Diffusion Sampler Theory: ODE Solvers and Acceleration

Module: Generative AI | Difficulty: Advanced

Probability Flow ODE

Shares marginals with the SDE but is deterministic.

DDIM

DPM-Solver++ (2nd Order)

Consistency Models

class DPM_Solver:
    def __init__(self, model, alphas):
        self.model, self.alphas = model, alphas
    def step(self, x_t, t_i, t_next):
        a_i, a_n = self.alphas[t_i], self.alphas[t_next]
        s_i, s_n = (1-a_i).sqrt(), (1-a_n).sqrt()
        eps = self.model(x_t, t_i)
        x0 = (x_t - s_i*eps)/a_i
        return a_n*x0 + s_n*eps