Hi, thank you for releasing the code. I was reading the released code and trying to map it to the paper’s constrained generation modules. From what I can see, the model has a flow/trajectory head that predicts v_pred, which corresponds to the base flow-matching velocity field. I also see CF-like anchor/state replacement during sampling, and in V10 I see ESDF/energy conditioning that looks related to RFE. However, I could not clearly find the paper’s CVF velocity-field correction and the explicit RFE objective L_RFE = E_theta(x^(1)) - E_theta(x_1) in the training code. Could you please point me to where CVF, CF, and RFE are implemented, or clarify if some of these modules are omitted or only partially included in the public release?
Hi, thank you for releasing the code. I was reading the released code and trying to map it to the paper’s constrained generation modules. From what I can see, the model has a flow/trajectory head that predicts v_pred, which corresponds to the base flow-matching velocity field. I also see CF-like anchor/state replacement during sampling, and in V10 I see ESDF/energy conditioning that looks related to RFE. However, I could not clearly find the paper’s CVF velocity-field correction and the explicit RFE objective L_RFE = E_theta(x^(1)) - E_theta(x_1) in the training code. Could you please point me to where CVF, CF, and RFE are implemented, or clarify if some of these modules are omitted or only partially included in the public release?