duplicate detecton works by comparing MD5 hashes (which happen to become the filenames). If a picture is even one pixel off, or contains extra meta-data it won't be detected as a duplicate. Visual similarity is a much harder problem than exact bit-for-bit duplicate detection.
This one is smaller in terms of filesize. If you compare them the higher compression on this one is quite noticeable, look at the edges of Nanoha's hair for example.
Interesting, if I put the two duplicates into photoshop, calculate the difference, and play with the colors & contrast, I can get some halfway decent sketchy line-art. It's pretty cool. I guess it just means that lossy compression like JPG wreaks the most havoc with the edges, but still, it's interesting.
