search

atong01 / conditional-flow-matching

TorchCFM: a Conditional Flow Matching library
https://arxiv.org/abs/2302.00482
MIT License
875 stars 60 forks source link

Confusion about the SB-CFM #118

Closed csufangyu closed 3 weeks ago

csufangyu commented 4 weeks ago

Thank you for your contribution, I have some confusion about SB-CFM, I find when I remove OT, the results are better than with OT in my task. As discussed in the issue, I am also a condition-generating task, if the method in the original article still right after removing OT. I can't understand the derivation in your paper, especially Equation 19. It still seems to be correct even if the OT is removed, but not sure, so I'm asking for your advice, looking forward to your reply!

atong01 commented 3 weeks ago

Hi! Generation quality is still theoretically guaranteed without the OT. This is no longer close to a Schrödinger Bridge, but still should match your x0 and x1 distributions just fine.

csufangyu commented 3 weeks ago

thanks very much!