Paul Engstler

I work on 3D Scene Generation and Understanding, supervised by Andrea Vedaldi, Christian Rupprecht, and Iro Laina. Here are a few selected publications:

SynCity: Training-Free Generation of 3D Worlds

Paul Engstler*, Aleksandar Shtedritski*, Iro Laina, Christian Rupprecht, Andrea Vedaldi

Under Submission

Fully navigable 3D worlds beyond the small "bubble" worlds generated by prior methods. We prompt pre-trained 2D image and image-to-3D generators to create immersive environments without the need for any training.

arXiv Project Page Demo GitHub

Invisible Stitch: Generating Smooth 3D Scenes with Depth Inpainting

Paul Engstler, Andrea Vedaldi, Iro Laina, Christian Rupprecht

3DV 2025

Geometrically consistent scene generation through depth inpainting. We utilise depth information in the generation process, simplifying previous pipelines while creating more coherent 3D scenes.

arXiv Project Page 🤗 HF Demo GitHub

Understanding Self-Supervised Features for Learning Unsupervised Instance Segmentation

Paul Engstler*, Luke Melas-Kyriazi*, Christian Rupprecht, Iro Laina

NeurIPS 2023 Workshop: Self-Supervised Learning - Theory and Practice

Utilising features learned by self-supervised masked autoencoders (MAE) for unsupervised instance segmentation.

arXiv

* denotes equal contribution