Question about the model conditioning design

[Shi-Qi-Li]

Hi @suniique
Thanks for your great work and for open-sourcing this code!

I have a question regarding the model design of the RPF. I noticed that the input to the flow model consists of the point features of unposed parts concatenated with noise along the feature dimension.

I am curious why you chose to concatenate these features rather than injecting them as common conditional signals using AdaLN in the DiT block, which seems to be the standard design for conditional generation. Is there a specific intuition or advantage behind this design choice?

Any response is appreciated!