The issue of inconsistency between the calculation of x_prev, a.k.a. x_t-1, in the p-sample-ddim_deditional function in the code and the 5th line of Algorithm 2 in the paper

[csjunjun]

May I ask where the code corresponding to solving x_t-1 in line 5 of Algorithm 2 is located in the p_sample_ddim_conditional?

From the paper we know that the coefficient of the x_t term in line 5 of Algorithm 2 is (1+0.5*beta), but I calculated this coefficient using the code, a.k.a the coefficient for the variable x_prev is 1/sqrt(beta), which is inconsistent.

Thank you very much!

[JamieCR1999]

I have the same question, here is the code from ddim.py.
Is c3 here equal to the hyperparameter beta? It seems like a case of variance in DDIM.

[HolmesShuan]

@SerJamie @csjunjun The code in ddim.py is based on the derivation of DPS from the paper “Diffusion Posterior Sampling for General Noisy Inverse Problems”. The definition of the hyperparameter c3 can be found in the official DPS code, where its connection between DDIM and DDPM is established.

Additionally, you could consider treating “Freedom” as replacing measurement models (the red line in the DPS Algorithm) with losses related to style, face ID, etc., while keeping the rest of the pipeline intact.