PhD candidate · Computer Vision and Learning Lab · Heidelberg University
I work on generative image and video models. My research focuses on how to control them, how to make them cheaper to train and run, and how to use them to generate useful training data.
My PhD is supervised by Carsten Rother and Stefan Roth and funded by a scholarship from the Konrad Zuse School of Excellence in Learning and Intelligent Systems (ELIZA). Since January 2026, I have also been a visiting researcher in Stefan's group at TU Darmstadt.
From March to September 2026, I was an ML Research Intern in Vladlen Koltun's group at Apple in Zurich, where I worked on large-scale text-to-image and text-to-video diffusion and flow models, with a particular focus on video editing.
InterestsDiffusion & flow models · Controllable generation · Efficient large-scale models · Representation learning · Synthetic data
ECCV 2026
Compresses diffusion transformers such as FLUX by assigning ranks per block according to importance, truncating the least important blocks with SVD, and recovering quality through distillation. It keeps near-full quality at about 68% of the parameters.
ICLR 2026
PQGAN brings product quantisation into the VQGAN framework, giving image autoencoders higher reconstruction quality for downstream image synthesis.
3DV 2026
An online version of Video Depth Anything that borrows caching and frame masking from LLMs. It runs at 42 FPS on an A100 and 20 FPS on a Jetson.
ICPR 2026 Oral
Adapts an off-the-shelf diffusion model to a target domain using only imperfect pseudo-labels, then generates aligned training images from semantic maps. It gains up to 8.0 mIoU points over prior translation methods.
ACCV 2024
Uses a single-step Stable Diffusion "preimage" as a rich feature representation for fast, accurate monocular depth estimation.
ECCV 2024 Oral
Treats image control as a feedback-control problem, with information flowing both ways between the control network and the base model. The control network can be as small as 1% of the base model's parameters.
Visual Inference Lab, TU Darmstadt · Host: Stefan Roth
Apple, Zurich · Group of Vladlen Koltun
Research on large-scale text-to-image and text-to-video diffusion and flow models. Co-started a video-editing project and built its first data and training pipeline. When it grew into a larger team effort, I co-led it.
CompVis Group, Heidelberg University
Implemented deep learning methods for an industry collaboration on autonomous driving.
Heidelberg University
Efficient and controllable data synthesis, quality assessment of synthetic data. Funded by an ELIZA PhD scholarship.
Heidelberg University
Thesis: Continuous arbitrary-scale image synthesis with patch-wise diffusion models (Prof. Björn Ommer)
Heidelberg University
Thesis: Nonparametric prediction of plausible object motion (Prof. Björn Ommer)
Heidelberg University
Heidelberg University
Four theses on generative modelling, model efficiency, synthetic data and generative-model evaluation