Yanwu Xu, Ph.D.
Advancing fast, compact, and on-device generative media for image and video synthesis and editing.
I am a Senior Software Engineer at Google, where I focus on efficient and on-device generative models across text-to-image generation, generalist image editing, and video synthesis. Previously, I worked as a Senior Research Scientist at FinBone building generative tools for AI filmmaking workflows, and as a Research Scientist at Snap Inc. (Creative Vision) researching mobile-friendly image and video models.
I received my Ph.D. in Electrical and Computer Engineering from Boston University in 2024, advised by Dr. Kayhan Batmanghelich and co-advised by Dr. Mingming Gong, with a dissertation on improving the efficiency of conditional generative models. Earlier, I earned an M.S. in Intelligent Systems from the University of Pittsburgh (2022) and completed multiple research internships and student researcher terms at Google (2022–2023).
Second Gear Generation (Fast Generation)
Interactive Demo · Denoising Gear Shift
From 50-Step Diffusion to Second Gear (1-Step)
Distills iterative denoising into a single forward pass via Diffusion-GAN & adversarial training (UFOGen, SF-V), enabling subsecond on-device image and video synthesis.
01 / Acceleration
Few- & One-Step Diffusion
Hybrid Diffusion-GAN training and adversarial distillation for real-time synthesis (SIDDMs, UFOGen, SF-V).
02 / Mobile & Edge
On-Device Image & Video
Compact architectures and inference optimization for mobile deployment (BlazeEdit, MobileDiffusion, SnapGen, SnapGen-V, Mobile-VTON).
03 / Foundations & 3D
Generative Theory & Medical AI
Conditional GAN alignment, unpaired image translation, and text-guided 3D volumetric medical synthesis (MedSyn, TAC-GAN).
Featured Work & Public Releases
Publicly released models, research milestones, and product contributions spanning Google DeepMind, Google, and Snap Research.
Veo Lite
Contributor to Veo Lite, working on few-step model distillation and low-latency optimization for production video generation.
BlazeEdit
A 195M-parameter generalist on-device image-to-image diffusion editor consolidating five tasks with 290 ms inference on Pixel 10.
SnapGen & SnapGen-V
379M-parameter on-device 1024×1024 text-to-image generation in 1.4s (Highlight), extended to generating 5s mobile videos within 5s.
UFOGen
You Forward Once: Diffusion-GAN hybrid enabling single-step large-scale text-to-image generation and editing.
MobileDiffusion
Subsecond text-to-image diffusion on mobile hardware via streamlined UNet attention and sampling optimization.
SF-V
Single Forward Video Generation Model achieving single-step video diffusion through adversarial fine-tuning.
Selected Publications
Recent representative work on efficient diffusion, mobile media generation, and medical imaging.
-
BlazeEdit: Generalist Image Editing on Mobile Devices with Image-to-Image Diffusion Models
-
Mobile-VTON: High-Fidelity On-Device Virtual Try-On
-
SnapGen: Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training
-
SnapGen-V: Generating a Five-Second Video within Five Seconds on a Mobile Device
-
UFOGen: You Forward Once Large Scale Text-to-Image Generation via Diffusion GANs
-
MedSyn: Text-Guided Anatomy-Aware Synthesis of High-Fidelity 3D CT Images
Experience & Education
Industry engineering, research leadership, and academic training.
-
Sep 2025 – Present · Sunnyvale, CA
Senior Software Engineer
Google
Contributor to Veo Lite (few-step video model distillation and low-latency production delivery) and BlazeEdit; own model and runtime optimization for efficient visual generation and editing.
-
Jun 2025 – Sep 2025 · Los Angeles, CA
Software Engineer
Google
Built on-device image and video generation pipelines balancing latency, memory footprint, and fidelity.
-
Jan 2025 – May 2025 · Los Angeles, CA
Senior Research Scientist
FinBone
Researched generative image and video models for controllable AI filmmaking workflows.
-
Aug 2024 – Dec 2024 · Los Angeles, CA
Research Scientist, Creative Vision
Snap Inc.
Worked on fast visual generation for mobile and server deployment (SF-V, SnapGen, SnapGen-V).
-
May 2022 – Nov 2023 · Mountain View, CA & Boston, MA
Research Intern & Student Researcher
Google
Researched Diffusion-GAN hybrids, 2D/3D visual stylization, and subsecond mobile diffusion (SIDDMs, UFOGen, MobileDiffusion).
-
2023 – 2024 · Boston, MA
Ph.D., Electrical & Computer Engineering
Boston University
Advised by Dr. Kayhan Batmanghelich and co-advised by Dr. Mingming Gong.
-
2019 – 2022 · Pittsburgh, PA
M.S., Intelligent Systems Program
University of Pittsburgh
Research on generative modeling, domain generalization, and medical image analysis.
-
2013 – 2017 · Changsha, China
B.S., Mechanical & Electronic Engineering
Central South University
News & Milestones
Recent announcements and recognitions.
- 2026 BlazeEdit accepted to CVPR 2026 EDGE Workshop; Mobile-VTON accepted to CVPR 2026; MatrixFSDP released on arXiv.
- 2025 Joined Google full-time (Jun) and promoted to Senior Software Engineer (Sep); SnapGen (CVPR 2025 Highlight) & SnapGen-V accepted to CVPR 2025.
- 2024 Completed Ph.D. at Boston University; joined Snap Inc.; SF-V (NeurIPS), MobileDiffusion (ECCV), UFOGen (CVPR), and MedSyn (IEEE TMI) published.
- 2023 SIDDMs accepted to NeurIPS 2023; Shortest Path Regularization accepted to CVPR 2023.
- 2022 MSPC accepted to CVPR 2022; Adversarial Consistency accepted to MICCAI 2022; Hierarchical Amortized GAN published in IEEE JBHI; started Google internship.
- 2021 Papers accepted to ICCV 2021, ACM MM 2021, and IJCAI WSRL Workshop.
- 2018–20 Two papers at AAAI 2020; TAC-GAN (NeurIPS 2019 Spotlight); Frontiers in Neuroscience 2020; ACCV 2018 & MICCAI BraTS 2018.
Published Patent Applications
Publicly published patent applications assigned to Google LLC.
US20250157008A1 · 2025
Image Generation with Minimal Denoising Diffusion Steps
US20250165756A1 · 2025
Resource-Efficient Diffusion Models
WO2024254328A1 · 2024