Diffusion Probabilistic Modeling for Video Generation

Abstract

Denoising diffusion probabilistic models are a promising new class of generative models that mark a milestone in high-quality image generation. This paper showcases their ability to sequentially generate video, surpassing prior methods in perceptual and probabilistic forecasting metrics. We propose an autoregressive, end-to-end optimized video diffusion model inspired by recent advances in neural video compression. The model successively generates future frames by correcting a deterministic next-frame prediction using a stochastic residual generated by an inverse diffusion process. We compare this approach against six baselines on four datasets involving natural and simulation-based videos. We find significant improvements in terms of perceptual quality and probabilistic frame forecasting ability for all datasets.

1. Introduction

The ability to anticipate future frames of a video is intuitive for humans but challenging for a computer [1]. Applications of such video prediction tasks include anticipating events [2], model-based reinforcement learning [3], video interpolation [4], predicting pedestrians in traffic [5], precipitation nowcasting [6], neural video compression [7,8,9,10,11,12], and many more.

The goals and challenges of video prediction include (i) generating multi-modal, stochastic predictions that (ii) accurately reflect the high-dimensional dynamics of the data long-term while (iii) identifying architectures that scale to high-resolution content without blurry artifacts. These goals are complicated by occlusions, lighting conditions, and dynamics on different temporal scales. Broadly speaking, models relying on sequential variational autoencoders [13,14,15] tend to be stronger in goals (i) and (ii), while sequential extensions of generative adversarial networks [16,17,18] tend to perform better in goal (iii). A probabilistic method that succeeds in all three desiderata on high-resolution video content is yet to be found.

Recently, diffusion probabilistic models have achieved considerable progress in image generation, with perceptual qualities comparable to GANs while avoiding the optimization challenges of adversarial training [19,20,21,22,23]. In this paper, we extend diffusion probabilistic models for stochastic video generation. Our ideas are inspired by the principles of predictive coding [24,25] and neural compression algorithms [26] and draw on the intuition that residual errors are easier to model than dense observations [27]. Our architecture relies on two prediction steps: first, we employ a deterministic convolutional RNN to deterministically predict the next frame conditioned on a sequence of frames. Second, we correct this prediction by an additive residual generated by a denoising diffusion process, also conditioned on a temporal context (see Figure 1a,b). This approach is scalable to high-resolution video, stochastic, and relies on likelihood-based principles. Our ablation studies strongly suggest that predicting video frame residuals instead of naively predicting the next frames improves generative performance. By investigating our architecture on various datasets and comparing it against multiple baselines, we prove superior results on both probabilistic (CRPS) and perceptual (LPIPS, FID) metrics. In more detail, our achievements are as follows:

1.

We show how to use diffusion probabilistic models to generate videos autoregressively. This enables a new path towards probabilistic video forecasting while achieving perceptual qualities better than or comparable with likelihood-free methods such as GANs.

2.

We also study our model based on three metrics, FVD/LPIPS/CRPS, that can cover both forecasting ability and perceptual quality. The result demonstrates that our model performs better than modern GAN and VAE baselines such as IVRNN, SVG-LP, SLAMP, RetroGAN, DVD-GAN and FutureGAN [14,15,16,17,28,29].

3.

Our ablation studies demonstrate that modeling residuals from the predicted next frame yields better results than directly modeling the next frames. This observation is consistent with recent findings in neural video compression. Figure 1a summarizes the main idea of our approach (Figure 1b has more details).

The structure of our paper is as follows. We first describe our method, which is followed by a discussion of our experimental findings along with ablation studies. We then discuss connections to the literature and summarize our contributions.

2. Related Work

Our paper combines ideas from video generation, diffusion probabilistic models, and neural video compression. As follows, we discuss related work along these lines.

2.1. Video Generation Models

Since the advent of modern deep learning, video generation and prediction has been a topic of ongoing interest; see [1] and this paper’s introduction. Videos can be generated based on side information of various types, such as images [30,31], text [32,33,34,35,36,37] or other videos [38]. Alternatively, videos can also be generated unconditionally, e.g., from white noise [29,39,40]. This survey focuses on conditional video generation, where one conditions the generation of future frames in the context of past frames.

Conditional video prediction is sometimes treated as a supervised problem, where the focus is often on error metrics such as PSNR and SSIM [41,42,43]. Early works leverage deterministic methods to predict the most likely next frames [42,43,44,45,46,47]. Generally speaking, the downside of supervised approaches is that real-world videos display multi-modal behavior, i.e., the future is not uniquely predictable from the past (e.g., a traffic light may switch from yellow to red or green, a new object may or may not enter the scene, etc.). Treating video prediction as a supervised problem can, therefore, lead to mode averaging, perceived as blurriness.

Most recent methods focus on stochastic generation using deep generative models. In contrast to learning the average video dynamics, these methods try to match the conditional distribution of future frames given the past, typically by minimizing a divergence measure (as in GANs) or by optimizing a variational bound to a model’s log-likelihood (as in VAEs). Consequently, the evaluation has shifted to held-out likelihoods or perceptual metrics, such as FID or LPIPS. A large body of video generation research relies on variational deep sequential latent variable models [13,28,29,48,49,50,51,52,53,54,55,56], to name a few. These works often draw on earlier works for modeling stochastic dynamics, e.g., Bayer and Osendorfer [57], Chung et al. [58], who included latent variables in recurrent neural networks. Later work [14] extended the sequential VAE by incorporating more expressive priors conditioned on a longer frame context. IVRNN [15] further enhanced the generation quality by working with a hierarchy of latent variables, which, to our knowledge, is currently the best end-to-end trained sequential VAE model that can be further refined by greedy fine-tuning [59]. Normalizing flow-based models for video have been proposed while typically suffering from high demands on memory and compute [49]. Some works [25,27,60,61] explored the use of residuals for improving video generation in sequential VAEs but did not achieve state-of-the-art results.

Overall, the downside of VAE-based models is that they are trained to reconstruct the data. Blau and Michaeli [62] theoretically showed that generative models are typically in conflict between data reconstruction tasks and achieving a high degree of realism (defined as matching the target distribution unconditionally, without artifacts). This suggests that VAEs may not be the final answer when it comes to video prediction.

Another line of sequential models relies on GANs [16,17,40,63,64,65], which—at inference time—can be either deterministic or stochastic. In contrast to VAE-based models, these models tend to show fewer blurry artifacts. A downside of GANs is that their loss function is mode-seeking (as opposed to mass covering), meaning that the data distribution is not covered at sufficient breadth, reflected in their typically worse performance in probabilistic distribution matching and forecasting metrics.

2.2. Diffusion Probabilistic Models

DDPMs have recently shown impressive performance in high-fidelity image generation. Sohl-Dickstein et al. [19] first introduced and motivated this model class by drawing on a non-equilibrium thermodynamic perspective. Song and Ermon [20] proposed a single-network model for score estimation, using annealed Langevin dynamics for sampling. Furthermore, Song et al. [22] used stochastic differential equations (related to diffusion processes) to train a network to transform random noise into the data distribution.

DDPM by Ho et al. [21] is the first instance of a diffusion model scalable to high-resolution images. This work also showed the equivalence of DDPM and denoising score-matching methods described above. Subsequent work includes extensions of these models to image super-resolution [66] or hybridizing these models with VAEs [67]. Apart from the traditional computer vision tasks, diffusion models were proven to be effective in audio synthesis [68,69], while Luo and Hu [70] hybridized normalizing flows and diffusion model to generative 3D point cloud samples.

To the best of our knowledge, TimeGrad [71] is the first sequential diffusion model for time-series forecasting. Their architecture was not designed for video but for traditional lower-dimensional correlated time-series datasets. Two concurrent preprints also study a video diffusion model [72,73]. Both works are based on alternative architectures and focus primarily on perceptual metrics.

We will extend this survey to the camera-ready version.

2.3. Neural Video Compression Models

Video compression models typically employ frame prediction methods optimized for minimizing code length and distortion. In recent years, sequential generative models were proven to be effective on video compression tasks [7,8,9,10,11]. Some of these models show impressive rate-distortion performance with hierarchical structures that separately encode the prediction and error residual. Although compression models have different goals from generative models, both benefit from predictive sequential priors [26]. Note, however, that these models are ill-suited for generation since compression models typically have a small spatio-temporal context and are constructed to preserve local information rather than to generalize [12].

3. A Diffusion Probabilistic Model for Video

We begin by reviewing the relevant background on diffusion probabilistic models. We then discuss our design choices for extending these models to sequential models for video.

3.1. Background on Diffusion Probabilistic Models

Denoising diffusion probabilistic models (DDPMs) are a recent class of generative models with promising properties [19,21]. Unlike GANs, these models rely on the maximum likelihood training paradigm (and are thus stable to train) while producing samples of comparable perceptual quality as GANs [74].

Similar to hierarchical variational autoencoders (VAEs) [75], DDPMs are deep latent variable models that model data ${\mathbf{x}}_0$ in terms of an underlying sequence of latent variables ${\mathbf{x}}_{1:N}$ such that $p_{\theta }\left({\mathbf{x}}_0\right)=\int p_{\theta }\left({\mathbf{x}}_{0:N}\right)d{\mathbf{x}}_{1:N}$. The main idea is to impose a diffusion process on the data that incrementally destroys the structure. The diffusion process’s incremental posterior yields a stochastic denoising process that can be used to generate structure [19,21]. The forward, or diffusion process is given by

$$\begin{array}{cc}\hfill q\left({\mathbf{x}}_{1:N}\right|{\mathbf{x}}_0)& =\prod _{n=1}^{N}q\left({\mathbf{x}}_n\right|{\mathbf{x}}_{n-1});\hfill \\ \hfill q\left({\mathbf{x}}_n\right|{\mathbf{x}}_{n-1})& =\mathcal{N}\left({\mathbf{x}}_n\right|\sqrt{1-{\beta }_n}{\mathbf{x}}_{n-1},{\beta }_n\mathbf{I}).\hfill \end{array}$$

(1)

Besides a predefined incremental variance schedule with ${\beta }_n\in (0,1)\mathrm{for}n\in \{1,\cdots ,N\}$, this process is parameter-free [20,21]. The reverse process is called denoising process,

$$\begin{array}{cc}\hfill p_{\theta }\left({\mathbf{x}}_{0:N}\right)& =p\left({\mathbf{x}}_N\right)\prod _{n=1}^N{p}_{\theta }\left({\mathbf{x}}_{n-1}\right|{\mathbf{x}}_n);\hfill \\ \hfill p_{\theta }\left({\mathbf{x}}_{n-1}\right|{\mathbf{x}}_n)& =\mathcal{N}\left({\mathbf{x}}_{n-1}\right|M_{\theta }({\mathbf{x}}_n,n),\gamma \mathbf{I}).\hfill \end{array}$$

(2)

The reverse process can be thought of as approximating the posterior of the diffusion process. Typically, one fixes the covariance matrix (with hyperparameter $\gamma$) and only learns the posterior mean function $M_{\theta }({\mathbf{x}}_n,n)$. The prior $p\left({\mathbf{x}}_N\right)=\mathcal{N}(\mathbf{0},\mathbf{I})$ is typically fixed. The parameter $\theta$ can be optimized by maximizing a variational lower bound on the log-likelihood, $L_{\mathrm{variational}}={\mathbb{E}}_q[-log\frac{p_{\theta }\left({\mathbf{x}}_{0:N}\right)}{q\left({\mathbf{x}}_{1:N}\right|{\mathbf{x}}_0)}]$. This bound can be efficiently estimated by stochastic gradients by subsampling time steps n at random since the marginal distributions $q\left({\mathbf{x}}_n\right|{\mathbf{x}}_0)$ can be computed in the closed form [21].

In this paper, we use a simplified loss due to Ho et al. [21], who showed that the variational bound could be simplified to the following denoising score-matching loss,

$$\begin{array}{cc}\hfill L\left(\theta \right)& ={\mathbb{E}}_{{\mathbf{x}}_0,n,ϵ}\left|\right|ϵ-f_{\theta }({\mathbf{x}}_n,n){\left|\right|}^2\hfill \\ \hfill \mathrm{where}{\mathbf{x}}_n& =\sqrt{{\overline{\alpha }}_n}{\mathbf{x}}_0+\sqrt{1-{\overline{\alpha }}_n}ϵ.\hfill \end{array}$$

(3)

We therefore define ${\overline{\alpha }}_n={\prod }_{i=1}^n(1-{\beta }_i)$, where $1-{\overline{\alpha }}_n$ is the variance schedule whose square root is the standard deviation to reparametrize the injected noise $ϵ$. We note that this schedule ensures ${\overline{\alpha }}_n\to 0$ when $n\to N$. The intuitive explanation of this loss is that $f_{\theta }$ tries to predict the noise $ϵ\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ at the denoising step n [21]. Once the model is trained, it can be used to generate data by ancestral sampling, starting with a draw from the prior $p\left({\mathbf{x}}_N\right)$ and successively generating increased structure through an annealed Langevin dynamics procedure [20,22].

3.2. Residual Video Diffusion Model

Experience shows that it is often simpler to model differences from our predictions than the predictions themselves. For example, masked autoregressive flows [76] transform random noise into an additive prediction error residual, and boosting algorithms train a sequence of models to predict the error residuals of earlier models [77]. Residual errors also play an important role in modern theories of the brain. For example, predictive coding [24] postulates that neural circuits estimate probabilistic models of other neural activity, iteratively exchanging information about error residuals. This theory has interesting connections to VAEs [25,27] and neural video compression [9,11], where one also compresses the residuals to the most likely next-frame predictions.

This work uses a diffusion model to generate residual corrections to a deterministically predicted next frame, adding stochasticity to the video generation task. Both the deterministic prediction as well as the denoising process are conditioned on a long-range context provided by a convolutional RNN. We call our approach “Residual Video Diffusion” (RVD). Details will be explained next.

Notation. We consider a frame sequence ${\mathbf{x}}^{0:T}$ and a set of latent variables ${\mathbf{y}}^{1:T}\equiv {\mathbf{y}}_{0:N}^{1:T}$ specified by a diffusion process over the lower indices. We refer to ${\mathbf{y}}_0^{1:T}$ as the (scaled) frame residuals.

Generative Process. We consider a joint distribution over ${\mathbf{x}}^{0:T}$ and ${\mathbf{y}}^{1:T}$ of the following form:

$$\begin{array}{c}\hfill p({\mathbf{x}}^{0:T},{\mathbf{y}}^{1:T})=p\left({\mathbf{x}}^0\right)\prod _{t=1}^{T}p\left({\mathbf{x}}^t\right|{\mathbf{y}}^t,{\mathbf{x}}^{<t})p\left({\mathbf{y}}^t\right|{\mathbf{x}}^{<t}).\end{array}$$

(4)

We first specify the data likelihood term $p\left({\mathbf{x}}^t\right|{\mathbf{y}}^t,{\mathbf{x}}^{<t})$, which we model autoregressively as a Masked Autoregressive Flow (MAF) [76] applied to the frame sequence. This involves an autoregressive prediction network outputting ${\mu }_{\varphi }$ and a scale parameter $\sigma$,

$${\mathbf{x}}^t={\mu }_{\varphi }\left({\mathbf{x}}^{<t}\right)+\sigma \odot {\mathbf{y}}_0^t\iff {\mathbf{y}}_0^t=\frac{{\mathbf{x}}^t-{\mu }_{\varphi }\left({\mathbf{x}}^{<t}\right)}{\sigma }.$$

(5)

Conditioned on ${\mathbf{y}}_0^t$, this transformation is deterministic. The forward MAF transform ($\mathbf{y}\to \mathbf{x}$) converts the residual into the data sequence; the inverse transform ($\mathbf{x}\to \mathbf{y}$) decorrelates the sequence. The temporally decorrelated, sparse residuals ${\mathbf{y}}_0^{1:T}$ involve a simpler modeling task than generating the frames themselves. Although the scale parameter $\sigma$ can also be conditioned on past frames, we did not find a benefit in practice.

The autoregressive transform in Equation (5) has also been adapted in a VAE model [27] as well as in neural video compression architectures [9,10,11,26]. These approaches separately compress latent variables that govern the next-frame prediction as well as frame residuals, therefore achieving state-of-the-art rate-distortion performance on high-resolution video content. Although these works focused on compression, this paper focuses on generation.

We now specify the second factor in Equation (4), the generative process of the residual variable, as

$$\begin{array}{c}\hfill \begin{array}{cc}\hfill p_{\theta }\left({\mathbf{y}}_{0:N}^t\right|{\mathbf{x}}^{<t})& =p\left({\mathbf{y}}_N^t\right)\prod _{n=1}^N{p}_{\theta }\left({\mathbf{y}}_{n-1}^t\right|{\mathbf{y}}_n^t,{\mathbf{x}}^{<t}).\hfill \end{array}\end{array}$$

(6)

We fix the top-level prior distribution to be a multivariate Gaussian with identity covariance. All other denoising factors are conditioned on past frames and involve prediction networks $M_{\theta }$,

$$\begin{array}{c}\hfill p_{\theta }\left({\mathbf{y}}_{n-1}\right|{\mathbf{y}}_n,{\mathbf{x}}^{<t})=\mathcal{N}\left({\mathbf{y}}_{n-1}\right|M_{\theta }({\mathbf{y}}_n,n,{\mathbf{x}}^{<t}),\gamma \mathbf{I}).\end{array}$$

(7)

As in Equation (2), $\gamma$ is a hyperparameter. Our goal is to learn $\theta$.

Inference Process. Having specified the generative process, we next specify the inference process conditioned on the observed sequence ${\mathbf{x}}^{0:T}$:

$$\begin{array}{c}\hfill q_{\varphi }\left({\mathbf{y}}_{0:N}^t\right|{\mathbf{x}}^{\le t})=q_{\varphi }\left({\mathbf{y}}_0^t\right|{\mathbf{x}}^{\le t})\prod _{n=1}^{N}q\left({\mathbf{y}}_n^t\right|{\mathbf{y}}_{n-1}^t).\end{array}$$

(8)

Since the residual noise is a deterministic function of the observed and predicted frame, the first factor is deterministic and can be expressed as $q_{\varphi }\left({\mathbf{y}}_0^t\right|{\mathbf{x}}^{\le t})=\delta ({\mathbf{y}}_0^t-\frac{{\mathbf{x}}^t-{\mu }_{\varphi }\left({\mathbf{x}}^{<t}\right)}{\sigma })$. The remaining N factors are identical to Equation (1) with ${\mathbf{x}}_n$ being replaced by ${\mathbf{y}}_n$. Following Nichol and Dhariwal [78], we use a cosine schedule to define the variance ${\beta }_n\in (0,1)$. The architecture is shown in Figure 1b.

Equations (7) and (8) generalize and improve the previously proposed TimeGrad [71] method. This approach showed promising performance in forecasting the time series of comparatively smaller dimensions such as electricity prices or taxi trajectories and not video. Besides differences in architecture, this method neither models residuals nor considers the temporal dependency in posterior, which we identify as a crucial aspect to make the model competitive with strong VAE and GAN baselines (see Section 4.6 for an ablation).

Optimization and Sampling. In analogy to time-independent diffusion models, we can derive a variational lower bound that we can optimize using stochastic gradient descent. In analogy to the derivation of Equation (3) [21] and using the same definitions of ${\overline{\alpha }}_n$ and $ϵ$, this results in

$$\begin{array}{cc}\hfill L(\theta ,\varphi )& ={\mathbb{E}}_{\mathbf{x},n,ϵ}\sum _{t=1}^T\left|\right|ϵ-f_{\theta }({\mathbf{y}}_n^t\left(\varphi \right),n,{\mathbf{x}}^{<t}){\left|\right|}^2;\hfill \\ \hfill {\mathbf{y}}_n^t\left(\varphi \right)& =\sqrt{{\overline{\alpha }}_n}{\mathbf{y}}_0^t\left(\varphi \right)+\sqrt{1-{\overline{\alpha }}_n}ϵ;{\mathbf{y}}_0^t\left(\varphi \right)=\frac{{\mathbf{x}}^t-{\mu }_{\varphi }\left({\mathbf{x}}^{<t}\right)}{\sigma }.\hfill \end{array}$$

(9)

We can optimize this function using the reparameterization trick [75], i.e., by randomly sampling $ϵ$ and n and taking stochastic gradients with respect to $\varphi$ and $\theta$. For a practical scheme involving multiple time steps, we also employ teacher forcing [79]. See Algorithms 1 and 2 for the detailed training and sampling procedure, where we abbreviated $f_{\theta ,\varphi }({\mathbf{y}}_n^t,n,{\mathbf{x}}^{<t})\equiv f_{\theta }({\mathbf{y}}_n^t\left(\varphi \right),n,{\mathbf{x}}^{<t})$.

Algorithm 1: Training

Algorithm 2: Generation

4. Experiments

We compare Residual Video Diffusion (RVD) against five strong baselines, including three GAN-based models and two sequential VAEs. We consider four different video datasets and consider both probabilistic (CRPS) and perceptual (FVD, LPIPS) metrics, discussed below. Our model achieves a new state of the art in terms of perceptual quality while being comparable with or better than the best-performing sequential VAE in its frame forecasting ability.

4.1. Datasets

We consider four video datasets of varying complexities and resolutions. Among the simpler datasets of frame dimensions of $64\times 64$, we consider the BAIR Robot Pushing arm dataset [80] and KTH Actions [81]. Among the high-resolution datasets (frame sizes of $128\times 128$), we use CityScape [82], a dataset involving urban street scenes, and a two-dimensional Simulation dataset for turbulent flow of our own making that has been computed using the Lattice Boltzmann Method [83]. These datasets cover various complexities, resolutions, and types of dynamics.

Preprocessing. For KTH and BAIR, we preprocess the videos as commonly proposed [14,27]. For CityScape, we download the portion titled leftImg8bit_sequence_trainvaltest from the official website (https://www.cityscapes-dataset.com/downloads/ (accessed on 1 March 2022)). Each video is a 30-frame sequence from which we randomly select a sub-sequence. All the videos are center-cropped and downsampled to 128 × 128. For the simulation dataset, we use an LBM solver to simulate the flow of a fluid (with pre-specified bulk and shear viscosity and rate of flow) interacting with a static object. We extract 10,000 frames sampled every 128 ticks, using 8000 for training and 2000 for testing.

4.2. Training and Testing Details

The diffusion models are trained with 8 consecutive frames for all the datasets of which the first two frames as used as context frames. We set the batch size to 4 for all high-resolution videos and to 8 for all low-resolution videos. The pixel values of all the video frames are normalized to $[-1,1]$. The models are optimized using the Adam optimizer with an initial learning rate of $5\times {10}^{-5}$, which decays to $2\times {10}^{-5}$. All the models are trained on four NVIDIA RTX Titan GPUs in parallel for around 4–5 days. The number of diffusion depth is fixed to $N=1600$, and the scale term is set to $\sigma =2$. For testing, we use 4 context frames and predict 16 future frames for each video sequence. Wherever applicable, these frames are recursively generated. The model size is about 123 megabytes (32-bit float numbers) for high-resolution models (128 × 128). To sample a 128 × 128 video frame, the model takes 18.5 s per 1000 iterations.

4.3. Baseline Models

SVG-LP [14] is an established sequential VAE baseline. It leverages recurrent architectures in all of the encoder and decoder, and prior to capturing the dynamics in videos. We adopt the official implementation from the authors while replacing all the LSTM with ConvLSTM layers, which helps the model scale to different video resolutions. IVRNN [15] is currently the state-of-the-art video-VAE model trained end-to-end from scratch. The model improves SVG by involving a hierarchy of latent variables. We use the official codebase to train the model. SLAMP [28] is a recent algorithm yielding stochastic predictions similar to ours. It is also similar in spirit to our idea because it incorporates “motion history” to predict the dynamics for future frames. FutureGAN [16] relies on an encoder-decoder GAN model that uses spatio-temporal 3D convolutions to process video tensors. To make the quality of the output more perceptually appealing, the paper employs the concept of progressively growing GANs. We use the official codebase to train the model. Retrospective Cycle GAN [17] employs a single generator that can predict both future and past frames given a context and enforces retrospective cycle constraints. Besides the usual discriminator that can identify fake frames, the method also introduces sequence discriminators to identify sequences containing the said fake frames. We used an available third-party implementation (https://github.com/SaulZhang/Video_Prediction_ZOO/tree/master/RetrospectiveCycleGAN (accessed on 1 March 2022)). DVD-GAN [29] proposes an alternative dual-discriminator architecture for video generation on complex datasets. We also adapt a third-party implementation of the model to conduct our experiment (https://github.com/Harrypotterrrr/DVD-GAN (accessed on 1 March 2022)).

4.4. Evaluation Metrics

We address two key aspects for determining the quality of generated sequences: perceptual quality and the models’ probabilistic forecasting ability. For the former, we adopt FVD [84] and LPIPS [85], while the latter is evaluated using CRPS [86] to assess the marginal (pixel-based) predictions of future frames.

Fréchet Video Distance (FVD) compares sample realism by calculating 2-Wasserstein distance between the ground truth video distribution and the distribution defined by the generative model. Typically, an I3D network pretrained on an action-recognition dataset is used to capture low-dimensional feature representations, the distributions of which are used in the metric. Learned Perceptual Image Patch Similarity (LPIPS), on the other hand, computes the ${\ell }_2$ distance between deep embeddings across all the layers of a pretrained network which are then averaged spatially. The LPIPS score is calculated based on individual frames and then averaged.

Another desirable property of video prediction methods is to forecast future frames reliably. Since ground truth videos exhibit multi-modal conditional distributions (e.g., a traffic light may switch from yellow to green or red), such multi-modality is best captured by proper Scoring Rules such as the Continuous Ranked Probability Score (CRPS). A brief introduction about the metric is available in Appendix A These metrics are frequently used in probabilistic forecasting problems in, e.g., meteorology or finance [87,88].

In a nutshell, CRPS compares a single-sample estimate of the ground truth CDF (a step function) with the model’s CDF for the next frame. The latter can be efficiently estimated in one dimension by repeatedly sampling from the model. In expectation, CRPS not only rewards high accuracy of the mean prediction but also good uncertainty estimates of the model. Although CRPS is not commonly used in evaluating video prediction methods, we argue that it adds a valuable perspective on a model’s uncertainty calibration.

4.5. Qualitative and Quantitative Analysis

Using the perceptual and probabilistic metrics mentioned above, we compare test set predictions of our video diffusion architecture against a wide range of baselines, which model underlying data density both explicitly and implicitly.

Table 1. Test set perceptual (FVD, LPIPS) and forecasting (CRPS) metrics, lower is better (see Section 4 for details). Bold numbers denote the best performance.

	FVD↓				LPIPS↓				CRPS↓
	KTH	BAIR	Sim	City	KTH	BAIR	Sim	City	KTH	BAIR	Sim	City
RVD (ours)	1351	1272	20	997	0.06	0.06	0.01	0.11	6.51	12.86	0.58	9.84
IVRNN	1375	1337	24	1234	0.08	0.07	0.008	0.18	6.17	11.74	0.65	11.00
SLAMP	1451	1749	2998	1853	0.05	0.08	0.30	0.23	6.18	24.8	2.53	23.6
SVG-LP	1783	1631	21	1465	0.12	0.08	0.01	0.20	18.24	13.96	0.75	19.34
RetroGAN	2503	2038	28	1769	0.28	0.07	0.02	0.20	27.49	19.42	1.60	20.13
DVD-GAN	2592	3097	147	2012	0.18	0.10	0.06	0.21	12.05	27.2	1.42	21.61
FutureGAN	4111	3297	319	5692	0.33	0.12	0.16	0.29	37.13	27.97	6.64	29.31

lists all the metric scores for our model and the baselines. Our model performs best in all cases in terms of FVD, a no-reference metric that measures frame quality irrespective of context and without reference to the ground truth. For LPIPS, a reference metric, our model also performs best in 3 out of 4 datasets. The perceptual performance is also verified visually in Figure 2, where RVD shows higher clarity on the generated frames and shows less blurriness in regions that are less predictable due to the fast motion.

We also reported CRPS scores in Table 1. Figure 3 shows 1/CRPS (higher is better) as a function of the frame index, revealing a monotonically decreasing trend along the time axis. This follows our intuition that long-term predictions become worse over time for all models. Our method performs best in 3 out of 4 cases. We can resolve this score also spatially in the images, as we do in Figure 4. Areas of distributional disagreement within a frame are shown in blue (right). See Appendix D for the generated videos on other datasets.

4.6. Ablation Studies

We consider two ablations of our model. The first one studies the impact of applying the diffusion generative model for modeling residuals as opposed to directly predicting the next frames. The second ablation studies the impact of the number of frames that the model sees during training.

Modeling Residuals vs. Dense Frames Our proposed method uses a denoising diffusion generative model to generate residuals to a deterministic next-state prediction (see Figure 1b). A natural question arises whether this architecture is necessary or whether it could be simplified by directly generating the next frame ${\mathbf{x}}_0^t$ instead of the residual ${\mathbf{y}}_0^t$. To address this, we make the following adjustment. Since ${\mathbf{y}}_0^t$ and ${\mathbf{x}}_0^t$ have equal dimensions, the ablation can be realized by setting ${\mu }^t=0$ and $\sigma =1$. To distinguish from our proposed “Residual Video Diffusion” (RVD), we call this ablation “Video Diffusion” (VD). Please note that this ablation can be considered a customized version of TimeGrad [71] applied to video.

Table 2. Ablation studies on (1) modeling residuals (RVD, proposed) versus future frames (VD) and (2) training with different sequence lengths, where $(p+q)$ denotes p context frames and q future frames for prediction. Bold numbers denote the best performance.

	FVD↓				LPIPS↓				CRPS↓
	KTH	BAIR	Sim	City	KTH	BAIR	Sim	City	KTH	BAIR	Sim	City
VD (2 + 6)	1523	1374	37	1321	0.066	0.066	0.014	0.127	6.10	12.75	0.68	13.42
SimpleRVD (2 + 6)	1532	1338	45	1824	0.065	0.063	0.024	0.163	5.80	12.98	0.65	10.78
RVD (2 + 6)	1351	1272	20	997	0.066	0.060	0.011	0.113	6.51	12.86	0.58	9.84
RVD (2 + 3)	1663	1381	33	1074	0.072	0.072	0.018	0.112	6.67	13.93	0.63	10.59
IVRNN (4 + 8)	1375	1337	24	1234	0.082	0.075	0.008	0.178	6.17	11.74	0.65	11.00
IVRNN (4 + 4)	2754	1508	150	3145	0.097	0.074	0.040	0.278	7.25	13.64	1.36	18.24

shows the results. Across all perceptual metrics, the residual model performs better on all data sets. In terms of CRPS, VD performs slightly better on the simpler KTH and BAIR datasets but worse on the more complex Simulation and CityScape data. We, therefore, confirm our earlier claims that modeling residuals over frames is crucial for obtaining better performance, especially on more complex high-resolution video.

Frame Differences vs. Prediction Residuals One may also wonder if similar results could have been obtained by modeling the residual relative to the last frame as opposed to modeling the residual relative to a predicted frame. We called this ablation “SimpleRVD” (shown in Table 2), where we set ${\mu }_t={\widehat{\mathbf{x}}}_{t-1}$. Although the model can run more efficiently with this simplified scheme, Table 2 shows that the video quality is typically negatively affected. We conjecture that, since the predicted next frame may be already motion-compensated, the resulting residual is sparser and, hence, easier to capture by the diffusion model. Similar observations have been made in neural video compression [9,26].

Influence of Training Sequence Length We train both our diffusion model and IVRNN on video sequences of varying lengths. As Table 2 reveals, we find that the diffusion model maintains a robust performance, showing only a small degradation on significantly shorter sequences. In contrast, IVRNN is more sensitive to sequence length. We note that in most experiments, we outperform IVRNN even though we trained our model on shorter sequences. We also note that IVRNN leverages dense-connected hierarchical latent variables to capture long-sequence dependency. Hence, optimizing the high-level latent variables can be challenging when the training sequence is not sufficiently long. Although the diffusion model is also hierarchical, the model only learns a denoising mapping, which is a much simpler scheme than dense-connected latent variables (as shown in Figure 1).

5. Discussion

We proposed “Residual Video Diffusion”: a new model for stochastic video generation based on denoising diffusion probabilistic models. Our approach uses a denoising process, conditioned on the context vector of a convolutional RNN, to generate a residual to a deterministic next-frame prediction. We showed that such residual prediction yields better results than directly predicting the next frame.

To benchmark our approach, we studied a variety of datasets of different degrees of complexity and pixel resolution, including CityScape and a physics simulation dataset of turbulent flow. We compared our approach against two state-of-the-art VAE and three GAN baselines in terms of both perceptual and probabilistic forecasting metrics. Our method leads to a new state of the art in perceptual quality while being competitive with or better than state-of-the-art hierarchical VAE and GAN baselines in terms of probabilistic forecasting. Our results provide several promising directions and could improve world model-based RL approaches as well as neural video codecs.

Limitations The autoregressive setup of the proposed model allows conditional generation with at least one context frame pre-selected from the test dataset. To achieve the unconditional generation of a complete video sequence, we need an auxiliary image generative model to sample the initial context frames. It is also worth mentioning that we only conduct the experiments on single-domain datasets with monotonic contents (e.g., CityScape dataset only contains traffic video recorded by a camera installed in the front of the car), as training a large model for multi-domain datasets like Kinetics [89] is demanding for our limited computing resources. Finally, diffusion probabilistic models tend to be slow in training, which could be accelerated by incorporating DDIM sampling [90] or model distillation [91].

Potential Negative Impacts Just as other generative models, video generation models pose the danger of being misused for generating deepfakes, running the risk of being used for spreading misinformation. Note, however, that a probabilistic video prediction model could also be used for anomaly detection (scoring anomalies by likelihood) and hence may help to detect such forgery.