https://wiki.mozilla.org/Identity/CryptoIdeas/01-PBKDF-scryp...
But unless you have a crazy long passphrase, you're not going to get 128 bits, let alone 256.
As I explained in a late edit to my comment, this is distinct from your "password-derived key" case because it eliminates the major drawback of that case: "This is not as secure as the previous setting, since most user passwords are not nearly as strong as full-strength crypto keys." If you generate high-entropy passphrases, that problem goes away.
128 bits is overkill for defense against brute force. You can do maybe 2³⁰ crypto operations per second in a custom crypto-cracking processor, pack maybe 2²⁰ of them onto a custom chip, use maybe 2²⁰ custom chips in your Crypto Cracking Center in your evil genius volcano base, and let it run for maybe a year, 2²⁵ seconds, before you get bored. That's still only enough to search 2⁹⁵ keys, so you should be pretty safe with keys that need 2¹⁰⁰ operations to crack, at least for a few years. Or, if you don't have a supervillain or major intelligence agency devoting their full computational resources to reading your browser bookmarks, 2⁸⁰.
I do think it's actually feasible for someone to memorize an 11-word phrase encoding a 128-bit key, but it would take several minutes and careful practice over the next few weeks to be sure they didn't forget it, and using a decent PBKDF with a 7- or 8-word passphrase is probably a better option.
http://lists.canonical.org/pipermail/kragen-hacks/2012-April... demonstrates encoding an 80-bit random number (which is plenty secure with a reasonable key derivation function) as each of "point pleased intense de maybe fairly arms", "bejuso jejigi nububi bidoda gahano", "ADD DOTE BID HILT LAUD MAIN CALF CITY", and "仴薦肨縨猯鹽", any of which is eminently practical to memorize. I use this program to generate my login passwords these days.
(80 bits is not enough for a key for something like AES, because you can try a lot of different keys per second. It's plenty if you have a decent key derivation function to add a 25–35-bit work factor.)
This is different from a user-chosen password because users are often highly predictable in their password choice.