code is inconsistent with your paper

[pengzhangzhi]

diffab/diffab/modules/diffusion/transition.py

Line 98 in 6febe02

def add_noise(self, v_0, mask_generate, t):

Hi. In the forward diffusion process, given e_0, the initial rotations, this function generates the noised rotations by R_noisy = E_scaled @ R0_scaled. where E_scaled is sampled from a prior distribution. I wonder why e_normal is not used but declared?

diffab/diffab/modules/diffusion/transition.py

Line 112 in 6febe02

e_normal = e_scaled / (c1 + 1e-8)

In your paper, the R_noisy should be the interpolation between c0 * v_0(scaled v_0) and c1 * e_scaled. But your code does not use c1 at all.

[luost26]

I named the variable scaled because e_scaled is already scaled by c1. The scaling factor c1 is set as the std of the distribution.

diffab/diffab/modules/diffusion/transition.py

Line 90 in 6febe02

self.angular_distrib_fwd = ApproxAngularDistribution(c1.tolist(), **angular_distrib_fwd_opt)

e_normal is for debugging. Just to check whether it comes back to a standard normal after division.

[pengzhangzhi]

My bad! Thank you soooo much.

[pengzhangzhi]

@luost26 , Hi. I have another question regarding the posterior formula of orientation

Why the mean of the posterior is the predicted O_0 instead of the interpolation between O_t and predicted O_0. The later one is more popular in Diffusion Model, including original DDPM.