Had a quick skim.

I was really struggling to see the practical advantage of this. Because you can convert different formats to what the model needs easily.

It feels like some strawman are set up. In the world of the paper, people have to painstakingly convert the image to the correct format for the model, and "hand-crafting a model stem for each modality"

But once you have set up a model, and got it working for say .tiff then to work on any other format you can just use ImageMagick or something? Unless you want the meta data too?

I think the use case for a model that works on bytes is as a kind of "easy to install" package on local devices, like a security camera, regular camera etc that will work with whatever the local file format is.

Also, a dollar into the "is/are all you need" jar please.