Consistency Models

Yang Song’ Prafulla Dhariwal1 Mark Chen’ Ilya Sutskever1

Abstract

Diffusion models have made significant breakthroughs in image,audio,and video generation, but they depend on an iterative generation process that causes slow sampling speed and caps their potential for real-time applications.To overcome this limitation,we propose consistency models, a new family of generative models that achieve high sample quality without adversarial training. They support fast one-step generation by design, while still allowing for few-step sampling to trade compute for sample quality. They also support zero-shot data editing,like image inpainting, colorization,and super-resolution,without requiring explicit training on these tasks. Consistency models can be trained either as a way to distill pre-trained diffusion models,or as standalone generative models. Through extensive experiments, we demonstrate that they outperform existing distillation techniques for diffusion models in oneand few-step generation.For example,we achieve the new state-of-the-art FID of 3.55 on CIFAR-10 and 6.20 on ImageNet 64 × 64 for one-step generation. When trained as standalone generative models, consistency models also outperform single-step,non-adversarial generative models on standard benchmarks like CIFAR-10, ImageNet 64 × 64 and LSUN 256 × 256.

1. Introduction

Diffusion models (Sohl-Dickstein et al., 2O15; Song & Ermon, 2019; 2020; Ho et al., 2020; Song et al., 2021), also known as score-based generative models, have achieved unprecedented success across multiple fields, including image generation (Dhariwal & Nichol, 2021; Nichol et al., 2021; Ramesh et al., 2022; Saharia et al., 2022; Rombach et al.,2022),audio synthesis (Kong et al., 202O; Chen et al., 2021; Popov et al., 2O21),and video generation (Ho et al., 2022b;a).Unlike Generative Adversarial Networks (GANs,

Figure 1: Given a Probability Flow (PF) ODE that smoothly converts data to noise, we learn to map any point (e.g., Xt, Xt,and xT) on the ODE trajectory to its origin (e.g., Xo) for generative modeling. Models of these mappings are called consistency models,as their outputs are trained to be consistent for points on the same trajectory.

Goodfellow et al. (2O14)), these models do not rely on adversarial training and are thus less prone to issues such as unstable training and mode collapse.Additionally, diffusion models do not impose the same strict constraints on model architectures as in autoregressive models (Bengio & Bengio, 1999; Uria et al., 2013; 2016; Van Den Oord et al., 2016), variational autoencoders (VAEs, Kingma & Welling (2014); Rezende et al. (2O14)), or normalizing flows (Dinh et al., 2015; 2017; Kingma & Dhariwal, 2018).

Key to the success of diffusion models is their iterative sampling process which progressively removes noise from a random noise vector. This iterative refinement procedure repetitively evaluates the diffusion model,allowing for the trade-off of compute for sample quality: by using extra compute for more iterations,a small-sized model can unroll into a larger computational graph and generate higher quality samples. Iterative generation is also crucial for the zeroshot data editing capabilities of diffusion models, enabling them to solve challenging inverse problems ranging from image inpainting,colorization,stroke-guided image editing, to Computed Tomography and Magnetic Resonance Imaging (Song & Ermon,2019; Song et al., 2021; 2022; 2023; Kawar et al., 2021; 2022; Chung et al., 2023; Meng et al., 2021). However, compared to single-step generative models like GANs, VAEs,and normalizing flows, the iterative generation procedure of diffusion models typically requires 10-2000 times more compute (Song & Ermon, 2020; Ho et al., 2020; Song et al., 2021; Zhang & Chen, 2022; Lu et al.,2O22), causing slow inference and limiting their real-

致终身学习者社群

行业报告/思维导图/电子书/资讯情报

收钱文案、增长黑客、产品运营、品牌企划、营销战略、办公软件、会计财务、广告设计、摄影修图、视频剪辑、直播带货、电商运营、投资理财、汽车房产、餐饮烹饪、职场经验、演讲口才、风水命理、心理思维、恋爱情趣、美妆护肤、健身瘦身、格斗搏击、漫画手绘、声乐训练、自媒体打造、效率软件工具、游戏影音.....

扫码先加好友，以备不时之需

time applications.

Our objective is to create generative models that facilitate efficient, single-step generation without sacrificing important advantages of iterative refinement. These advantages include the ability to trade-off compute for sample quality when necessary,as well as the capability to perform zeroshot data editing tasks.As illustrated in Fig.1,we build on top of the probability flow (PF) ordinary differential equation (ODE) in continuous-time diffusion models (Song et al.,2O21),whose trajectories smoothly transition the data distribution into a tractable noise distribution. We propose to learn a model that maps any point at any time step to the trajectory's starting point.A notable property of our model is self-consistency: points on the same trajectory map to the same initial point.We therefore refer to such models as consistency models. Consistency models allow us to generate data samples (initial points of ODE trajectories, e.g.,Xo in Fig.1) by converting random noise vectors (endpoints of ODE trajectories, e.g., XT in Fig.1) with only one network evaluation. Importantly,by chaining the outputs of consistency models at multiple time steps,we can improve sample quality and perform zero-shot data editing at the cost of more compute, similar to what iterative refinement enables for diffusion models.

To train a consistency model, we offer two methods based on enforcing the self-consistency property. The first method relies on using numerical ODE solvers and a pre-trained diffusion model to generate pairs of adjacent points on a PF ODE trajectory. By minimizing the difference between model outputs for these pairs,we can effectively distill a diffusion model into a consistency model, which allows generating high-quality samples with one network evaluation. By contrast, our second method eliminates the need for a pre-trained diffusion model altogether,allowing us to train a consistency model in isolation.This approach situates consistency models as an independent family of generative models.Crucially,neither approach requires adversarial training,and both training methods permit flexible neural network architectures for consistency models.

We demonstrate the efficacy of consistency models on several challenging image benchmarks, including CIFAR-10 (Krizhevsky et al., 2009), ImageNet 64 × 64 (Deng et al., 2009),and LSUN 256 × 256 (Yu et al., 2015). Empirically,we observe that as a distillation approach,consistency models outperform progressive distillation(Salimans & Ho,2O22) across a variety of datasets and number of sampling steps. On CIFAR-10,consistency models reach new state-of-the-art FIDs of 3.55 and 2.93 for one-step and two-step generation.On ImageNet 64 × 64,it achieves record-breaking FIDs of 6.2O and 4.7O with one and two network evaluations respectively.When trained as standalone generative models,consistency models achieve comparable performance to progressive distillation for single step generation, despite having no access to pre-trained diffusion models.They are able to outperform many GANs,and all other non-adversarial, single-step generative models across multiple datasets.We also show that consistency models can be used to perform zero-shot data editing tasks,including image denoising, interpolation, inpainting, colorization, super-resolution, and stroke-guided image editing (SDEdit, Meng et al. (2021)).

2. Diffusion Models

Consistency models are heavily inspired by the theory of (continuous-time) diffusion models (Song et al., 2O21).Diffusion models generate data by progressively perturbing data to noise via Gaussian perturbations,then creating samples from noise via sequential denoising steps. Let pdata (x) denote the data distribution. Diffusion models start by diffusing Pdata(x) with a stochastic differential equation (SDE) (Song et al., 2021; Karras et al., 2022)

$$\tag{1}$$

where t ∈ [O,T],T > O is a fixed constant,μ(·,·) and g(·） are the drift and diffusion coefficients respectively, and {wt}te[o,T] denotes the standard Brownian motion. We denote the distribution of Xt as Pt(x) and as a result Po(x)= Pdata(x).A remarkable property of this SDE is the existence of an ordinary differential equation (ODE), dubbed the Probability Flow (PF) ODE by Song et al. (2021),whose solution trajectories sampled at t are distributed according to pt (x):

$$\tag{2}$$

Here V log pt(x) is the score function of pt(x); hence diffusion models are also known as score-based generative models (Song & Ermon, 2019; 2020; Song et al., 2021).

Typically, the SDE in Eq. (1) is designed such that pr(x) is close to a tractable Gaussian distribution π(x). We hereafter adopt the configurations in Karras et al. (2O22), who set μ(x,t) = O and σ(t) = √2t. In this case,we have Pt(x)= Pdata(x) N(0,t²I),where 区 denotes the convolution operation,and π(x) = N(O,T²I). For sampling, we first train a score model s𝜙(x,t) ≈ Vlog pt(x) via score matching (Hyvärinen & Dayan, 2005; Vincent, 2011; Song et al., 2019; Song & Ermon, 2019; Ho et al., 2020), then plug it into Eq. (2) to obtain an empirical estimate of the PF ODE,which takes the form of

$$\tag{3}$$

We call Eq. (3) the empirical PF ODE. Next, we sample XT \~ π = N(O,T²I) to initialize the empirical PF ODE and solve it backwards in time with any numerical ODE solver, such as Euler (Song et al.,202O; 2021) and Heun solvers (Karras et al.,2O22),to obtain the solution trajectory {Xt}te[0,T]. The resulting Xo can then be viewed as an approximate sample from the data distribution pdata(x). To avoid numerical instability, one typically stops the solver at t = ∈,where ε is a fixed small positive number,and instead accepts Xe as the approximate sample.Following Karras et al. (2022), we rescale pixel values in images to [-1,1], then set T = 80,and e = 0.002.

Figure 2: Consistency models are trained to map points on any trajectory of the PF ODE to the trajectory's origin.

Diffusion models are bottlenecked by their slow sampling speed. Clearly, using ODE solvers for sampling requires many evaluations of the score model s𝜙(x,t),which is computationally costly. Existing methods for fast sampling include faster numerical ODE solvers (Song et al., 2020; Zhang & Chen, 2022; Lu et al.,2022; Dockhorn et al., 2022), and distillation techniques (Luhman & Luhman, 2O21; Salimans & Ho, 2022; Meng et al., 2022; Zheng et al., 2022). However, ODE solvers still need more than 1O evaluation steps to generate competitive samples.Most distillation methods like Luhman & Luhman (2O21) and Zheng et al. (2022) rely on collecting a large dataset of samples from the diffusion model prior to distillation,which itself is computationally expensive.To our best knowledge, the only distillation approach that does not suffer from this drawback is progressive distillation (PD, Salimans & Ho (2O22)), with which we compare consistency models extensively in our experiments.

3. Consistency Models

We propose consistency models,a new type of generative models that support single-step generation at the core of its design, while still allowing iterative generation for zero-shot data editing and trade-offs between sample quality and compute.Consistency models can be trained in either the distillation mode or the isolation mode.In the former case, consistency models distill the knowledge of pre-trained diffusion models into a single-step sampler, significantly improving other distillation approaches in sample quality, while allowing zero-shot image editing applications. In the latter case,consistency models are trained in isolation, with no dependence on pre-trained diffusion models.This makes them an independent new class of generative models.

Below we introduce the definition, parameterization,and sampling of consistency models,plus a brief discussion on their applications to zero-shot data editing.

Definition Given a solution trajectory {Xt}te[e,T] of the PF ODE in Eq. (2),we define the consistency function as f : (Xt,t) → Xe. A consistency function has the property of self-consistency: its outputs are consistent for arbitrary pairs of (xt,t) that belong to the same PF ODE trajectory, i.e.,f(xt,t) = f(xt,t') for all t,t'∈ [e,T].With fixed time argument, f(·,t) is always an invertible function.As illustrated in Fig.2, the goal of a consistency model, symbolized as fe,is to estimate this consistency function f from data by learning to enforce the self-consistency property (details in Sections 4 and 5).

Parameterization For any consistency function f(·,·),we have f(xe,∈)= Xe,i.e.,f(·,e) is an identity function. We call this constraint the boundary condition. A valid consistency model has to respect this boundary condition. For consistency models based on deep neural networks, we discuss two ways to implement this boundary condition almost for free. Suppose we have a free-form deep neural network Fe(x,t) whose output has the same dimensionality as x.The first way is to simply parameterize the consistency model as

$$\tag{4}$$

The second method is to parameterize the consistency model using skip connections, that is,

$$\tag{5}$$

where Cskip(t) and Cout(t) are differentiable functions such that Cskip(∈)= 1,and Cout(∈)= O. This way, the consistency model is differentiable at t = ∈ifFe(x,t) and scaling coefficients are differentiable,which is critical for training continuous-time consistency models (Appendices B.1 and B.2). The parameterization in Eq.(5) bears strong resemblance to many successful diffusion models (Karras et al., 2022; Balaji et al., 2022),making it easier to borrow powerful diffusion model architectures for constructing consistency models.We therefore follow the second parameterization in all experiments.

Sampling With a well-trained consistency model fe(*,:), we can generate samples by sampling from the initial distribution XT \~ (O,T²I) and then evaluating the consistency model for Xe = fe(xT,T). This involves only one forward pass through the consistency model and therefore generates samples in a single step. Importantly, one can also evaluate the consistency model multiple times by alternating denoising and noise injection steps for improved sample quality. Summarized in Algorithm 1, this multistep sampling procedure provides the flexibility to trade compute for sample quality. It also has important applications in zero-shot data editing. In practice,we find time points in Algorithm 1 with a greedy algorithm,where the time points are pinpointed one at a time using ternary search to optimize the FID of samples obtained from Algorithm 1.

Algorithm 1 Multistep Consistency Sampling
Input: Consistency model fe(·,·)，sequence of time
points T1 > T2 > ·.> TN-1,initial noise XT
X ←fe(xt,T)
for n=1 to N-1do
Sample z \~ N(0,I)
xTπ←x+√T²-∈²z
X←fθ(xTn,Tn)
end for
Output: x

Zero-Shot Data Editing Consistency models enable various data editing and manipulation applications in zero shot; they do not require explicit training to perform these tasks. For example,consistency models define a one-to-one mapping from a Gaussian noise vector to a data sample. Similar to latent variable models like GANs, VAEs,and normalizing flows,consistency models can easily interpolate between samples by traversing the latent space (Fig.11).As consistency models are trained to recover Xe from any noisy input Xt Where t ∈ [e,T],they can perform denoising for various noise levels (Fig.12). Moreover, the multistep generation procedure in Algorithm1is useful for solving certain inverse problems in zero shot by using an iterative replacement procedure similar to that of diffusion models (Song & Ermon, 2019; Song et al., 2021; Ho et al., 2022b). This enables many applications in the context of image editing,including inpainting (Fig.1O),colorization (Fig. 8), super-resolution (Fig. 6b) and stroke-guided image editing (Fig. 13) as in SDEdit (Meng et al., 2021). In Section 6.3,we empirically demonstrate the power of consistency models on many zero-shot image editing tasks.

4. Training Consistency Models via Distillation

We present our first method for training consistency models based on distilling a pre-trained score model s(x,t). Our discussion revolves around the empirical PF ODE in Eq. (3), obtained by plugging the score model s𝜙(x,t) into the PF ODE. Consider discretizing the time horizon [e,T] into N -1 sub-intervals,with boundaries t1 = ∈ < t2 < ·. < tN = T. In practice,we follow Karras et al. (2O22) to determine the boundaries with the formula ti = (e1/p + i-1/N-1(T1/p - ɕ1/p)),where ρ = 7.When N is sufficiently large,we can obtain an accurate estimate of Xtn from Xtn+1 by running one discretization step of a numerical ODE solver. This estimate,which we denote as x,is defined by

$$\tag{6}$$

where Φ(.. ;) represents the update function of a onestep ODE solver applied to the empirical PF ODE. For example, when using the Euler solver, we have Φ(x,t; Φ) = -ts(x,t) which corresponds to the following update rule

For simplicity,we only consider one-step ODE solvers in this work.It is straightforward to generalize our framework to multistep ODE solvers and we leave it as future work.

Due to the connection between the PF ODE in Eq. (2) and the SDE in Eq.(1) (see Section 2),one can sample along the distribution of ODE trajectories by first sampling X ～ Pdatas then adding Gaussian noise to x. Specifically, given a data point X,we can generate a pair of adjacent data points (nXtn+1) on the PF ODE trajectory efficiently by sampling x from the dataset, followed by sampling Xtn+1 from the transition density of the SDE N(x,tn+1I),and then using one discretization step of the numerical ODE solver according to Eq. (6). Afterwards, we train the consistency model by minimizing its output differences on the pair (xn, Xtn+1). This motivates our following consistency distillation loss for training consistency models.

Definition 1. The consistency distillation loss is defined as

$$\tag{7}$$

where the expectation is taken with respect to X \~ pdata, n \~ U[1,N-1], and Xtn+1 \~ N(x;tn+1I). HereU[1,N-1] denotes the uniform distribution over {1,2,..,N -1}, X(-) ∈ R+ is a positive weighting function, xt is given by Eq. (6), 0- denotes a running average of the past values of 0 during the course ofoptimization, and d(·,:) is a metric function that satisfies ∀x,y : d(x,y) ≥ O and d(x,y) = 0 if and only if x = y.

Unless otherwise stated, we adopt the notations in Definition 1 throughout this paper,and use E[·] to denote the expectation over all relevant random variables. In our experiments, we consider the squared l2 distance d(x,y） = |x-yl²,l1 distance d(x,y) = |x-yll1,and the Learned Perceptual Image Patch Similarity (LPIPS, Zhang et al. (2018)). We find X(tn) = 1 performs well across all tasks and datasets. In practice,we minimize the objective by stochastic gradient descent on the model parameters 0,while updating 0- with exponential moving average (EMA).That is,given a decay rate O ≤ μ <1,we perform the following update after each optimization step:

Algorithm 2 Consistency Distillation (CD) Algorithm 3 Consistency Training (CT)Input: dataset D,initial model parameter 0,learning rate Input: dataset D, initial model parameter 0, learningn, ODE solver Φ(-, ; Φ),d(.,-), λ(-),and μ rate η, step schedule N(·),EMA decay rate schedule0-←θ μ(-),d(-,·), and λ(-)repeat 0-←θandk←0Sample x ~ D and n ~U[1,N - 1] repeatSample Xtn+1 ~ N(x;t²+1I) Sample x ~ D,and n ~U[1,N(k) - 1]xtn←Xtn+1 +(tn-tn+1)Φ(xtn+1,tn+1;) Sample z ~ N(0, I)L(0,0-) ←L(0,0−;Φ)←λ(tn)d(fe(x+tn+1z,tn+1),fθ-(x+tnz,tn)X(tn)d(fθ(xtn+1,tn+1),fθ-(xn,tn)) 0←0-nVθL(0,0-)0←0-n∀θl(0,0-;Φ) 0− ← stopgrad(μ(k)0- +(1- μ(k)0)0- ← stopgrad(μ0- + (1- μ)0) k←k+1until convergence until convergence

$$\tag{8}$$

The overall training procedure is summarized in Algorithm 2. In alignment with the convention in deep reinforcement learning (Mnih et al., 2013; 2015; Lillicrap et al., 2015) and momentum based contrastive learning (Grill et al., 2020; He et al., 202O), we refer to fe- as the “target network”, and fe as the“online network".We find that compared to simply setting 0- = 0, the EMA update and “stopgrad" operator in Eq. (8) can greatly stabilize the training process and improve the final performance of the consistency model.

Below we provide a theoretical justification for consistency distillation based on asymptotic analysis.

Theorem 1. Let △t := maXn∈[1,N-1]{tn+1 - tnl},and f(·, ;Φ) be the consistency function of the empirical PF ODE in Eq. (3). Assume fe satisfies the Lipschitz condition: there exists L > O such that for all t ∈ [ε,T], x, and y, we have |lfe(x,t) - fe(y,t)ll2 ≤ L|/x-yll2. Assume further that for alln∈[1,N-1], the ODE solver called at tn+1 has local error uniformly bounded by O((tn+1 - tn)p+1) with p ≥ 1. Then,ifLN(0,0;Φ) = 0, we have

Proof. The proof is based on induction and parallels the classic proof of global error bounds for numerical ODE solvers (Suli & Mayers,2O03). We provide the full proof in Appendix A.2.

Since θ- is a running average of the history of 0, we have 0- = θ when the optimization of Algorithm 2 converges.

That is, the target and online consistency models will eventually match each other.If the consistency model additionally achieves zero consistency distillation loss,then Theorem 1 implies that,under some regularity conditions,the estimated consistency model can become arbitrarily accurate, as long as the step size of the ODE solver is sufficiently small.

The consistency distillation loss LN(0,0-;Φ) can be extended to hold for infinitely many time steps (N → oo) if 0- = 0 or 0- = stopgrad(0). The resulting continuoustime loss functions do not require specifying N nor the time steps {t1,t2,·.,tN}.Nonetheless,they involve Jacobianvector products and require forward-mode automatic differentiation for efficient implementation,which may not be well-supported in some deep learning frameworks. We provide these continuous-time distillation loss functions in Theorems 3 to 5,and relegate details to Appendix B.1.

5. Training Consistency Models in Isolation

Consistency models can be trained without relying on any pre-trained diffusion models.This differs from diffusion distillation techniques,making consistency models a new independent family of generative models.

In consistency distillation, we use a pre-trained score model S𝜙(x,t) to approximate the ground truth score function Vlog Pt(x). To get rid of this dependency, we need to seek other ways to estimate the score function. In fact, there exists an unbiased estimator of Vlog Pt(xt） due to the following identity (Lemma 1 in Appendix A):

where X \~ Pdata and Xt \~ N(x; t²I). That is, given X and Xt,we can form a Monte Carlo estimate of Vlog Pt(Xt) with -(Xt - x)/t². We now show that this estimate actually suffices to replace the pre-trained diffusion model in consistency distillation,when using the Euler method (or any higher order method) as the ODE solver in the limit of N→8.

More precisely, we have the following theorem.

Theorem 2. Let △t := maXn∈[1,N-1]{ltn+1 - tnl}. Assume d and fe- are both twice continuously differentiable with bounded second derivatives, the weighting function λ(-）is bounded,and E[|Vlog ptn(xtn)ll²] < oo. Assume further that we use the Euler ODE solver, and the pre-trained score model matches the ground truth,i.e., Vt ∈[e,T] : s(x,t) = Vlog pt(x). Then,

$$\tag{9}$$

where the expectation is taken with respect to X \~ Pdata, n \~ U[1,N-1], and xtn+1\~N(x;t+11). The conistency training objective, denoted by LNr(0,0-), is defined as

$$\tag{10}$$

where z \~ N(0,I). Moreover, LNr(0,0-) ≥ O(△t) if infN LN(0,0-;Φ) >0.

Proof. The proof is based on Taylor series expansion and properties of score functions (Lemma 1).A complete proof is provided in Appendix A.3.

We refer to Eq. (1O) as the consistency training (CT) loss. Crucially, L(0,0-）only depends on the online network fe,and the target network fe-，while being completely agnostic to diffusion model parameters .The loss function L(0,0-）≥ O(△t) decreases at a slower rate than the remainder o(△t) and thus will dominate the loss in Eq. (9) as N → ∞ and △t → 0.

For improved practical performance, we propose to progressively increase N during training according to a schedule function N(-). The intuition (cf.,Fig.3d) is that the consistency training loss has less “variance”but more“bias”with respect to the underlying consistency distillation loss (i.e., the left-hand side of Eq. (9)) when N is small (i.e., △t is large), which facilitates faster convergence at the beginning of training. On the contrary, it has more “variance”but less "bias”when N is large (i.e.,△t is small), which is desirable when closer to the end of training. For best performance, we also find that μ should change along with N,according to a schedule function μ(·). The full algorithm of consistency training is provided in Algorithm 3,and the schedule functions used in our experiments are given in Appendix C.

Similar to consistency distillation, the consistency training loss LNr(0,0-) can be extended to hold in continuous time (i.e.,N → o) if 0- = stopgrad(0),as shown in Theorem 6. This continuous-time loss function does not require schedule functions for N or μ, but requires forward-mode automatic differentiation for efficient implementation. Unlike the discrete-time CT loss, there is no undesirable“bias” associated with the continuous-time objective,as we effectively take △ → O in Theorem 2.We relegate more details to Appendix B.2.

6. Experiments

We employ consistency distillation and consistency training to learn consistency models on real image datasets, including CIFAR-10 (Krizhevsky et al., 2009), ImageNet 64 × 64 (Deng et al., 2009),LSUN Bedroom 256 × 256, and LSUN Cat 256 × 256 (Yu et al., 2015). Results are compared according to Fréchet Inception Distance (FID, Heusel et al. (2O17),lower is better), Inception Score (Is, Salimans et al. (2O16), higher is better), Precision (Prec., Kynkäänniemi et al. (2O19), higher is better),and Recall (Rec., Kynkäänniemi et al. (2O19), higher is better). Addi-tional experimental details are provided in Appendix C.

6.1. Training Consistency Models

We perform a series of experiments on CIFAR-1O to understand the effect of various hyperparameters on the performance of consistency models trained by consistency distillation (CD) and consistency training (CT).We first focus on the effect of the metric function d(·,:), the ODE solver, and the number of discretization steps N in CD, then investigate the effect of the schedule functions N(-) and μ(-) in CT.

To set up our experiments for CD,we consider the squared l2 distance d(x,y) = |x - yl²2,l1 distance d(x,y) = |x -yll1,and the Learned Perceptual Image Patch Similarity (LPIPS, Zhang et al. (2O18)) as the metric function. For the ODE solver, we compare Euler's forward method and Heun's second order method as detailed in Karras et al. (2022). For the number of discretization steps N,we compare N ∈ {9,12,18,36,50,60,80,120}. All consistency models trained by CD in our experiments are initialized with the corresponding pre-trained diffusion models,whereas models trained by CT are randomly initialized.

As visualized in Fig.3a, the optimal metric for CD is LPIPS, which outperforms both l1 and l2 by a large margin over all training iterations.This is expected as the outputs of consistency models are images on CIFAR-1O,and LPIPS is specifically designed for measuring the similarity between natural images.Next, we investigate which ODE solver and which discretization step N work the best for CD.As shown in Figs.3b and 3c,Heun ODE solver and N = 18 are the best choices.Both are in line with the recommendation of Karras et al. (2O22) despite the fact that we are training consistency models,not diffusion models.Moreover, Fig.3b shows that with the same N,Heun's second order solver uniformly outperforms Euler's first order solver. This corroborates with Theorem 1,which states that the optimal consistency models trained by higher order ODE solvers have smaller estimation errors with the same N.The results of Fig. 3c also indicate that once N is sufficiently large, the performance of CD becomes insensitive to N. Given these insights,we hereafter use LPIPS and Heun ODE solver for CD unless otherwise stated. For N in CD,we follow the suggestions in Karras et al. (2O22) on CIFAR-1O and ImageNet 64 × 64. We tune N separately on other datasets (details in Appendix C).

(a) Metric functions in CD.

(b) Solvers and N in CD.

(d) Adaptive N and μ in CT.

Figure 3: Various factors that affect consistency distillation (CD) and consistency training (CT) on CIFAR-1O. The best configuration for CD is LPIPS,Heun ODE solver, and N = 18. Our adaptive schedule functions for N and μ make CT converge significantly faster than fixing them to be constants during the course of optimization.

(a) CIFAR-10

(b) ImageNet 64 × 64

(d) Cat 256 × 256
Figure 4: Multistep image generation with consistency distillation (CD). CD outperforms progressive distillation (PD) across all datasets and sampling steps. The only exception is single-step generation on Bedroom 256 × 256.

Due to the strong connection between CD and CT, we adopt LPIPS for our CT experiments throughout this paper. Unlike CD, there is no need for using Heun's second order solver in CT as the loss function does not rely on any particular numerical ODE solver. As demonstrated in Fig.3d, the convergence of CT is highly sensitive to N—smaller Nleads to faster convergence but worse samples,whereas larger N leads to slower convergence but better samples upon convergence.This matches our analysis in Section 5,and motivates our practical choice of progressively growing N and μ for CT to balance the trade-off between convergence speed and sample quality. As shown in Fig. 3d,adaptive schedules of N and μ significantly improve the convergence speed and sample quality of CT. In our experiments,we tune the schedules N(·) and μ(·) separately for images of different resolutions,with more details in Appendix C.

6.2. Few-Step Image Generation

Distillation In current literature, the most directly comparable approach to our consistency distillation (CD) is progressive distillation (PD, Salimans & Ho (2022)); both are thus far the only distillation approaches that do not construct synthetic data before distillation.In stark contrast, other distillation techniques, such as knowledge distillation (Luhman & Luhman, 2021) and DFNO (Zheng et al., 2022), have to prepare a large synthetic dataset by generating numerous samples from the diffusion model with expensive numerical ODE solvers.We perform comprehensive comparison between PD and CD on CIFAR-10, ImageNet 64 × 64, and LSUN 256 × 256,with all results reported in Fig. 4. All methods distill from an EDM(Karras et al., 2022) model that we pre-trained in-house.We note that across all sampling iterations,using the LPIPS metric uniformly improves PD compared to the squared l2 distance in the original paper of Salimans & Ho (2022). Both PD and CD improve as we take more sampling steps.We find that CD uniformly outperforms PD across all datasets,sampling steps,and metric functions considered, except for single-step generation on Bedroom 256 × 256,where CD with l2 slightly underperforms PD with l2.As shown in Table 1, CD even outperforms distillation approaches that require synthetic dataset construction, such as Knowledge Distillation (Luhman & Luhman, 2021) and DFNO (Zheng et al., 2022).

Table 1: Sample quality on CIFAR-10. *Methods that require synthetic data construction for distillation.

METHOD	NFE(↓)	FID(↓)	IS(↑)
Diffusion + Samplers DDIM (Song et al., 2020)	50	4.67
DDIM (Song et al., 2020)	20	6.84
DDIM (Song et al.,2020)	10	8.23
DPM-solver-2 (Lu et al.,2022)	12	5.28
DPM-solver-3 (Lu et al.,2022)	12	6.03
3-DEIS (Zhang & Chen,2022)	10	4.17
Diffusion + Distillation
Knowledge Distillation*(Luhman & Luhman,2021)		9.36
DFNO* (Zheng et al., 2022)	1
1-Rectified Flow (+distil)* (Liu et al., 2022)	1	4.12	9.08
2-RectifiedFlow (+distil)*(Liu etal.,22)	1	6.18	9.01
3-Rectified Flow (+distill)*(Liu et al.,2022)	1	4.85	8.79
PD (Salimans & Ho,2022)	1 1	5.21	8.69
CD		8.34 3.55	9.48
PD (Salimans & Ho,2022)	1 2	5.58	9.05
CD	2	2.93	9.75
Direct Generation
BigGAN (Brock etal.,2019)			9.22
	1	14.7
CR-GAN (Zhang et al., 2019)	1	14.6	8.40
AutoGAN (Gong et al.,2019)	1	12.4	8.55
E2GAN (Tian et al., 2020)	1	11.3	8.51
ViTGAN (Lee et al., 2021)	1	6.66	9.30
TransGAN (Jiang et al.,2021) StyleGAN2-ADA (Karras et al.,2020)	1	9.26	9.05
	1	2.92	9.83
StyleGAN-XL (Sauer et al., 2022)	1	1.85
Score SDE (Song et al.,2021)	2000	2.20	9.89
DDPM (Ho et al.,2020)	1000	3.17	9.46
LSGM (Vahdat et al.,2021)	147	2.10
PFGM (Xu et al., 2022)	110	2.35	9.68
EDM (Karras et al., 2022)	36	2.04	9.84
1-Rectified Flow (Liu et al.,2022)	1	378	1.13
Glow (Kingma & Dhariwal, 2018)	1	48.9	3.92
Residual Flow (Chen et al.,2019a)	1	46.4
GLFlow (Xiao et al., 2019)	1	44.6
DenseFlow (Grcic et al.,2021)	1	34.9
DC-VAE (Parmar et al., 2021)	1	17.9	8.20
CT	1	8.70	8.49
CT	2	5.83	8.85

Table 2: Sample quality on ImageNet 64 × 64, and LSUN Bedroom & Cat 256 × 256.tDistillation techniques.

METHOD ImageNet 64 × 64	NFE(↓)	FID (↓)	Prec.(↑)	Rec.(↑)
PD+ (Salimans & Ho,2022)	1	15.39	0.59	0.62
DFNOt* (Zheng et al.,2022)	1	8.35
CD	1	6.20	0.68	0.63
PD† (Salimans & Ho,2022)	2	8.95	0.63	0.65
CDt	2	4.70	0.69	0.64
ADM (Dhariwal & Nichol,2021)	250	2.07	0.74	0.63
EDM (Karras et al., 2022)	79	2.44	0.71	0.67
BigGAN-deep (Brock et al., 2019)	1	4.06	0.79	0.48
CT	1	13.0	0.71	0.47
CT	2	11.1	0.69	0.56
LSUN Bedr00m 256 × 256
PDf (Salimans & Ho,2022)	1	16.92	0.47	0.27
PD† (Salimans & Ho,2022)	2	8.47	0.56	0.39
CD†	1	7.80	0.66	0.34
CD	2	5.22	0.68	0.39
DDPM (Ho et al., 2020)	1000	4.89	0.60	0.45
ADM (Dhariwal & Nichol, 2021)	1000	1.90	0.66	0.51
EDM (Karras et al., 2022)	79	3.57	0.66	0.45
SS-GAN (Chen et al.,2019b)	1	13.3
PGGAN (Karras et al., 2018)	1	8.34
PG-SWGAN (Wu et al.,2019)	1	8.0
StyleGAN2 (Karras et al.,2020)	1	2.35	0.59	0.48
CT	1	16.0	0.60	0.17
CT	2	7.85	0.68	0.33
LSUN Cat 256× 256
PD† (Salimans & Ho,2022)	1	29.6	0.51	0.25
PD+ (Salimans & Ho,2022)	2	15.5	0.59	0.36
CDt	1	11.0	0.65	0.36
CD	2	8.84	0.66	0.40
DDPM (Ho et al.,2020)	1000	17.1	0.53	0.48
ADM (Dhariwal & Nichol,2021)	1000	5.57	0.63	0.52
EDM (Karras et al., 2022)	79	6.69	0.70	0.43
PGGAN (Karras et al., 2018)	1	37.5
StyleGAN2 (Karras et al.,020)	1	7.25	0.58	0.43
CT	1	20.7	0.56	0.23
CT	2	11.7	0.63	0.36

Figure 5: Samples generated by EDM (top), CT + single-step generation (middle),and CT + 2-step generation (Bottom). All corresponding images are generated from the same initial noise.

(a) Left: The gray-scale image. Middle: Colorized images. Right: The ground-truth image.

Figure 6: Zero-shot image editing with a consistency model trained by consistency distillation on LSUN Bedroom 256 × 256.

Direct Generation In Tables 1 and 2,we compare the sample quality of consistency training (CT) with other generative models using one-step and two-step generation.We also include PD and CD results for reference. Both tables report PD results obtained from the l2 metric function,as this is the default setting used in the original paper of Salimans & Ho (2O22).For fair comparison, we train both PD and CD to distill the same EDM models. In Tables 1 and 2, we observe that CT outperforms all single-step, non-adversarial generative models,i.e.，VAEs and normalizing flows,by a significant margin on CIFAR-1O.Moreover, CT obtains comparable quality to PD for single-step generation without relying on distillation. In Fig.5,we provide EDM samples (top),single-step CT samples (middle),and two-step CT samples (bottom). In Appendix E,we show additional samples for both CD and CT in Figs.14 to 21. Importantly, all samples obtained from the same initial noise vector share significant structural similarity,even though CT and EDM models are trained independently from one another. This indicates that CT is unlikely to suffer from mode collapse, as EDMs do not.

6.3. Zero-Shot Image Editing

Similar to diffusion models,consistency models allow zero-shot image editing by modifying the multistep sampling process in Algorithm 1. We demonstrate this capability with a consistency model trained on the LSUN bedroom dataset using consistency distillation.In Fig. 6a,we show such a consistency model can colorize gray-scale bedroom images at test time,even though it has never been trained on colorization tasks.In Fig. 6b,we show the same consistency model can generate high-resolution images from low-resolution inputs. In Fig. 6c,we additionally demonstrate that it can generate images based on stroke inputs created by humans,as in SDEdit for diffusion models (Meng et al.,2021). Again, this editing capability is zero-shot, as the model has not been trained on stroke inputs. In Appendix D,we additionally demonstrate the zero-shot capability of consistency models on inpainting (Fig. 10), interpolation (Fig.11) and denoising (Fig.12),with more examples on colorization (Fig. 8), super-resolution (Fig. 9) and stroke-guided image generation (Fig. 13).

7. Conclusion

We have introduced consistency models,a type of generative models that are specifically designed to support one-step and few-step generation. We have empirically demonstrated that our consistency distillation method outshines the existing distillation techniques for diffusion models on multiple image benchmarks and various sampling iterations.Furthermore,as a standalone generative model, consistency models outdo other available models that permit single-step generation,barring GANs.Similar to diffusion models,they also allow zero-shot image editing applications such as inpainting,colorization,super-resolution,denoising,interpolation, and stroke-guided image generation.

In addition, consistency models share striking similarities with techniques employed in other fields,including deep Q-learning (Mnih et al., 2O15) and momentum-based contrastive learning (Grill et al., 202O; He et al., 202O). This offers exciting prospects for cross-pollination of ideas and methods among these diverse felds.

Acknowledgements

We thank Alex Nichol for reviewing the manuscript and providing valuable feedback, Chenlin Meng for providing stroke inputs needed in our stroke-guided image generation experiments,and all members of the OpenAI Algorithms team for helpful suggestions during the project.

References

Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., Karras,T.,and Liu,M.-Y. ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324,2022.

Bengio,Y.and Bengio,S.Modeling high-dimensional discrete data with multi-layer neural networks. Advances in Neural Information Processing Systems,12,1999.

Brock,A., Donahue, J.,and Simonyan, K. Large scale GAN training for high fidelity natural image synthesis. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum? id=B1xsqj09Fm.

Chen,N., Zhang, Y., Zen, H., Weiss, R.J., Norouzi, M., and Chan, W. Wavegrad: Estimating gradients for waveform generation. In International Conference on Learning Representations (ICLR),2021.

Chen,R.T., Behrmann,J., Duvenaud, D.K.,and Jacobsen, J.-H. Residual flows for invertible generative modeling. In Advances in Neural Information Processing Systems, pp. 9916-9926,2019a.

Chen,T., Zhai,X.,Ritter, M., Lucic,M.,and Houlsby, N. Self-supervised gans via auxiliary rotation loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.12154-12163,2019b.

Chung, H., Kim, J.,Mccann, M. T., Klasky, M.L.,and Ye, J. C. Diffusion posterior sampling for general noisy inverse problems. In International Conference on Learning Representations, 2023. URL https : / /openreview . net/forum?id=OnD9zGAGT0k.

Deng, J.,Dong,W., Socher,R., Li,L.-J.,Li, K.,and Fei-Fei, L. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition,pp.248-255.Ieee,2009.

Dhariwal, P. and Nichol, A. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems (NeurIPS),2021.

Dinh,L., Krueger, D., and Bengio, Y. NICE: Non-linear independent components estimation. International Conference in Learning Representations Workshop Track, 2015.

Dinh,L., Sohl-Dickstein,J.,and Bengio, S. Density estimation using real NVP. In 5th International Conference on Learning Representations,ICLR 2Ol7, Toulon, France, April 24-26, 20l7, Conference Track Proceedings. OpenReview.net, 2017. URL ht tps : / /open review . net/forum?id=HkpbnH9lx.

Dockhorn, T., Vahdat, A., and Kreis,K. Genie: Higherorder denoising diffusion solvers.arXiv preprint arXiv:2210.05475,2022.

Gong, X., Chang, S., Jiang, Y., and Wang, Z. Autogan: Neural architecture search for generative adversarial networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 3224-3234,2019.

Goodfellow, I.， Pouget-Abadie, J.，Mirza， M.， Xu,B., Warde-Farley, D., Ozair, S., Courville,A., and Bengio, Y.Generative adversarial nets. In Advances in neural information processing systems, pp.2672-2680, 2014.

Grcic,M., Grubisic,I.,and Segvic, S. Densely connected normalizing flows.Advances in Neural Information Processing Systems,34:23968-23982,2021.

Grill, J.-B., Strub, F., Altché,F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar,M., et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems,33:21271-21284, 2020.

He, K., Fan, H., Wu, Y., Xie,S., and Girshick, R.Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 9729-9738, 2020.

Heusel,M.,Ramsauer, H., Unterthiner, T.,Nessler,B., and Hochreiter, S. GANs trained by a two time-scale update rule converge to a local Nash equilibrium. In Advances in Neural Information Processing Systems,pp. 6626-6637, 2017.

Ho, J., Jain,A.,and Abbeel, P. Denoising Diffusion Probabilistic Models.Advances in Neural Information Processing Systems,33,2020.

Ho, J., Chan, W., Saharia, C., Whang,J., Gao,R., Gritsenko, A.,Kingma, D.P., Poole, B., Norouzi, M., Fleet, D. J., et al.Imagen video:High definition video generation with diffusion models.arXiv preprint arXiv:2210.02303, 2022a.

Ho,J., Salimans, T., Gritsenko,A. A., Chan, W., Norouzi, M.,and Fleet,D.J.Video diffusion models.In ICLR Workshop on Deep Generative Models for Highly Structured Data, 2022b. URL https://openreview . net/forum?id=BBelR2NdDZ5.

Hyvärinen, A.and Dayan, P. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research (JMLR),6(4),2005.

Jiang, Y., Chang, S., and Wang, Z. Transgan: Two pure transformers can make one strong gan,and that can scale up. Advances in Neural Information Processing Systems, 34:14745-14758, 2021.

Karras,T.,Aila, T., Laine, S.,and Lehtinen, J. Progres-sive growing of GANs for improved quality, stability, and variation. In International Conference on Learning Representations,2018. URL https : / /openreview . net/forum?id=Hk99zCeAb.

Karras, T., Laine, S., Aitala, M., Hellsten, J., Lehtinen, J., and Aila, T. Analyzing and improving the image quality of stylegan. 2020.

Karras,T., Aitala, M., Aila, T.,and Laine, S. Elucidating the design space of diffusion-based generative models. In Proc.NeurIPS,2022.

Kawar, B., Vaksman, G.,and Elad, M. Snips: Solving noisy inverse problems stochastically. arXiv preprint arXiv:2105.14951,2021.

Kawar, B., Elad, M., Ermon, S.,and Song,J. Denoising diffusion restoration models.In Advances in Neural Information Processing Systems,2022.

Kingma, D.P. and Dhariwal, P. Glow: Generative flow with invertible 1x1 convolutions. In Bengio, S.，Wallach,H.,Larochelle,H., Grauman, K., Cesa-Bianchi, N., and Garnett, R. (eds.), Advances in Neural Information Processing Systems 31, pp.10215-10224.2018.

Kingma, D.P. and Welling,M. Auto-encoding variational bayes. In International Conference on Learning Representations,2014.

Kong,Z., Ping, W., Huang, J., Zhao,K., and Catanzaro, B.DiffWave:A Versatile Diffusion Model for Audio Synthesis. arXiv preprint arXiv:2009.09761,2020.

Krizhevsky,A., Hinton, G., et al. Learning multiple layers of features from tiny images.2009.

Kynkäänniemi, T., Karras, T.,Laine, S., Lehtinen, J., and Aila, T. Improved precision and recall metric for assessing generative models. Advances in Neural Information Processing Systems,32,2019.

Lee,K., Chang, H., Jiang,L., Zhang, H., Tu, Z.,and Liu, C.Vitgan: Training gans with vision transformers. arXiv preprint arXiv:2107.04589,2021.

Lillicrap, T.P.,Hunt, J.J., Pritzel,A., Heess,N.,Erez, T., Tassa, Y., Silver, D.,and Wierstra, D. Continuous control with deep reinforcement learning. arXiv preprint arXiv:1509.02971,2015.

Liu,L., Jiang, H., He, P., Chen, W., Liu, X., Gao, J., and Han, J. On the variance of the adaptive learning rate and beyond. arXiv preprint arXiv:1908.03265,2019.

Liu,X., Gong, C.,and Liu, Q. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003,2022.

Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., and Zhu, J. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 1O steps.arXiv preprint arXiv:2206.00927,2022.

Luhman,E.and Luhman,T. Knowledge distillation in iterative generative models for improved sampling speed. arXiv preprint arXiv:2101.02388,2021.

Meng, C., Song, Y., Song, J., Wu, J., Zhu, J.-Y.,and Ermon, S. Sdedit: Image synthesis and editing with stochastic differential equations.arXiv preprint arXiv:2108.01073, 2021.

Meng, C., Gao, R., Kingma, D.P., Ermon, S., Ho, J., and Salimans,T. On distillation of guided diffusion models. arXiv preprint arXiv:2210.03142,2022.

Mnih，V.，Kavukcuoglu，K.， Silver，D.，Graves，A., Antonoglou, I., Wierstra, D.,and Riedmiller, M. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602,2013.

Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G., Graves,A., Riedmiller, M., Fidjeland,A.K., Ostrovski, G.,et al. Human-level control through deep reinforcement learning. nature,518(7540): 529-533,2015.

Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P.,McGrew,B.，Sutskever, I.,and Chen,M.Glide: Towards photorealistic image generation and editing with text-guided diffusion models.arXiv preprint arXiv:2112.10741,2021.

Parmar, G.,Li, D.,Lee,K.,and Tu, Z. Dual contradistinctive generative autoencoder. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 823-832,2021.

Popov, V., Vovk, I., Gogoryan, V., Sadekova, T.,and Kudinov, M. Grad-TTS:A diffusion probabilistic model for text-to-speech. arXiv preprint arXiv:2105.06337,2021.

Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., and Chen, M.Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125,2022.

Rezende,D.J.,Mohamed, S., and Wierstra, D. Stochastic backpropagation and approximate inference in deep generativemodels.In Proceedingsof the 3lst International Conference on Machine Learning,pp.1278-1286,2014.

Rombach,R.,Blattmann,A., Lorenz,D., Esser, P.,and Ommer, B.High-resolution image synthesis with latent diffusion models.In Proceedingsof the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 10684-10695,2022.

Saharia, C., Chan, W., Saxena, S.,Li,L., Whang,J.,Denton, E., Ghasemipour, S.K. S., Ayan, B. K., Mahdavi, S. S., Lopes,R.G., et al.Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487,2022.

Salimans, T.and Ho, J. Progressive distillation for fast sampling of diffusion models. In International Conference on Learning Representations, 2022. URL ht tps : //openreview.net/forum?id=TIdIXIpzhoI.

Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford,A.,and Chen, X. Improved techniques for training gans. In Advances in neural information processing systems,pp.2234-2242,2016.

Sauer, A., Schwarz, K.,and Geiger, A. Stylegan-xl: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH 2022 conference proceedings, pp.1-10, 2022.

Sohl-Dickstein,J.,Weiss,E.,Maheswaranathan,N.,and Ganguli, S. Deep Unsupervised Learning Using Nonequilibrium Thermodynamics. In International Conference on Machine Learning, pp.2256-2265,2015.

Song,J., Meng,C.,and Ermon, S. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502,2020.

Song， J.，Vahdat，A.，Mardani，M.，and Kautz，J. Pseudoinverse-guided diffusion models for inverse problems.In International Conference on Learning Representations, 2023. URL https: //openreview.net / forum?id=9_gsMA8MRKQ.

Song,Y.and Ermon, S. Generative Modeling by Estimating Gradients of the Data Distribution.In Advances in Neural Information Processing Systems,pp.11918-11930,2019.

Song,Y. and Ermon, S.Improved Techniques for Training Score-Based Generative Models.Advances in Neural Information Processing Systems,33,2020.

Song,Y., Garg,S., Shi, J.,and Ermon, S. Sliced score matching: A scalable approach to density and score estimation. In Proceedings of the Thirty-Fifth Conference on Uncertainty in Artificial Intelligence, UAI 2019,Tel Aviv, Israel, July 22-25,2019,pp.204,2019.

Song,Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S.,and Poole,B. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum? id=PxTIG12RRHS.

Song,Y., Shen,L., Xing,L.,and Ermon, S. Solving inverse problems in medical imaging with score-based generative models. In International Conference on Learning Representations,2022. URL https : / /openreview . net/forum?id=vaRCHVj0uGI.

Suli,E.and Mayers,D.F. An introduction to numerical analysis. Cambridge university press,2003.

Tian, Y., Wang, Q., Huang, Z., Li, W., Dai, D., Yang, M., Wang, J.,and Fink, O. Off-policy reinforcement learning for efficient and effective gan architecture search. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pp.175-192. Springer,2020.

Uria,B., Murray, I.,and Larochelle,H. Rnade: the realvalued neural autoregressive density-estimator. In Proceedings of the 26th International Conference on Neural Information Processing Systems-Volume 2, pp.2175- 2183,2013.

Uria，B.， Coté，M.-A.， Gregor， K.， Murray， I.，andLarochelle,H. Neural autoregressive distribution esti-mation. The Journal of Machine Learning Research,17(1):7184-7220, 2016.

Vahdat, A., Kreis, K.,and Kautz, J. Score-based generative modeling in latent space.Advances in Neural Information Processing Systems,34:11287-11302,2021.

Van Den Oord, A., Kalchbrenner, N., and Kavukcuoglu, K. Pixel recurrent neural networks. In Proceedings of the 33rd International Conference on International Conference on Machine Learning - Volume 48, ICML'16, pp. 1747-1756.JMLR.org,2016. URL http: //dl .acm. org/citation.cfm?id=3045390.3045575.

Vincent, P. A Connection Between Score Matching and Denoising Autoencoders. Neural Computation, 23(7): 1661-1674, 2011.

Wu, J., Huang, Z., Acharya, D.,Li, W., Thoma, J., Paudel, D.P.,and Gool, L. V. Sliced wasserstein generative models.In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.3713- 3722,2019.

Xiao,Z., Yan, Q.,and Amit, Y. Generative latent flow.arXiv preprint arXiv:1905.10485,2019.

Xu, Y., Liu, Z., Tegmark, M., and Jaakkola, T. S. Pois-son flow generative models. In Oh, A. H., Agarwal, A., Belgrave,D.,and Cho,K. (eds.),Advances in Neural Information Processing Systems, 2022. URL https : //openreview.net/forum?id=voV_TRqcWh.

Yu, F., Seff, A., Zhang, Y., Song, S., Funkhouser, T., and Xiao,J. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365,2015.

Zhang,H., Zhang, Z., Odena, A.,and Lee,H. Consistency regularization for generative adversarial networks.arXiv preprint arXiv:1910.12027,2019.

Zhang，Q.and Chen，Y.Fast sampling of diffusion models with exponential integrator.arXiv preprint arXiv:2204.13902,2022.

Zhang,R., Isola,P., Efros,A.A., Shechtman,E.,and Wang, O. The unreasonable efectiveness of deep features as a perceptual metric. In CVPR, 2018.

Zheng,H., Nie, W., Vahdat, A., Azizzadenesheli, K.,and Anandkumar,A. Fast sampling of diffusion models via operator learning. arXiv preprint arXiv:2211.13449, 2022.

1 Introduction 1
2Diffusion Models 2
3Consistency Models 3
4Training Consistency Models via Distillation 4
5Training Consistency Models in Isolation 5
6Experiments 6
6.1Training Consistency Models 6
6.2Few-Step Image Generation 7
6.3Zero-Shot Image Editing 9
7Conclusion 9
Appendices 15
Appendix A Proofs 15
A.1 Notations 15
A.2 Consistency Distillation 15
A.3 Consistency Training 16
Appendix B Continuous-Time Extensions 18
B.1 Consistency Distillation in Continuous Time . 18
B.2 Consistency Training in Continuous Time 21
B.3Experimental Verifications 23
Appendix C Additional Experimental Details 24
Model Architectures 24
Parameterization for Consistency Models 24
Schedule Functions for Consistency Training 25
Training Details． 25
Appendix D Additional Results on Zero-Shot Image Editing 25
Inpainting. 26
Colorization 26
Super-resolution 27
Stroke-guided image generation 2
Denoising. 2
Interpolation 2
Appendix E Additional Samples from Consistency Models 2

Appendices

A. Proofs

A.1. Notations

We use fe(x,t) to denote a consistency model parameterized by 0,and f(x,t;Φ) the consistency function of the empirical PF ODE in Eq. (3). Here symbolizes its dependency on the pre-trained score model s𝜙(x,t). For the consistency function of the PF ODE in Eq. (2),we denote it as f(x,t). Given a multi-variate function h(x,y), we let dih(x,y) denote the Jacobian of h over x,and analogously Ozh(x,y) denote the Jacobian of h over y. Unless otherwise stated, x is supposed to be a random variable sampled from the data distribution Pdata(x),n is sampled uniformly at random from [1,N -1],and Xtn is sampled from N(x;tnI). Here [1,N -1] represents the set of integers {1,2,..,N -1}. Furthermore,recall that we define

where Φ(.. ; Φ) denotes the update function of a one-step ODE solver for the empirical PF ODE defined by the score model s𝜙(x,t). By default,E[·] denotes the expectation over al relevant random variables in the expression.

A.2. Consistency Distillation

Theorem 1. Let △t := maxne[1,N-1]{tn+1- tnl}, and f(*,:; Φ) be the consistency function ofthe empirical PF ODE in Eq. (3). Assume fe satisfies the Lipschitz condition: there exists L > O such that for allt ∈ [ε,T], x, and y, we have llfe(x,t)- fe(y,t)ll2 ≤L|x-yll2. Assume further that for all n ∈ [1,N-1]], the ODE solver called at tn+1 has local error uniformly bounded by O((tn+1-tn)p+1) with p ≥ 1. Then,ifLD(0,0;Φ)= 0, we have

Proof. From CN(0,0;) = 0, we have

$$\tag{11}$$

According to the definition, we have ptn(Xtn) = Pdata(x) N(O,tnI) where tn ≥ ∈ > O. It follows that Ptn (Xtn)>0 for every Xtn and1 ≤ n ≤ N. Therefore,Eq.(11) entails

$$\tag{12}$$

Because X(·) > O and d(x,y)= O ↔ X = y,this further implies that

$$\tag{13}$$

Now let en represent the error vector at tn,which is defined as

We can easily derive the following recursion relation

$$\tag{14}$$

where (i) is due toEq. (13)and f(xtn+1,tn+1; Φ) = f(xtn,tn;Φ). Because fe(-,tn) has Lipschitz constant L,we have

where (i) holds because the ODE solver has local error bounded by O(tn+1 -tn)p+1). In addition, we observe that e1 = 0, because

Here (i) is true because the consistency model is parameterized such that f(xt1,t1; ) = Xt and (ii) is entailed by the definition of f(·,:; ). This allows us to perform induction on the recursion formula Eq.(14) to obtain

which completes the proof.

A.3. Consistency Training

The following lemma provides an unbiased estimator for the score function, which is crucial to our proof for Theorem 2.
Lemma1. Let x \~ Pdata(x), xt \~ N(x; t²I),and pt(xt) = Pdata(x)N(0,t²I). We have Vlog Pt(x)= -E[Xt× | xt].

Proof. According to the definition of pt(xt), we have Vlog pt(xt) = Vxt log SPdata(x)p(xt | x) dx, where p(xt | x) = N(Xt; X,t²I). This expression can be further simplified to yield

where (i) is due to Bayes' rule.

Theorem 2. Let △t := maxne[1,N-1}{tn+1 - tnl}. Assume d and fe- are both twice continuously diferentiable with bounded second derivatives,the weighting function X(.) is bounded, and E[|Vlog Ptn(Xtr)ll2] < o. Assume further that we use the Euler ODE solver, and the pre-trained score model matches the ground truth, i.e., Vt ∈ [E,T] : sΦ(x,t) = V log Pt(x). Then,

where the expectationis taken with respect to x\~Pdata, n \~U[1,N-1],and Xtn+1\~ N(x;tn+1I). The consistency training objective, denoted by LNr(0,0-), is defined as

where z \~ N(O,1). Moreover, LNr(0,0-) ≥ O(△t) ifinfN LND(0,0-;Φ) > 0.

Proof. With Taylor expansion, we have

$$\tag{15}$$

Then, we apply Lemma l to Eq. (15) and use Taylor expansion in the reverse direction to obtain

$$\tag{16}$$

where(i)isduetftotaleptationdisstributedacdigtesada distribution. This implies LND(0,0-;Φ) = LNr(0,0-) + o(△t) and thus completes the proof for Eq. (9). Moreover, we have LN(0,0-) ≥ O(△t) whenever infN LN(0,0-;Φ) > 0. Otherwise,LN(0,0-)

B. Continuous-Time Extensions

The consistency distilation and consistency training objectives can be generalized to hold for infinite time steps (N → ) under suitable conditions.

B.1. Consistency Distillation in Continuous Time

Depending on whether 0- = 0 or 0- = stopgrad(0)(same as setting μ = O),there are two possible continuous-time extensions for the consistency distillation objective LN(0,0-; ). Given a twice continuously diffrentiable metric function d(x,y), we define G(x) as a matrix, whose (i,j)-th entry is given by

Similarly, we define H(x) as

The matrices G and H play a crucial role in forming continuous-time objectives for consistency distillation. Additionally, we denotetheJacobanoffe(x,t)withrespectoxsf(t).

When 0- = 0 (with no stopgrad operator), we have the following theoretical result.

Theorem 3. Let tn = N-1
differentiable with bounded third derivatives,and fe is twice continuously differentiable with bounded frst and second derivatives.AsmefurhertattheweightingfunctionX()isuded,andsupxte[e]|ls(x,t)llO.Supposeweuse the Euler ODE solver. Then,

$$\tag{17}$$

where

$$\tag{18}$$

Here the expectation above is taken over X \~ Pdatas t \~U[e,T], and Xt \~ N(x,t²I).

Proof. First, we can derive the following equation with Taylor expansion:

$$\tag{19}$$

Applying Taylor expansion again, we get

$$\tag{20}$$

where we obtain (i) by first expanding d(fe(Xtn+1,tn+1),:) to second order, and then noting that d(x,x) = O and Vyd(x,y)ly=x = O. We obtain (i) using Eq. (19). By taking the limit for both sides of Eq. (2O) as △t → O, we arrive at Eq. (17), which completes the proof. □

Remark 1. Although Theorem 3 assumes uniform timesteps and the Euler ODE solver for technical simplicity, we believe an analogous result can be derived without those assumptions,as all time boundaries and ODE solvers should perform similarly as N → o. We leave a more general version of Theorem 3 as future work.

Remark 2. Theorem 3 implies that consistency models can be trained by minimizing LD(0,0; Φ). In particular, when d(x,y) = |/x - yl², we have

$$\tag{21}$$

However, this continuous-time objective requires computing Jacobian-vector products as a subroutine to evaluate the loss function, which can be slow and laborious to implement in deep learning frameworks that do not support forward-mode automatic differentiation.

Remark 3. If fe(x,t) matches the ground truth consistency function for the empirical PF ODE of sφ(x,t), then

and therefore LCD(0,0;Φ)= 0. This can be proved by noting that fe(xt,t) = Xe for allt ∈ [ε,T], and then taking the time-derivative of this identity:

The above observation provides another motivation for LD(0,0; Φ), as it is minimized if and only if the consistency model matches the ground truth consistency function.

For some metric functions,such as the ll norm, the Hessian G(x) is zero so Theorem 3 is vacuous. Below we show that a non-vacuous statement holds for the l1 norm with just a small modification of the proof for Theorem 3.

Theorem 4. Let tn = entiable with bounded first and second derivatives. Assume further that the weighting function X(-) is bounded, and suPx,te[e,T] lls𝜙(x,t)ll2

$$\tag{22}$$

where

where the expectation above is taken over X \~ pdata, t \~ U[e,T], and Xt \~ N(x,t²I).

Proof. We have

$$\tag{23}$$

where (i) is obtained by plugging Eq.(19) into the previous equation. Taking the limit for both sides of Eq. (23) as △t → 0 leads to Eq. (22),which completes the proof. □

Remark 4. According to Theorem 4, consistency models can be trained by minimizing LCD,e1(0,0; Φ). Moreover, the same reasoning in Remark3can be applied to show that LCD,e1(0,0;Φ)= Oifand only if fe(xt,t)= Xefor allxt ∈ Rd and te[e,T].

In the second case where 0- = stopgrad(0), we can derive a pseudo-objective whose gradient matches the gradient of LN(0,0-;Φ) in thelimit of N→ oo. Minimizing this pseudo-objective with gradient descent gives another way to train consistency models via distillation. This pseudo-objective is provided by the theorem below.

Theorem 5. Let tn = (n-1) ously differentiable with bounded third derivatives, and fe is twice continuously differentiable with bounded first and second derivatives. Assume further that the weighting function X(-) is bounded, supx,te[e.T] lls(x,t)ll2 < o， and supx,te[e,T] l/Vθfe(x,t)ll2 < oo. Suppose we use the Euler ODE solver, and 0- = stopgrad(0). Then,

$$\tag{24}$$

where

$$\tag{25}$$

Here the expectation above is taken over X \~ pdatas t \~ U[e,T], and Xt \~ N(x,t²I).

Proof. First, we leverage Taylor series expansion to obtain

$$\tag{26}$$

where (i) is derived by first expanding d(.,fe-(xn,tn)） to second order, and then noting that d(x,x) = O and Vyd(y,x)ly=x = O. Next, we compute the gradient of Eq. (26) with respect to 0 and simplify the result to obtain

Here (i) results from the chain rule,and (ii) follows from Eq.(19) and fe(x,t) = fe-(x,t), since 0- = stopgrad(0). Taking the limit for both sides of Eq.(27) as △t → O yields Eq.(24),which completes the proof. □

Remark 5. When d(x,y) = |lx -yll²2,the pseudo-objective LD(0,0-;Φ)can be simplified to

$$\tag{28}$$

Remark 6. The objective LD(0,0-;Φ) defined in Theorem 5 is only meaningful in terms of its gradient-one cannot measure the progress of training by tracking the value of LD(0,0-;Φ), but can still apply gradient descent to this objective to distill consistency models from pre-trained diffusion models. Because this objective is not a typical loss function,we refer to it as the “pseudo-objective"for consistency distillation.

Remark 7. Following the same reasoning in Remark 3，we can easily derive that LD(0,0-;Φ) = O and VθLD(0,0-;Φ) = 0 if fe(x,t) matches the ground truth consistency function for the empirical PF ODE that involves S(x,t). However the converse does not hold true in general. This distinguishes LD(0,0-;Φ)from L(0,0;Φ), the latter of which is a true loss function.

B.2. Consistency Training in Continuous Time

A remarkable observation is that the pseudo-objective in Theorem 5 can be estimated without any pre-trained diffusion models,which enables direct consistency training of consistency models.More precisely,we have the following result.

Theorem 6. Let tn = (n-1) (n-1)(T-e) + ∈ and △t = uously differentiable with bounded third derivatives,and fe is twice continuously differentiable with bounded frst and second derivatives. Assume further that the weighting function X(-) is bounded, E[lVlog Ptn(Xtn)ll2] < O, and suPx,te[e,T] lVθfe(x,t)ll2 < o. Suppose we use the Euler ODE solver and 0- = stopgrad(0). Then,

$$\tag{29}$$

where

$$\tag{30}$$

Here the expectation above is taken over X \~ Pdata, t \~U[e,T], and Xt \~ N(x,t²I).

Proof. The proof mostly follows that of Theorem 5. First, we leverage Taylor series expansion to obtain

$$\tag{31}$$

where (i) is derived by first expanding d(:, fe-(x + tnz,tn)） to second order, and then noting that d(x,x) = O and Vyd(y,x)ly=x = O. Next, we compute the gradient of Eq. (31) with respect to 0 and simplify the result to obtain

$$\tag{32}$$

Here (i) results from the chain rule,and (ii) follows from Taylor expansion. Taking the limit for both sides of Eq. (32) as △t → O yields Eq. (29),which completes the proof. □

(a) Consistency Distillation
(b) Consistency Training
Figure 7: Comparing discrete consistency distillation/training algorithms with continuous counterparts.

Remark 8. Note that Lr(0,0-) does not depend on the difusion model parameter Φ and hence can be optimized without any pre-trained diffusion models.

Remark 9. When d(x,y) = |x -yl/2,the continuous-time consistency training objective becomes

$$\tag{33}$$

Remark 10. Under suitable regularity conditions, we can combine the insights of Theorems 2 and 6 to derive

As in Theorems 2 and 6, here Φ represents diffusion model parameters that satisfy sΦ(x,t) = Vlog Pt(x), and 0- = stopgrad(0).

Remark 11. Similar to LCD(0,0-;Φ) in Theorem 5, Lr(0,0-) is a pseudo-objective; one cannot track training by monitoring the value of Lr(0,0-)， but can still apply gradient descent on this loss function to train a consistency model fe(x,t) directly from data. Moreover, the same observation in Remark 7 holds true: L(0,0-) = O and VeLr(0,0-) = O if fe(x,t) matches the ground truth consistency function for the PF ODE.

B.3.Experimental Verifications

To experimentally verify the efficacy of our continuous-time CD and CT objectives,we train consistency models with a variety of loss functions on CIFAR-1O. All results are provided in Fig.7. Unless otherwise noted, models are trained with hyperparameters given in Table 3. We occasionally modify some hyperparameters for improved performance. For distillation, we compare the following objectives:

· CD (l2): Consistency distilltion CND with N = 18 and the l2 metric.

· CD (l1): Consistency distillation LND with N = 18 and the l1 metric. We set the learning rate to 2e-4.

· CD (LPIPS): Consistency distilation CN with N = 18 and the LPIPS metric.

·CD (l2): Consistency distilation LD in Theorem 3 with the l2 metric. We set the learning rate to le-3 and dropout to 0.13.

· CD (l1): Consistency distilation L in Theorem 4 with the l1 metric. We set the learning rate to 1e-3and dropout to 0.3.

· CD (stopgrad,C2): Consistency distilation L&D in Theorem 5 with the l2 metric. We set the learning rate to 5e-6.

Table 3: Hyperparameters used for training CD and CT models

Hyperparameter	CIFAR-10		ImageNet 64×64		LSUN 256 × 256
Hyperparameter	CD	CT	CD	CT	CD	CT
Learning rate	4e-4	4e-4	8e-6	8e-6	1e-5	1e-5
Batch size	512	512	2048	2048	2048	2048
从	0		0.95		0.95
μ0		0.9		0.95		0.95
S0		2		2		2
S1		150		200		150
EMA decay rate	0.9999	0.9999	0.999943	0.999943	0.999943	0.999943
Training iterations	800k	800k	600k	800k	600k	1000k
Mixed-Precision (FP16)	No	No	Yes	Yes	Yes	Yes
Dropout probability	0.0	0.0	0.0	0.0	0.0	0.0
Number of GPUs	8	8	64	64	64	64

· CD (stopgrad,LPIPS): Consistency distillation LD in Theorem 5 with the LPIPS metric.We set the learning rate to 5e-6.

We did not investigate using the LPIPS metric in Theorem 3 because minimizing the resulting objective would require back-propagating through second order derivatives of the VGG network used in LPIPS, which is computationally expensive and prone to numerical instability. As revealed by Fig.7a, the stopgrad version of continuous-time distillation (Theorem 5) works better than the non-stopgrad version (Theorem 3) for both the LPIPS and l2 metrics,and the LPIPS metric works the best for all distillation approaches.Additionally,discrete-time consistency distillation outperforms continuous-time consistency distillation, possibly due to the larger variance in continuous-time objectives,and the fact that one can use effective higher-order ODE solvers in discrete-time objectives.

For consistency training (CT), we find it important to initialize consistency models from a pre-trained EDM model in order to stabilize training when using continuous-time objectives. We hypothesize that this is caused by the large variance in our continuous-time loss functions.For fair comparison, we thus initialize all consistency models from the same pre-trained EDM model on CIFAR-10 for both discrete-time and continuous-time CT, even though the former works well with random initialization. We leave variance reduction techniques for continuous-time CT to future research.

We empirically compare the following objectives:

· CT (LPIPS): Consistency training CN with N =120 and the LPIPS metric. We set the learning rate to 4e-4,and the EMA decay rate for the target network to 0.99. We do not use the schedule functions for N and μ here because they cause slower learning when the consistency model is initialized from a pre-trained EDM model.

· CT∞ (l2): Consistency training L&T with the l2 metric. We set the learning rate to 5e-6.

· CT (LPIPS): Consistency training L& with the LPIPS metric. We set the learning rate to 5e-6.

As shown in Fig. 7b, the LPIPS metric leads to improved performance for continuous-time CT. We also find that continuoustime CT outperforms discrete-time CT with the same LPIPS metric. This is likely due to the bias in discrete-time CT, as △t > Oin Theorem 2 for discrete-time objectives, whereas continuous-time CT has no bias since it implicitly drives △t to 0.

C. Additional Experimental Details

Model ArchitecturesWe follow Song et al. (2O21); Dhariwal & Nichol (2021) for model architectures. Specifically, we use the NCSN++ architecture in Song et al. (2O21) for all CIFAR-10 experiments,and take the corresponding network architectures from Dhariwal & Nichol (2021) when performing experiments on ImageNet 64 × 64,LSUN Bedroom 256 × 256 and LSUN Cat 256 × 256.

Parameterization for Consistency ModelsWe use the same architectures for consistency models as those used for EDMs.The only difference is we slightly modify the skip connections in EDM to ensure the boundary condition holds for consistency models. Recallthat in Section 3 we propose to parameterize a consistency model in the following form:

In EDM (Karras et al., 2022),authors choose

where Odata = O.5. However, this choice of Cskip and Cout does not satisfy the boundary condition when the smallest time instant é ≠ O. To remedy this issue,we modify them to

which clearly satisfies Cskip(ε) = 1 and Cout(ε) = 0.

Schedule Functions for Consistency Training As discussed in Section 5,consistency generation requires specifying schedule functions N(·)and μ(·) for best performance. Throughout our experiments,we use schedule functions that take the form below:

where K denotes the total number of training iterations,so denotes the initial discretization steps,s1 > so denotes the target discretization steps at the end of training,and μo > O denotes the EMA decay rate at the beginning of model training.

Training DetailsIn both consistency distillation and progressive distillation, we distill EDMs (Karras et al.,2O22). We trained these EDMs ourselves according to the specifications given in Karras et al. (2O22). The original EDM paper did not provide hyperparameters for the LSUN Bedroom 256 × 256 and Cat 256 × 256 datasets, so we mostly used the same hyperparameters as those for the ImageNet 64 × 64 dataset. The difference is that we trained for 6OOk and 3OOk iterations for the LSUN Bedroom and Cat datasets respectively, and reduced the batch size from 4096 to 2048.

We used the same EMA decay rate for LSUN 256 × 256 datasets as for the ImageNet 64 × 64 dataset. For progressive distillation, we used the same training settings as those described in Salimans & Ho (2O22) for CIFAR-1O and ImageNet 64 × 64. Although the original paper did not test on LSUN 256 × 256 datasets,we used the same settings for ImageNet 64 × 64 and found them to work well.

In alldistilltion experiments,we initialized the consistency model with pre-trained EDM weights.For consistency training, we initialized the model randomly, just as we did for training the EDMs.We trained all consistency models with the Rectified Adam optimizer (Liu et al.,2O19), with no learning rate decay or warm-up,and no weight decay. We also applied EMA to the weights of the online consistency models in both consistency distillation and consistency training,as well as to the weights of the training online consistency models according to Karras et al. (2022). For LSUN 256 × 256 datasets, we chose the EMA decay rate to be the same as that for ImageNet 64 × 64, except for consistency distilltion on LSUN Bedroom 256 × 256, where we found that using zero EMA worked better.

When using the LPIPS metric on CIFAR-10 and ImageNet 64 × 64, we rescale images to resolution 224 × 224 with bilinear upsampling before feeding them to the LPIPS network. For LSUN 256 × 256, we evaluated LPIPS without rescaling inputs. In addition,we performed horizontal flips for data augmentation for all models and on all datasets.We trained all models on a cluster of Nvidia A1Oo GPUs.Additional hyperparameters for consistency training and distillation are listed in Table 3.

D. Additional Results on Zero-Shot Image Editing

With consistency models, we can perform a variety of zero-shot image editing tasks.As an example, we present additional results on colorization (Fig. 8),super-resolution (Fig. 9), inpainting (Fig.10),interpolation (Fig.11),denoising (Fig.12), and stroke-guided image generation (SDEdit, Meng et al. (2O21),Fig.13). The consistency model used here is trained via consistency distillation on the LSUN Bedroom 256 × 256.

Algorithm 4 Zero-Shot Image Editing
1: Input: Consistency model fe(·,·),sequence of time points t1 > t2 > ... > tN,reference image y,invertible linear
transformation A,and binary image mask Ω
2: y ← A-1[(Ay)0(1-Ω)+0Ω]
3: Sample x \~ N(y,t²I)
4: X ← fe(x,t1)
5: x ← A-1[(Ay)O(1-Ω)+ (Ax)①Ω]
6: for n = 2 to N do
7:Samplex \~ N(x,(tn-ε2)I)
8:x←fe(x,tn)
9:x ← A-1[(Ay)⊙(1-Ω)+(Ax)①Ω]
10: end for
11: Output: x

All these image editing tasks, except for image interpolation and denoising,can be performed via a small modification to the multistep sampling algorithm in Algorithm 1. The resulting pseudocode is provided in Algorithm 4. Here y is a reference image that guides sample generation,Ω is a binary mask, O computes element-wise products,and A is an invertible linear transformation that maps images into a latent space where the conditional information in y is infused into the iterative generation procedure by masking with Ω. Unless otherwise stated, we choose

in our experiments, where N = 40 for LSUN Bedroom 256 × 256.

Below we describe how to perform each task using Algorithm 4.

InpaintingWhen using Algorithm 4 for inpainting, we let y be an image where missing pixels are masked out, Ω be a binary mask where 1 indicates the missing pixels,and A be the identity transformation.

ColorizationThe algorithm for image colorization is similar, as colorization becomes a special case of inpainting once we transform data intoa decoupled space. Specificall,let y ∈Rhxwx3 be a gray-scale image that we aim tocolorize, where all channels of y are assumed to be the same,i.e., y[:,:,O] = y[:,:,1] = y[:, :,2] in NumPy notation. In our experiments, each channel of this gray scale image is obtained from a colorful image by averaging the RGB channels with

We define Ω ∈ {0,1}h×w x3 to be a binary mask such that

Let Q ∈ R3×3 be an orthogonal matrix whose first column is proportional to the vector (0.2989,0.5870,0.1140). This orthogonal matrix can be obtained easily via QR decomposition, and we use the following in our experiments

We then define the linear transformation A : x ∈ Rh×w ×3 → y ∈ Rh×w×3, where

Because Qis orthogonal, theinversion A-1 : y∈Rxw → x ∈Rh×w×3 is easy to compute, where

With A and Ω defined as above, we can now use Algorithm 4 for image colorization.

Super-resolutionWith a similar strategy, we employ Algorithm 4 for image super-resolution. For simplicity, we assume that the down-sampled image is obtained by averaging non-overlapping patches of size p × p. Suppose the shape of full resolutionimages ish × w × 3.Lety∈Rhxw3denotealow-resolutionimage naivelyup-sampled tofullrsolution, whrepielshveedoall{ mask such that

Similar to image colorization,super-resolution requires an orthogonal matrix Q eRp2×p2 whose first column is (1/p,1/p,... ,1/p). This orthogonal matrix can be obtained with QR decomposition. To perform super-resolution, we define the linear transformation A : x ∈ Rhxwx3 → y ∈ Rh/pxw/pxp2×3, where

The inversetransformation A-1 : y ∈ Rh/pxw/pxp2×3 → x ∈Rhxwx3 is easy to derive, with

Above definitions of A and Ω allow us to use Algorithm 4 for image super-resolution.

Stroke-guided image generationWe can also use Algorithm 4 for stroke-guided image generation as introduced in SDEdit (Meng et al.,2021). Specifically, we let y ∈ Rhxw ×3 be a stroke painting. We set A = I,and define Ω ∈ Rh x w ×3 as a matrix of ones. In our experiments, we set t1 = 5.38 and t2 = 2.24, with N = 2.

Denoising It is possible to denoise images perturbed with various scales of Gaussian noise using a single consistency model. Suppose the input image x is perturbed with N(O; g²I). As long as σ ∈ [e,T], we can evaluate fe(x,σ) to produce the denoised image.

InterpolationWe can interpolate between two images generated by consistency models. Suppose the first sample X1 is produced by noise vector Z1,and the second sample X2 is produced by noise vector Zz. In other words, X1 = fe(z1,T) and X2 = fe(z2,T). To interpolate between X1 and X2, we first use spherical linear interpolation to get

zz2

E. Additional Samples from Consistency Models

We provide additional samples from consistency distillation (CD) and consistency training (CT) on CIFAR-1O (Figs. 14 and 18),ImageNet 64 × 64 (Figs.15 and 19),LSUN Bedroom 256 × 256 (Figs.16 and 20) and LSUN Cat 256 × 256 (Figs. 17 and 21).

Figure 8: Gray-scale images (left), colorized images by a consistency model (middle),and ground truth (right).

Figure 9: Downsampled images of resolution 32 × 32 (left),full resolution (256 × 256) images generated by a consistency model (middle),and ground truth images of resolution 256 × 256 (right).

Figure 10: Masked images (left), imputed images by a consistency model (middle),and ground truth (right).

Figure 11: Interpolating between leftmost and rightmost images with spherical linear interpolation. All samples are generated by a consistency model trained on LSUN Bedroom 256 × 256.

Figure 12: Single-step denoising with a consistency model. The leftmost images are ground truth. For every two rows, the top row shows noisy images with different noise levels, while the bottom row gives denoised images.

Figure 13: SDEdit with a consistency model. The leftmost images are stroke painting inputs.Images on the right side are the results of stroke-guided image generation (SDEdit).

(a) EDM (FID=2.04)

(b) CD with single-step generation (FID=3.55)

(c) CD with two-step generation (FID=2.93)
Figure 14: Uncurated samples from CIFAR-10 32 × 32. All corresponding samples use the same initial noise.

(a) EDM (FID=2.44)

(b) CD with single-step generation (FID=6.20)

(c) CD with two-step generation (FID=4.70)
Figure 15: Uncurated samples from ImageNet 64 × 64. All corresponding samples use the same initial noise.

(a) EDM (FID=3.57)

(b) CD with single-step generation (FID=7.80)

(c) CD with two-step generation (FID=5.22)
Figure 16: Uncurated samples from LSUN Bedroom 256 × 256. All corresponding samples use the same initial noise.

(a) EDM (FID=6.69)

(b) CD with single-step generation (FID=10.99)

(c) CD with two-step generation (FID=8.84)
Figure 17: Uncurated samples from LSUN Cat 256 × 256. Allcorresponding samples use the same initial noise.

(a) EDM (FID=2.04)

(b) CT with single-step generation (FID=8.73)

(c) CT with two-step generation (FID=5.83)
Figure 18: Uncurated samples from CIFAR-10 32 × 32. All corresponding samples use the same initial noise.

(a) EDM (FID=2.44)

(b) CT with single-step generation (FID=12.96)

(c) CT with two-step generation (FID=11.12)
Figure 19: Uncurated samples from ImageNet 64 × 64. All corresponding samples use the same initial noise.

(a) EDM (FID=3.57)

(b) CT with single-step generation (FID=16.00)

(c) CT with two-step generation (FID=7.80)
Figure 20: Uncurated samples from LSUN Bedroom 256 × 256. All corresponding samples use the same initial noise.

(a) EDM (FID=6.69)

(b) CT with single-step generation (FID=20.70)

(c) CT with two-step generation (FID=11.76)
Figure 21: Uncurated samples from LSUN Cat 256 × 256. Allcorresponding samples use the same initial noise.

致终身学习者社群

行业报告/思维导图/电子书/资讯情报

收钱文案、增长黑客、产品运营、品牌企划、营销战略、办公软件、会计财务、广告设计、摄影修图、视频剪辑、直播带货、电商运营、投资理财、汽车房产、餐饮烹饪、职场经验、演讲口才、风水命理、心理思维、恋爱情趣、美妆护肤、健身瘦身、格斗搏击、漫画手绘、声乐训练、自媒体打造、效率软件工具、游戏影音...

Consistency Models

常见问题