Wonbong Jang / Won

Hello! My name is Wonbong, you can also call me Won. I am currently a Postdoctoral Researcher at Meta's London office, on the MGenAI team, working on video generation and camera understanding. I completed my PhD at University College London under Prof. Lourdes Agapito, working on reconstructing 3D scenes from RGB images.

I was drawn to 3D computer vision because understanding the world from images is genuinely hard for machines, even though humans do it effortlessly, and because the field rests on beautiful mathematics and matrices. At Meta, I work on video diffusion models, teaching them to handle one of the oldest problems in computer vision (where the cameras are) without losing their generative capabilities, by learning camera trajectories as an additional modality alongside video.

Looking forward, I'm curious about how we can bring other modalities into these models, not only vision, but also language and action, and how we move beyond passive perception toward systems that decide where and how to look.

Before moving into computer vision and machine learning, I had the privilege of working for the Korean government for six years. Outside of research, I live in London and enjoy exploring its outskirts, drinking cortados, listening to podcasts, reading widely, and solving math problems. I'm always happy to chat about research, work, or anything adjacent. Feel free to reach out.

Email  /  Google Scholar  /  Twitter  /  LinkedIn  /  Threads

Rays as Pixels represents each camera as dense ray pixels (raxels) and denoises them jointly with video frames in a single video diffusion model. A decoupled self-cross attention mechanism lets one model predict camera trajectories from video, generate videos from one or more input images, and render novel views along a user-specified trajectory, with forward and inverse predictions that agree in a closed-loop self-consistency test. This is the first work that does video generation and camera pose prediction within a single model, which I've wanted to do since my early PhD days.

Kaleido pushes the idea of "3D perception is not a geometric problem, but a form of visual common sense" to novel view synthesis problem. It generates beautiful renderings and achieves per-scene optimization model with much fewer number of input images.

DT-NVS extends NViST by leveraging diffusion models, an approach well-suited for the inherently probabilistic nature of novel view synthesis in occluded regions. The method maintains consistency with the input image while generating diverse outputs for these unseen parts, as shown in the teaser figure with an occluded car. This research was completed during the final stages of my PhD studies and is available on arXiv.

DUSt3R generates 3D pointmaps from regular images without requiring camera poses. In practice, significant effort is put into camera calibration or deploying additional sensors to acquire point clouds. We present Pow3R, a single network capable of processing any subset of this auxiliary information. By incorporating priors, our method achieves more accurate and precise 3D reconstructions, multi-view depth estimation, and camera pose predictions. This approach opens new possibilities, such as processing images at native resolution and performing depth completion. Additionally, Pow3R generates pointmaps in two distinct coordinate systems, enabling the model to compute camera poses more quickly and accurately.

NViST turns in-the-wild single images into implicit 3D functions with a single pass using Transformers. Extending CodeNeRF to multiple real-world scenes, Feed-forward model, Transformers.

Disentangled NeRF, Conditional NeRF, Generalizing NeRF.