The problem that nobody will take a position on is this:
Is the metadata on an image file, or on an image?
This comes back to the more general question:
Does the LSID identify a digital instance of an object (database record, binary file), or does the LSID identify the "abstract" object (specimen, image, etc.) that the digital object serves as surrogate for?
In the context of images, my thinking is this:
The "abstract" object is the set of photons that struck a planar surface inside a camera (over a given period of time) after passing through a series of lenses. From this object, there might be a derivative physical object (e.g., a frame of celluloid film, which might have its own set of properties such as film type, etc.), and there might be derivative digital object (e.g., a binary RAW file, or whatever "most original" set of 1s and 0s extracted from the camera representing a digitally interpreted version of the set of photons that struck the planar surface -- which I henceforth refer to as "RAW" file for simplicity sake -- even though it might actually be a TIFF or a JPEG or a NEF or whatever). Either of these two derivative objects (the celluloid or the "RAW" file) might have derivatives of their own (e.g., dupes, prints, and scans for the celluloid object; crops, color corrections, resizes, other digital image formats for the "RAW" file).
I see a world where each of these things (at least the ones that have value in terms of information presentation) gets its own LSID, and part of the metadata for each primary and subsequent derivative would be a pointer to the LSID that identifies the "top-level" non-physical, non-digital "abstract" image object (i.e., set of photons striking the planar surface).
Once at that stage, the big questions become:
1) What metadata from derived objects gets inherited "upstream" to the "master" abstract LSID;
2) What metadata from the "master" object gets inherited "downstream" to the various derivatives; and
3) Which of the LSIDs from the various derivatives become incorporated into the metadata of the "master" abstract LSID?
Number 3 above is slightly different from number 1, in that number 1 is more about metadata content inheritance, whereas #3 is more about cross-linking among various LSIDs.
One would assume that all LSIDs of derived image-objects would have as part of their metadata pointers to the "master" abstract LSID from which the former were derived.
In this model, it becomes relatively straightforward which LSIDs get which metadata.
The problem, of course, is how to structure the inheritance/flow of metadata from one LSID to another other (i.e., "master" to derived, and vice versa).
Just some random thoughts....
Rich