AI researcher · currently at Amazon AGI
jiayewuw@gmail.comGoogle ScholarLinkedIn
Ten years ago I started reconstructing the world — its physical shapes and attributes. Along the way I began reimagining it — taking the measured world as material for creativity. Now I teach machines to reason about it — and measure how far they can go. I do science — to understand the world we live in.
About
I’m an Applied Scientist at Amazon AGI, where I work on post-training vision-language models: visual-token compression, synthetic data generation, and LLM evaluation.
I earned my Ph.D. in Computer Science from the University of Maryland, advised by David Jacobs. My research spans reconstruction of shape and physical attributes, and generative models for images, 3D shapes, and video. During my Ph.D., I was an Applied Scientist intern at Amazon Web Services, working on video vision-language models. I was also fortunate to collaborate with Roni Sengupta on inverse rendering.
As an undergraduate at Washington University in St. Louis, I worked with Yasutaka Furukawa on floorplan reconstruction.
Ph.D., Computer Science — University of Maryland, 2019–2025
B.S., Computer Science — Washington University in St. Louis, 2015–2019
Experience
Amazon AGI
Amazon Web Services
Research
Recovering geometry, materials, and lighting of indoor scenes under pronounced global illumination, captured with a moving, co-located light and camera.
Inserting new content into videos with realistic environmental effects, from shadows to splashes, controllable by artists through text prompts and masks.
Generating branching facial aging trees conditioned on lifestyle factors from a single photo via training-free diffusion.
Personalizing facial age transformation by adapting a global aging prior to an individual from a small personal photo collection.
Generating a personalized 3D face model of an individual from as few as 50 selfies.
Recovering geometry, materials, and lighting of indoor scenes under pronounced global illumination, captured with a moving, co-located light and camera.
A physically measured benchmark that closes a long-standing evaluation gap in intrinsic image decomposition.
Estimating object geometry and reflectance from just a few images captured at home with a camera and flashlight.
Reconstructing CAD floorplans from RGBD scans room by room via sequential shortest-path optimization.
3D reconstruction of residential buildings from aerial LiDAR scans by guiding procedural shape-grammar generation with neural networks.
A unified deep framework for vector-graphics floorplan reconstruction from raw 3D scans, recovering room structures and their semantics.
Building a NUMA-aware task-parallel runtime that provably retains classic work-stealing guarantees while mitigating work inflation.
Service
50 manuscript reviews across computer vision, graphics, and AI venues.
CVPR (2023–26) · ICCV (2023, 2025) · ECCV (2024, 2026) · WACV (2023, 2025–26) · AAAI (2025) · SIGGRAPH Asia (2024–25) · Eurographics (2026)
Journals: ACM TOG (2025) · IEEE TIP (2023, 2025) · CVIU (2024)
Get in touch
I’m always glad to connect about post-training multimodal models, physical intelligence, and creative AI.
jiayewuw@gmail.com ↗