Not a new concept, I remember years ago reading about the idea that language based image generation is the ultimate compression: e.g. a seal with a ball on his nose performing (ironically a few words is worth a million pixels).
Anyway, I'd be interested to know what a zip format that was allowed to have gigabytes of overhead would look like as a comparison. Like if you could analyze some stereotype of a large photo set to find the most common pixel sequences and then build a giant hash table (or some smarter version of this) so that you could encode many general images in very little data. That would be the fair comparison.
Not a new concept, I remember years ago reading about the idea that language based image generation is the ultimate compression: e.g. a seal with a ball on his nose performing (ironically a few words is worth a million pixels).
Anyway, I'd be interested to know what a zip format that was allowed to have gigabytes of overhead would look like as a comparison. Like if you could analyze some stereotype of a large photo set to find the most common pixel sequences and then build a giant hash table (or some smarter version of this) so that you could encode many general images in very little data. That would be the fair comparison.