhun Wang, and 6 more authors
    In NeurIPS, 2024
  3. NeurIPS
    doc.png
    Solving Inverse Problems via Diffusion Optimal Control
    Henry Li, and Marcus Pereira
    In NeurIPS, 2024
  4. ICLR (Spotlight)
    pcdm.gif
    Likelihood Training of Cascaded Diffusion Models via Hierarchical Volume-preserving Maps
    Henry Li, Ronen Basri, and Yuval Kluger
    In ICLR, 2024