Ideally images and videos are transcoded to optimal format and then several copies are created with different resolution etc. And that could be done on client or server. And if on server then there is trust involved. And then appropriate version of file has to be served depending on client needs. And then it should be content-addressable, with as little trust to third parties as possible. That's just what comes to mind, and it's already much wider scope than what blossom covers. I am not saying nip96 covers it all, but it's authors are trying to reconcile all this. Seems like a good effort to me. 

 I don't think the modified ones need to be content-addressable, just the originals. Imgproxy does this on-demand for images, so I don't see why you couldn't just use that or a service like that as a front-end for blossom. This would allow you to trade-off referential transparency for performance depending on reader use case. Of course, there are also many files that publishers don't care about censorship resistance and verifiability for, but in that case I would just use a traditional image host. 

 And if you want referential transparency for a transcoded one, then transcode it and then upload it to blossom