Zhihao (Zach) Xia

I am a research scientist on Marc Levoy's computational photography team at Adobe, where I work on computer vision, computational photography and machine learning.

I received my PhD from WashU advised by Ayan Chakrabarti. Prior to WashU, I got my Bachelors from School of the Gifted Young at USTC. I've also spent time at research labs at Google Research, Adobe Research, NUS and KAUST.

Internships: I'm always happy to host research interns at Adobe. If you are interested in interning with me, please send me an email describing your past experience and current research interests.

Email  /  CV  /  Google Scholar /  GitHub

My research interests include computer vision, computational photography and deep learning. I am particularly interested in the design of accurate and efficient algorithms for visual inference --- reasoning different aspects of visual appearance (geometry, light, colors, etc) from images and videos.

A generative model that, given a coarsely edited image, synthesizes a photorealistic output that follows the prescribed layout.

Control pre-trained video generation models to synthesize arbitrary camera and subject motion based only on a given start and end frame.

A method to directly apply a pre-trained diffusion model to blind image restoration, by constraining the generative space using a set of anchor images.

A self-supervised training strategy for burst super-resolution that only uses noisy low-resolution bursts during training.

A custom diffusion model that can "rig" the lighting, facial expression, and head pose in a portrait photo while maintaining identity and preserving high-frequency features.

A parametric harmonization method employing semi-supervised training to learn local appearance harmonization from unpaired real composites.

An end-to-end differentiable pipeline for rendering sRGB images from raw inputs controlled by meaningful parameters.

High-fidelty depth recovery for tabletop objects from a single smartphone shot.

Face surface normal and reflectance estimation under uncontrolled visible lighting with a "dark flash image".

A neural network-based method to denoise pairs of images taken in quick succession in low-light environments, with and without a flash.

A basis prediction network that predicts global basis kernels and corresponding per-pixel coefficients for burst denoising.

A common model that is only trained once for a variety of monocular depth applications.

Image denoising by using DNN to exploit self-similarity in natural images.

Unsupervised framework for training image estimation networks from only degraded or partial measurements.

A robust deep learning model for the identification of polyadenylation signal - a critical factor for gene expression.

A deep convolutional neural network for image reconstruction under multiple light scattering.

CSE559A: Computer Vision - Fall 2018
Teaching Assistant

This webpage is cool