Published on 17.09.2026
Tool Spotlight: GloViTa
Got a pile of images, scans, or video clips and a question like “which category is this?” or “how severe is this?”, but no interest in building a training pipeline from scratch? That’s the gap GloViTa is meant to fill.
Developed by the Engineering and Support Unit at DKFZ, GloViTa is a reusable framework for training deep learning models on classification and regression tasks — and it doesn’t care what field the images come from. Microscopy, satellite imagery, medical scans, wildlife camera traps, industrial inspection photos, pathology slides: as long as the task is “look at this and predict something about it,” the same framework applies. It supports 2D images, 3D volumes, video, and bag-of-patches setups like whole-slide pathology images. Instead of writing a new script for every project, you set a few basic options — where your data lives, how many classes you’re predicting, which pretrained model to start from — and run one command. It takes care of loading the data, training, and validating, and it automatically saves the exact settings used for each run, so results stay traceable later on.
It’s aimed at people who already have some comfort with running command-line tools and adapting settings, but don’t want to maintain custom training code for every new dataset.
It also doubles as a solid experimentation suite: because every setting is exposed the same simple way, trying out different learning rates, augmentation strengths, or model sizes is just a matter of changing one option and rerunning, rather than editing code each time. A range of regularization and finetuning techniques — from lighter adaptation methods that only touch part of a pretrained model, to full finetuning — come built in, so comparing approaches on your own data doesn’t require implementing any of them yourself first.
A solid starting point if imaging AI keeps showing up in your work, regardless of domain, and you’d rather not reinvent the training loop each time.
Apache-2.0 licensed.