You're ignoring half of my post (on purpose?): Now generate the collisions matching file sizes. Even that "under a second" concept relies in tiny files.

As files get larger matching both the MD5 and file size becomes more costly.