Why self.condition_mean is put after self.q_sample during training?

[CDLCHOI]

The IK guidance during generation process is clear. But I'm confused why the training process is like below:

x_t = self.q_sample(x_start, t, noise=noise) x_t = self.condition_mean_bfgs(x_t, num_condition)

Could you explain why you put self.condition_mean_bfgs here? Is that a theoretical approach or empirical approach?

Wait for your reply. Thanks

[zhenzhiwang]

It is an empirical approach. It is the implementation of the formula of $\mu$ (posterior) in equation (2) of our paper. If we remove this, it is the implementation of $x_0$ (clean motion).