Tim Kaiser

Nikolas Adaloglou, Tim Kaiser, Felix Michels et al.

Diffusion-based image generation models can enhance image quality when conditioned on ground truth labels. Here, we conduct a comprehensive experimental study on image-level conditioning for diffusion models using cluster assignments. We investigate how individual clustering determinants, such as the number of clusters and the clustering method, impact image synthesis across three different datasets. Given the optimal number of clusters with respect to image synthesis, we show that cluster-conditioning can achieve state-of-the-art performance, with an FID of 1.67 for CIFAR10 and 2.17 for CIFAR100, along with a strong increase in training sample efficiency. We further propose a novel empirical method to estimate an upper bound for the optimal number of clusters. Unlike existing approaches, we find no significant association between clustering performance and the corresponding cluster-conditional FID scores. The code is available at https://github.com/HHU-MMBS/cedm-official-wavc2025.

Nikolas Adaloglou, Tim Kaiser, Damir Iagudin et al.

Guidance is a widely used technique for diffusion models to enhance sample quality. Technically, guidance is realised by using an auxiliary model that generalises more broadly than the primary model. Using a 2D toy example, we first show that it is highly beneficial when the auxiliary model exhibits similar but stronger generalisation errors than the primary model. Based on this insight, we introduce \emph{masked sliding window guidance (M-SWG)}, a novel, training-free method. M-SWG upweights long-range spatial dependencies by guiding the primary model with itself by selectively restricting its receptive field. M-SWG requires neither access to model weights from previous iterations, additional training, nor class conditioning. M-SWG achieves a superior Inception score (IS) compared to previous state-of-the-art training-free approaches, without introducing sample oversaturation. In conjunction with existing guidance methods, M-SWG reaches state-of-the-art Frechet DINOv2 distance on ImageNet using EDM2-XXL and DiT-XL. The code is available at https://github.com/HHU-MMBS/swg_bmvc2025_official.

Tim Kaiser, Markus Kollmann

Standard evaluation metrics suggest that Denoising Diffusion Models based on U-Net or Transformer architectures generalize well in practice. However, as it can be shown that an optimal Diffusion Model fully memorizes the training data, the model error determines generalization. Here, we show that although sufficiently large denoiser models show increasing memorization of the training set with increasing training time, the resulting denoising trajectories do not follow this trend. Our experiments indicate that the reason for this observation is rooted in the fact that overfitting occurs at intermediate noise levels, but the distribution of noisy training data at these noise levels has little overlap with denoising trajectories during inference. To gain more insight, we make use of a 2D toy diffusion model to show that overfitting at intermediate noise levels is largely determined by model error and the density of the data support. While the optimal denoising flow field localizes sharply around training samples, sufficient model error or dense support on the data manifold suppresses exact recall, yielding a smooth, generalizing flow field. To further support our results, we investigate how several factors, such as training time, model size, dataset size, condition granularity, and diffusion guidance, influence generalization behavior.

Nikolas Adaloglou, Felix Michels, Tim Kaiser et al.

We present a comprehensive experimental study on pretrained feature extractors for visual out-of-distribution (OOD) detection, focusing on adapting contrastive language-image pretrained (CLIP) models. Without fine-tuning on the training data, we are able to establish a positive correlation ($R^2\geq0.92$) between in-distribution classification and unsupervised OOD detection for CLIP models in $4$ benchmarks. We further propose a new simple and scalable method called \textit{pseudo-label probing} (PLP) that adapts vision-language models for OOD detection. Given a set of label names of the training set, PLP trains a linear layer using the pseudo-labels derived from the text encoder of CLIP. To test the OOD detection robustness of pretrained models, we develop a novel feature-based adversarial OOD data manipulation approach to create adversarial samples. Intriguingly, we show that (i) PLP outperforms the previous state-of-the-art \citep{ming2022mcm} on all $5$ large-scale benchmarks based on ImageNet, specifically by an average AUROC gain of 3.4\% using the largest CLIP model (ViT-G), (ii) we show that linear probing outperforms fine-tuning by large margins for CLIP architectures (i.e. CLIP ViT-H achieves a mean gain of 7.3\% AUROC on average on all ImageNet-based benchmarks), and (iii) billion-parameter CLIP models still fail at detecting adversarially manipulated OOD images. The code and adversarially created datasets will be made publicly available.