They have lots of other resources, but this section is focusing solely on the 'checkpoints'.
This will limit the search to models only
Generally when downloading, use the pruned models.
They are smaller so you can load them faster in your VRAM, and taking less VRAM means more is available for the generation of the image.
Pruning means that the training data has been removed, this makes no difference when generating an image but makes them them less viable for retraining a new model (Not something most users will need to do).
Also the fp16 models are smaller and only generate marginally worse results, usually not even perceivable, so there is not really a reason not to use them.
Once a model is chosen it has to be added into the model weights folder, then it will show up on the list of available options in App (no restart required).
The SD 2.X models also need a config file to work properly, which can be found on the same download page.
Place this .yaml file in the same model weights folder.