I love the sound of my Kurzweil hardware synth piano and it is way better than most VST's as well. Firstly I am referring to a Kurzweil hardware synth and Studio One but the principles are the same.
I actually do it the other way around.
Playing the piano the long way around viw your DAW by using Local/OFF gives you the poorest latency setting. I prefer to leave Local/ON and allow the piano to be played from its own keyboard internally. It gives the fastest response that way while you are playing stuff in and that for me is more important. You should be able to send that data into the DAW and record the data onto the midi track at the same time. The only thing here is you need to
switch OFF any form of input monitoring on your midi track while recording. Then the piano won't be triggered via midi at all. You should only hear the piano sound engine being played via Local/ON.
Once recorded instant playback is possible as well. Because then the midi track playing back will trigger the piano as you want to hear. Playing along with that might be tricky as notes are being used up etc.
As long as any form of midi input monitoring is OFF then this process works rather well too. And you get to leave Local/ON in your piano too which is actually handy and better. You will never be caught out live wondering why your piano does not work!
It is important though if you are using external hardware synths then you get your midi timing DAW round trip latency as fast as possible too. It should not be related to the audio buffer latency setting either. Midi timing should be an independent parameter. It certainly is for me in Studio One. I have got a Steinberg Midex 8 interface running on its own USB port (and nothing else) from a PCI based USB card. (Under Win 7) The midi timing in record and playback should not be effected by how hard the audio side of the DAW is working as well. It can in some programs. What you can do though is investigate the options you may have in regard to shifting midi timing overall in relation to audio timing.
Then there is the option of advancing midi tracks in time to compensate for slower responding synths or presets. But this tends to work better for quantised material.