DFKI
↗ Website
LUMA is a lightweight, *backbone-agnostic* mask-transformer head that lets us fairly compare segmentation backbones by fixing the decoder. Benchmarking 20 backbones and 11 pretraining schemes, we find that "efficient" token mixers don't actually deliver efficiency and that the pretraining objective, not the architecture, governs segmentation quality.
CutMix assigns labels by patch area, not by visible object content, a systematic bias that mislabels 21.5% of samples and creates ghost labels in 17%. OA-CutMix replaces the label with one derived from object area, leaving the image mixing unchanged. It matches or beats 10+ static and dynamic mixing methods across 4 architectures and 6 datasets.
We show that newer text-to-image models are progressively worse as training data generators, despite better visual quality, because they collapse to a narrow aesthetic-centric distribution that diverges from real data.
We use a frozen text encoder on image captions as a lightweight training-time auxiliary objective for image classifiers. The text components are dropped at inference, leaving a fast, unimodal vision model. Accuracy on ImageNet improves by up to +2.7 p.p. and downstream transfer by +1.0 p.p. on average, outperforming vision knowledge distillation at a fraction of the compute.
We fix Forward-Forward's slow inference by replacing per-class passes with a single forward pass through hyperspherical prototype matching. Thus, we achieve 40× faster inference with competitive accuracy.
We introduce PRISM, a framework that disentangles architectural priors for dataset distillation, outperforming single-teacher setups.
We introduce SubZeroCore, a novel, training-free coreset selection method that integrates submodular coverage and density into a single, unified objective.
We present HyperCore, a lightweight adaptive coreset selection framework designed for noisy environments. HyperCore utilizes per class hypersphere models and adaptively selects pruning thresholds.
We extend pretrained super-resolution models to larger images by using local-aware prompts.
We improve dataset distillation by distilling only a representative coreset.
Image classification datasets carry compositional biases (objects are centered, at a typical size, on class-specific backgrounds) that models lean on and then break under distribution shift. ForAug segments and recombines foregrounds and backgrounds to break these correlations, making models across 10 architectures more accurate and more robust.
We conduct the first systematic study of dataset distillation for Super-Resolution.
We speed up diffusion classifiers by utilizing a label hierarchy and pruning unrelated paths.
We utilize the TaylorShift attention mechanism for global pixel-wise-attention in image super-resolution.