Once you need a custom client that's programmed to stitch together arbitrary bytes stored across N transactions to reproduce a bad image, it doesn't seem to be much different or any less savory than writing one that can find the same bytes from specific offsets across N articles on nytimes.com or one that plucks them out of a stream from /dev/random.
The inaccessibility of viewing this data seems to make the world of difference versus, say, looking at the images that appear in your web browser when you go to badsite.com.