scale factor b

[zhangdan8962]

@chentingpc You mentioned scale the input image x by factor b when we calculate the m_bits in Algorithm 1. I am wondering if we need to scale it back when we predict m_0 from m_t during sampling because it is not so obvious in your Algorithm 2. Did you actually implement it in DDIM sampler? I have similar question regarding your another paper On the Importance of Noise Scheduling for Diffusion Models.

If such inverted scaling is not necessary, could you please explain the reason behind this?