Yeah, this smells very fishy to me. It's almost trivially easy to gather a small validation dataset in humans for the paper. At my institution, it's about $800 an hour to scan someone. You can probably get enough data to validate the model with a half hour scan. Surely the group has enough grant funding to pop a few healthy controls in the scanner.

I haven't looked in super close detail to the paper, but their methods section says that they fit a video model (V-JEPA2) to the fMRI dataset in a voxelwise ridge regression, meaning that the baked in assumption is that the visual response affects each voxel independently. Voxelwise models are very nice for making statistical inferences, but are less good for prediction and modelling tasks, because our brains certainly do not work as collections of independent regions.

BOLD is intensely messy data, and their design is far too simple IMO to reflect anything of reality.