> Is that actually better (in practice, not in terms of entropy) than /dev/urandom?

It offers protection in the event that your /dev/urandom is compromised. Otherwise no.

(Of course, if your /dev/urandom is compromised then whatever process you use to compute a hash of a photo is likely compromised as well.)

If you're concerned about that, you can concat your JPEG with a few bytes from /dev/random and you'll get the security of whichever is stronger. In practice none of this will be your weakest link.

It is possible that using a JPEG with its known bytes could make the final effort weaker than if you just used /dev/random.

No, that would fundamentally break the way (good) hashing works if I could add known data and get it to reveal the secret data