It's not the JPG data you can copyright, but the recognisable image it produces. Every time you re-save or resize the image, the data changes, but the recognizable image remains.

It's not clear how this process applies to model weights. Once you run another training epoch on them, the data has changed. What is the essential copyrightable, trademarkable or patentable thing that remains? A legally untested question for sure.

It depends on how the JPG is made. Some JPGs are not copy writable, like that picture a monkey took of itself. Model weights are probably (legally) more like a selfie of a monkey than a photographer's photo.

Couldn't you just transform the weights in some predictable way and then untransform them before use in runtime? The weights at rest would be completely distinct from the meta weights.

But the data inside the jpg isn't generated by a machine, model weights are.