The original article has the correct images:

  $ curl -s http://www.fishtrap.co.uk/black.jpg.coll | md5
  b69dd1fd1254868b6e0bb8ed9fe7ecad
  $ curl -s http://www.fishtrap.co.uk/brown.jpg.coll | md5
  b69dd1fd1254868b6e0bb8ed9fe7ecad
  $ curl -s http://www.fishtrap.co.uk/white.jpg.coll | md5
  b69dd1fd1254868b6e0bb8ed9fe7ecad

Good catch. CloudFlare's optimizations are often too good for their own good. Fixing.

If you download the files to disk, you'll get the right values:

    $ curl -s https://blog.cloudflare.com/content/images/2015/08/white.jpg > file && md5 file                                                 
    MD5 (file) = b69dd1fd1254868b6e0bb8ed9fe7ecad

I've seen this same sort of thing happen with curl in other contexts, but I've never tracked down the details. I assume it has something to do with file stream handling in certain versions of curl. You'll see a discrepancy if you look at the output of piping to 'wc' vs. examining the downloaded file.

I ran the same curl command twice in a row and got files with different sizes! I think it actually does have to do with Cloudflare rather than with curl in this case.

Apparently sometimes the CDN is stripping the EXIF data