Yilun Xu
Superintelligence@Meta
I'm a Research Scientist at Meta (TBD Lab). Previously I was a Research Scientist on the Google DeepMind Gemini team, building some awesome generative models (Gemini Omni and Nano Banana Pro). Before that, I co-led the diffusion-distillation efforts at NVIDIA.
I received my PhD @ MIT EECS, advised by Tommi Jaakkola, and my BS from the Turing Class in EECS @ Peking University, working with Yizhou Wang (PKU) and Stefano Ermon (Stanford). During my time at GDM I worked with Tim Brooks; at NVIDIA I worked with Arash Vahdat and Ming-Yu Liu.
Email: aaronyilun.xu@gmail.com
Models
More coming soon
New models & releases on the way ✨
Muse Image / Video
Meta · Superintelligence LabFoundational models for agentic image and video generation.
Building the pre-training and post-training stack from scratch.
Read the announcement →.svg){kind=logo}
Try it on meta.ai
Gemini Omni
Google DeepMind · GeminiA natively multimodal foundation model that reasons jointly across text, image, video, and audio.
Core contributor in building initial pre-training, post-training and eval stack.
Read the announcement →
Nano Banana Pro
Google DeepMind · GeminiState-of-the-art (at release time) image generation and editing, delivering high-fidelity, controllable visuals.
Core contributor in post-training; building the distillation stack from scratch.
Read the announcement →Research (≤ PhD)
My research centers on deep generative modeling — designing new model families, improving training, accelerating sampling, and pushing discrete diffusion and distillation.
Publications (≤ PhD)
(*) denotes equal contribution.
COSMOS-Distilled
Distilled version of the COSMOS video foundation models. Featured in Jensen Huang's CES / GTC 2025 keynote — NVIDIA.
One-step Diffusion Models with f-Divergence Distribution Matching
Preprint, 2025 · code released as part of FastGen
Truncated Consistency Models
International Conference on Learning Representations (ICLR), 2025
Heavy-Tailed Diffusion Models
International Conference on Learning Representations (ICLR), 2025
Energy-Based Diffusion Language Models for Text Generation
International Conference on Learning Representations (ICLR), 2025
Think While You Generate: Discrete Diffusion with Planned Denoising
International Conference on Learning Representations (ICLR), 2025
Hamiltonian Score Matching and Generative Flows
Neural Information Processing Systems (NeurIPS), 2024
DisCo-Diff: Enhancing Continuous Diffusion Models with Discrete Latents
International Conference on Machine Learning (ICML), 2024
Particle Guidance: non-I.I.D. Diverse Sampling with Diffusion Models
International Conference on Learning Representations (ICLR), 2024; Deep Inverse & Diffusion Workshops, NeurIPS 2023 (Oral)
Restart Sampling for Improving Generative Processes
Neural Information Processing Systems (NeurIPS), 2023
Stable Target Field for Reduced Variance Score Estimation in Diffusion Models
International Conference on Learning Representations (ICLR), 2023
Poisson Flow Generative Models
Neural Information Processing Systems (NeurIPS), 2022 (Spotlight)
A Survey on Generative Diffusion Models
IEEE Transactions on Knowledge and Data Engineering (TKDE), 2023
Can Subnetwork Structure Be the Key to Out-of-Distribution Generalization?
International Conference on Machine Learning (ICML), 2021 (Long Talk)
TCGM: An Information-Theoretic Framework for Semi-Supervised Multi-Modality Learning
European Conference on Computer Vision (ECCV), 2020 (Oral)
Education
Massachusetts Institute of Technology
Sep 2020 – May 2024
Ph.D. in Computer Science · Advisor: Tommi Jaakkola
Peking University
Sep 2016 – Jul 2020
B.S. in Turing Class, Computer Science (summa cum laude) · Advisor: Yizhou Wang
Miscellaneous
I'm a national second-class table tennis player in China and the men's singles champion of the PKU Freshman's Cup (2016). I also played for the MIT table tennis team, taking 3rd place (group) at the 2022 NCTTA Upper New England Championship. 🏓