Untitled

However, this direct connection between DDPMs and numerical methods (e.g., forward Euler method, linear multi-step method and RungeKutta method (Timothy, 2017)) has weaknesses in both speed and effect (see Section 3.1).

Furthermore, we also notice that numerical methods can introduce noticeable noise at a high speedup rate, which makes high-order numerical methods (e.g., Runge-Kutta method) even less effective than DDIMs.

To tackle these problems, we design new numerical methods called pseudo numerical methods for diffusion models (PNDMs) to generate samples along a specific manifold in R n, which is the highdensity region of the data. We first compute the corresponding differential equations of diffusion models directly and self-consistently, which builds a theoretical connection between DDPMs and numerical methods. Considering that classical numerical methods cannot guarantee to generate samples on certain manifolds, we provide brand-new numerical methods called pseudo numerical methods based on our theoretical analyses. We also find that DDIMs are simple cases of pseudo numerical methods, which means that we also provide a new way to understand DDIMs better. Furthermore, we find that the pseudo linear multi-step method is the fastest method for diffusion models under similar generated quality.

BACKGROUND

STOCHASTIC DIFFERENTIAL EQUATION

there is another understanding of DDPMs. The diffusion process can be treated as solving a certain stochastic differential equation dx = (p p 1 − β(t) − 1)x(t)dt + β(t)dw.

Untitled

This is Variance Preserving stochastic differential equations (VP-SDEs). Here, we change the domain of t from [1, N] to [0, 1]. When N tends to infinity, {βi} N i=1, {xi} N i=1 become continuous functions β(t) and x(t) on [0, 1]. Song et al. (2020b) also show that this equation has an ordinary differential equation (ODE) version with the same marginal probability density as Equation (4):

Untitled

NUMERICAL METHOD

Many classical numerical methods can be used to solve ODEs, including the forward Euler method, Runge-Kutta method and linear multi-step method (Timothy, 2017).

Untitled

1 PSEUDO NUMERICAL METHOD FOR DDPM

FORMULA TRANSFORMATION

the reverse process of DDPMs and DDIMs satisfies:

Untitled

σt controls the ratio of random noise. If σt equals one, Equation (8) represents the reverse process of DDPMs; if σt equals zero, this equation represents the reverse process of DDIMs. And only when σt equals zero, this equation removes the random item and becomes a discrete form of a certain ODE.

Therefore, our work concentrate on the case σt equals zero.

To find the corresponding ODE of Equation (8), we replace discrete t − 1 with a continuous version t − δ according to (Song et al., 2020a) and change this equation into a differential form, namely, subtract xt from both sides of this equation: