There are two things here: reversing hash to "a message", and reversing a hash to "the message". Depending on "the message", it is way more impossible than just "a message".
The point pooya is making is that finding a collision is not the same as reversing the process. The input to SHA256 can be any message up to around 2 million terabytes. The output will always be 32 bytes. Therefore there are potentially million of messages which will give the same output. Even if you could search the entire space and find all the inputs which match a hash, you still haven't reversed the process.