Stan Szymanowicz

I am a PhD student at VGG group at University of Oxford, where I work on Computer Vision and Machine Learning with Andrea Vedaldi and Christian Rupprecht.

I spent the summer 2024 interning at Google San Francisco, continuing in London until April 2025.

Before my PhD, I worked at Microsoft Mixed Reality Lab (2021-2022) in Cambridge and received an MEng with Distinction from the University of Cambridge (2017-2021).

Email  /  Google Scholar  /  Twitter  /  Github

Research

I am interested in Computer Vision (with an emphasis on 3D), Computer Graphics and Generative Models.

Bolt3D allows feed-forward 3D scene generation in 6.25s on a single GPU. Our latent diffusion model directly outputs 3D Gaussians, and is trained on a large-scale dataset of reconstructed 3D scenes.

We show that leveraging a Monocular Depth Estimator and a Layered Representation allows building generalizable and efficient 3D scene reconstructors.

We design a fast (38FPS), simple, 2D network for single-view 3D reconstruction that represents shapes with 3D Gaussians. As a result, it can leverage Gaussian Splatting for rendering (588FPS), achieves state-of-the-art quality in several cases and is trains on just a single GPU.

Novel formulation of the denoising function in Diffusion Models lets us train 3D generative models from 2D data only. Our models can perform both few-view 3D reconstruction and 3D generation.

We unlock training NeRFs of 360◦ human heads from mobile phone captures. Synthetically-trained face keypoint detector allows to get a rough camera pose estimate which is then refined in the optimization process.

We can controllably animate Neural Radiance Fields by deforming tetrahedral cages. Tetrahedral connectivity allows us to run the animations in real-time.

We introduce an architecture for improved explainability of anomaly detection in CCTV videos.

We introduce a dataset and a method for detecting and explaining anomalies in CCTV videos.

Thank you to Jon Barron for the source code for the website!