The way OpenML Pytorch works is by using the OpenMl python extension API. The latter relies on a method called run_model_on_fold, which requires any other extensions to pass their information through this function. In essence, this means it is not currently possible to just have a "minimal" extension, where you can have any training pipeline and it "just works" with OpenML.
To counteract this, we have created two kinds of examples, one where you can ignore this limitation and just use data from OpenML along with your own pipelines example1 and the rest of them where you can use the API we provide to train your models and upload them to OpenML example2.
OpenML so far does not focus on image datasets, and as such, they are stored as a header file (think of it as a table with file names and say categories in the case of classification). For some datasets, OpenML provides a way to download the image folders directly, but this is not always the case. This will eventually be replaced by something nicer, but for now, this is the way it is.