| Oct 30, 2024 | Excited to shat that our paper, “AnomalyDINO: Boosting Patch-based Few-shot Anomaly Detection with DINOv2”, got accepted at WACV 2025!🎉 In this work, we show how DINOv2 representations can effectively address the few-shot (and even one-shot) visual anomaly detection task. You can find the full paper arXiv and access the code in our public GitHub repository. Once again, I want to give a huge shouthout to Simon (first author) and our supervisors Johannes and Asja for their incredible collabotation. |
| Aug 29, 2024 | I am thrilled to announce that I have successfully defended my PhD Thesis, titled Exploring Different Biases in Generative Models, with the highest distinction—summa cum laude. In this work, I demonstrate how my previous papers align with the general notion of bias in generative models. While my research on Benchmarking the Fairness of Image Upsampling Methods addresses the dataset bias, works such as Marginal Tail-Adaptive Normalizing Flows aim at specifying an inductive bias. Lastly, I introduce the novel concept of generator bias, which is used to perform model-attribution as shown in my recent ICML publication Single-Model Attribution of Generative Models Through Final-Layer Inversion. You can check out all the details in my thesis, which will be made public soon. |
| May 25, 2024 | AnomalyDINO is out! Our latest project delves into the high-quality features of DINOv2 and how these can be adapted for one- or few-shot visual anomaly detection. AnomalyDINO leverages patch similarities and, despite being methodologically simple and training-free, achieves state-of-the-art performance across various settings. This work is the product of an amazing collaboration with Simon and our supervisors Johannes and Asja. A big shoutout to all of you! You can find a preprint of our work on arXiv. |
| May 3, 2024 | Exciting news! Our paper, “Single-Model Attribution of Generative Models Through Final-Layer Inversion”, got accepted at ICML 2024 in Vienna! In this work, we address model attribution of generative models by leveraging modern anomaly detection techniques. To improve the performance further, we combine the latter with features extracted by final-layer inversion. Please find our preprint on arXiv. Thanks for the collaboration, Jonas! |
| Apr 10, 2024 | I am happy to share that our paper, “Benchmarking the Fairness of Image Upsampling Methods”, has been accepted at ACM FAccT 2024. 🇧🇷 In this work, we introduce a comprehensive framework for benchmarking the performance and fairness of conditional generative models. We explore the specific application of image upsampling and create a benchmark covering a wide variety of modern upsampling methods. Check out our preprint on arXiv. Thanks to my awesome collaborators at ETH Zurich, UHH, and RUB! |