An explorer from Quzhou

I am a senior research scientist at NVIDIA Deep Imagination Research Group, where I work on probabilistic modeling of high-dimensional data. Right now, I’m leading our world-model training efforts.

Prior to joining NVIDIA, I was:

My “generative AI” qualifications:

  • I lead NVIDIA’s large-scale diffusion model training. From infra ⚡implementing efficient file I/O, designing parallel strategies, hunting down the one buggy/slow GPU among 10 000 🤕 to ablating models and babysitting training runs. We delivered and open-sourced COSMOS world models and shipped closed-source Picasso in NVIDIA AI Foundations (and, after two years, a tech report finally saw the light).
  • I co-authored DEIS, the first to introduce exponential integrators for diffusion sampling—cutting sampling steps down to 10–15.
  • I democratized using generative diffusion models for sampling unnormalized probability densities (e.g., amortized molecular conformer generation) in this paper.

In general, my research approach centers on the representation, learning, and sampling of complex probability distributions.

Email: qsh.zh27 [at] gmail [dot] com

news

see all news →
Jul 1, 2025
Cosmos Predict2 release! SOTA world model for AV driving and robots.
Jan 11, 2025
Cosmos Predict1 release! First open source large scale world model
Dec 1, 2023
Graduated from GaTech and joined Nvidia as a research scientist working on NVIDIA GenAI, specifically Picasso.
May 16, 2022
ICML22 Wasserstein gradient flow in pixel space!
Apr 30, 2022
Training Free DEIS 10 steps to generate high-fidelity samples for diffusion model.

Selected publications

See Google Scholar for the full list of publications →
  1. Cosmos World Foundation Model Platform for Physical AI thumbnail
    Cosmos World Foundation Model Platform for Physical AI
    NVIDIA (led the model training and development)
    Whitepaper. 2025

    🎯 NVIDIA’s open-source video world model platform

    PDF Code Website Jensen Huang Keynote at CES 2025
  2. Masked Diffusion Models are Secretly Time-Agnostic Masked Models and Exploit Inaccurate Categorical Sampling thumbnail
    ICLR
    Masked Diffusion Models are Secretly Time-Agnostic Masked Models and Exploit Inaccurate Categorical Sampling
    Zheng Kaiwen, Chen Yongxin, Mao Hanzi, Liu Ming-Yu, Zhu Jun, and Zhang Qinsheng
    In International Conference on Learning Representations. 2025

    🎯 mask diffusion model; numerical issue; faster sampling

    PDF Website
  3. Fast Sampling of Diffusion Models with Exponential Integrator thumbnail
    ICLR
    Fast Sampling of Diffusion Models with Exponential Integrator
    Zhang Qinsheng, and Chen Yongxin
    In International Conference on Learning Representations. 2023

    🎯 exponential integrator for diffusion sampling; sampling with 10-15 steps

    arXiv url Code Website
  4. ediffi: Text-to-image diffusion models with an ensemble of expert denoisers thumbnail
    ediffi: Text-to-image diffusion models with an ensemble of expert denoisers
    Balaji Yogesh, Nah Seungjun, Huang Xun, Vahdat Arash, Song Jiaming, Zhang Qinsheng, Kreis Karsten, Aittala Miika, Aila Timo, Laine Samuli, Catanzaro Bryan, and others
    Whitepaper. 2022

    🎯 NVIDIA’s text-to-image foundation models

    arXiv Website

Student Mentees and Collaborators

I have been fortunate to work with many talented students and collaborators throughout my research journey