Ang Cao

I am a final-year Ph.D. student (2020-) in Computer Science and Engineering (CSE) at the University of Michigan, Ann Arbor, working with Prof. Justin Johnson and Prof. JJ (Jeong Joon) Park. I was very fortunate to work with Prof. Andrew Owens.

I work on generative model and 3D vision, especially interested in worling modeling and 3D world creation,including:
(1) 3D/4D recontruction and generation from 2D data:
HexPlane; FWD; Lightplane
(2) Injecting world knowledge into 3D/4D via generative/foundation models (e.g., VLMs, Image/Video Diffusion Models...)
Text2Room; LiftGS
I also enjoy exloring cute ideas in generative models: ViLP

Before that, I was a M.S. Student at UMich ECE (2018-2020) and I did my Bachelor's degree at Wuhan University in China (2014-2018).

I am actively looking for full-time/postdoc positions in 2025!

Email  /  CV  /  Google Scholar  /  Twitter  /  Github

Research Scientist Intern, Meta FAIR, MPK, USA

Work with Sasha Sax , Franziska Meier

May 2024 - December 2024

Research Scientist Intern, Meta GenAI, London, UK

Work with David Novotny, Andrea Vedaldi , Natalia Neverova

May 2023 - November 2023

Equal contribution is indicated by * and listed in alphabetical order.

We (pre-)train a 3D Visual Language Grounding (3D VLG) model with only 2D supervision, by distilling languages from 2D foundation models with render-supervision.

How could a generator help the agent? We explore the visual language priors in VLMs by constructing novel question-image-answer triplets from image diffusion models. Proposed Image-DPO to encourage the model to use more visual inputs.

We present a neat transformer-based 3D recon and camera pose est. pipeline, which can reconstruct 3D scenes from 1000+ images in one forward pass with ultra-high speed.

A new state-of-the-art, fast pipeline for text-to-3D asset generation

We investigate "flashattention" for NeRF: we present Lightplane Splatter and Lightplane Renderer, a pair of extremely memory efficient modules which can lift 2D images to 3D and render from theoretically any 3D hash representation with 4-5 orders of magnitude memory savings. We show its usage in a set of 3D recon and generation tasks.

We do 4D generation with Gaussian Splatting by distilling motions from video diffusion models.

We generate meshes of full 3D rooms using text-to-image models.

An elegant representation for dynamic 3D scenes using six feature planes.

We show point rasterization can be really fast for sparse view novel view synthesis.

Revealing intriguing properties of detectors by applying our layout inversion technique.

Teaching assistant (GSI/Head TA), working with JJ Park