chenwu98 / unified-generative-zoo Goto Github PK

In the line100 in langevin_dynamics.py, the z is updated according to the Langevin dynamics:
z = (z - self.step_size / 2 * grad + noise).detach().
The variable `grad' in code is computed as $\bigtriangledown_z -E(G(z)|C)$.

Refer to Sec 3.4 【UNIFIED PLUG-AND-PLAY GUIDANCE FOR GENERATIVE MODELS】 in paper 【UNIFYING DIFFUSION MODELS’ LATENT SPACE, WITH APPLICATIONS TO CYCLEDIFFUSION AND GUIDANCE】, the update of z should involve the term $\bigtriangledown_z \log(p_z(z^k))$.

I am wondering why this term can be omitted in practice or which codes compute this term. Can you give me some hints? Thank you for your great work and your kindly reply.