GLOBER: Coherent Non-autoregressive
Video Generation via GLOBal Guided Video DecodER

VAEs [23, 24, 25, 8] are widely used models that reduce search space for vision generation. For high-resolution image generation, VAEs typically encode each image into a latent feature and then decode it back to the original input image [10, 25, 26, 27]. To reduce the spatial details of videos in a similar way, we employ a pretrained KL-VAE [8] to encode each video frame individually. Specifically, the KL-VAE downsamples each video frame $x_{i}$ by a factor of $f_{frame}$, obtaining a frame latent feature $z_{i}\in R^{H^{\prime}\times W^{\prime}\times C^{\prime}}$, where $H^{\prime}$ and $W^{\prime}$ are $\frac{H}{f_{frame}}$ and $\frac{W}{f_{frame}}$, respectively, and $C^{\prime}$ represents the number of feature channels.

As illustrated in Fig. 2, the video encoder is composed of an embedding feature $e_{v}$, an input layer, a downsample module with a downsample factor of $f_{video}$, a mapping module, and an output module. The embedding feature $e_{v}\in R^{H^{\prime}\times W^{\prime}\times C^{\prime}}$ is randomly initialized and optimized jointly with the entire model. It should be noted that the temporal dimensions of the embedding feature and each frame feature are 1, which are omitted here for brevity. Since content redundancy exists between adjacent video frames [12], we select $K$ keyframes from each input video at equal intervals and encode them individually using KL-VAE. This process produces the corresponding frame features ${z_{q_{1}},z_{q_{2}},...,z_{q_{K}}}$, where $q_{k}$ denotes the index of the $k$-th selected keyframe. Then, we concatenate the embedding feature $e_{v}$ with keyframe features along the temporal dimension, obtaining the input feature $[e_{v}:z_{q_{1}}:...:z_{q_{K}}]\in R^{(K+1)\times H^{\prime}\times W^{\prime}\times C^{\prime}}$ for the video encoder.

The input layer employs a simple 2D spatial convolution with a kernel size of 3 to expand the available channel maps from $C^{\prime}$ to $D$, where $D$ represents the number of new channel maps. After the input layer, the downsample module processes these keyframe features to capture spatial and temporal information, while the mapping module follows. Specifically, the downsample module is composed of a residual layer, a spatial attention layer, a temporal attention layer, and a downsample convolution, while the mapping module follows a similar structure but replaces the downsample convolution with a temporal split operation. Finally, the output module comprises two spatial convolutions, a group normalization layer, and a SiLU activation function. It takes only the embedding part of the transformed feature as input and produces the mean and standard deviation (std) features of the global feature. Then the global feature can be sampled using the following equation:

$$v=v_{mean}+v_{std}*n$$

(1)

where $n$ is sampled from an isotropic Gauss distribution $\mathcal{N}(0,\textbf{I})$, $v_{mean},v_{std}\in R^{H^{\prime\prime}\times W^{\prime\prime}\times C^{\prime\prime}}$ are the mean and std features, $H^{\prime\prime}$ and $W^{\prime\prime}$ are $\frac{H^{\prime}}{f_{video}}$ and $\frac{W^{\prime}}{f_{video}}$ respectively, and $C^{\prime\prime}$ is the number of channels. Considering that we are going to model the global features for generation using a diffusion-based model, which will be specified in Sec. 3.2, an additional kl loss [8] is employed to force the distribution of global features towards an isotropic Gauss distribution:

$$\mathcal{L}_{kl}=\frac{1}{2}(v_{mean}^{2}+v_{std}^{2}-\log v_{std}^{2}-1)$$

(2)

We view the synthesis of video frames as a conditional diffusion-based generation task. As illustrated in Fig. 2, we employ UNet [8] as the backbone of the video decoder, which can be structured into the downsample, mapping, and upsample modules. Each module starts with a residual block and a spatial attention layer, while the downsample and upsample modules additionally include downsample and upsample convolutions respectively. Following [28], each frame feature $z_{i}^{0}$ (i.e. $z_{i}$) is corrupted by $T$ steps during the forward diffusion process using the transition kernel:

	$$\displaystyle q(z_{i}^{t}|z_{i}^{t-1})=\mathcal{N}(z_{i}^{t};\sqrt{1-\beta_{t}}z_{i}^{t-1},\beta_{t}\textbf{I})$$		(3)
	$$\displaystyle q(z_{i}^{t}|z_{i}^{0})=\mathcal{N}(z_{i}^{t};\sqrt{\bar{\alpha}_{t}}z_{i}^{0},(1-\bar{\alpha}_{t})\textbf{I})$$		(4)

where $\{\beta_{t}\in(0,1)\}_{t=1}^{T}$ is a set of hyper-parameters, $\alpha_{t}=1-\beta_{t}$ and $\bar{\alpha}_{t}=\prod_{i=1}^{t}\alpha_{i}$. Based on Eq. (4), we can obtain the corrupted feature $z_{i}^{t}$ directly given the timestep $t$ as follows:

$$z_{i}^{t}=\sqrt{\bar{\alpha}_{t}}z_{i}^{0}+(1-\bar{\alpha}_{t})n_{t}$$

(5)

where $n_{t}$ is a noise feature sampled from an isotropic Gauss distribution $\mathcal{N}(0,\textbf{I})$. The reverse diffusion process $q(z_{i}^{t-1}|z_{i}^{t},z_{i}^{0})$ has a traceable distribution:

$$q(z_{i}^{t-1}|z_{i}^{t},z_{i}^{0})=\mathcal{N}(z_{i}^{t-1}|\tilde{\mu}_{t}(z_{i}^{t},z_{i}^{0}),\tilde{\beta}_{t}\textbf{I})$$

(6)

where $\tilde{\mu}_{t}(z_{i}^{t},z_{i}^{0})=\frac{1}{\sqrt{\alpha_{t}}}(z_{i}^{t}-\frac{\beta_{t}}{\sqrt{1-\bar{\alpha}_{t}}}n_{t})$, $n_{t}\sim\mathcal{N}(0,\textbf{I})$, and $\tilde{\beta}_{t}=\frac{1-\bar{\alpha}_{t-1}}{1-\bar{\alpha}_{t}}\beta_{t}$.

Given that the added noise $n_{t}$ in the $t$-th step is the only unknown term in Eq. (6), we train the UNet to predict $n_{t}$ from $z_{i}^{t}$ condition on the timestep $t$, the global feature $v$, the frame index $i$, and the video description $c$ if exists. To achieve this, a timestep and an index embedding layers, additional modules, and a text encoder are employed to process these condition inputs. In particular, the timestep embedding layer first obtain the sinusoidal position encoding [29] of the diffusion timestep $t$ and then passes it through two linear functions with a SiLU activation interposed between them. Following previous works [28, 30], the timestep embedding is integrated with intermediate features by the residual block in each UNet module. The index embedding layer embeds the frame index in a similar way with the timestep embedding layer. The additional modules (i.e. downsample, mapping, and upsample modules) are utilized to incorporate the index embedding with the global feature and to extract multi-scale representations of the corresponding video frame. These representations are then added to the outputs of UNet modules after zero-convolution. When encoding video descriptions into text embeddings, a pretrained model, namely CLIP [31], is utilized as the text encoder. During attention calculation, these text embeddings are concatenated with flattened frame features to provide cross-modal instructions. Finally, the L2 distance between the predicted noise $n(z_{i}^{t},t,i,v,c)$ and the added noise $n_{t}$ is calculated as the training target:

$$\mathcal{L}_{rec}=\lVert n(z_{i}^{t},t,i,v,c)-n_{t}\rVert_{2}$$

(7)

where $\lVert*\rVert_{2}$ denotes the calculation of the L2 distance. The total training loss is:

$$\mathcal{L}=\mathcal{L}_{rec}+\lambda_{1}\mathcal{L}_{kl}+\lambda_{2}\mathcal{L}_{cra}^{G}$$

(8)

where $\lambda_{1}$ and $\lambda_{2}$ are hyper-parameters and $\mathcal{L}_{cra}^{G}$ is specified in the following paragraph. During generation, the video decoder synthesizes video frames given content features, frame indexes and video descriptions (if exist) by denoising noise features with multiple steps as in [28, 32].

We propose a novel Coherence and Realism Adversarial (CRA) loss to improve global coherence and local realism of synthesized video frames. To reduce computation complexity, we randomly synthesize two video frames with indexes $i$ and $j$, where $0\leq i<j\leq F$, using the video decoder. The synthesized video frame $\bar{x}_{i}$ can be decoded by KL-VAE from the frame feature $\bar{z}_{i}$, which is obtained directly in each training step with the following formulation:

$$\bar{z}_{i}^{0}=\frac{1}{\sqrt{\bar{\alpha}_{t}}}z_{i}^{t}-\frac{(1-\bar{\alpha}_{t})}{\sqrt{\bar{\alpha}_{t}}}n(z_{i}^{t},t,i,v,c)$$

(9)

where $n(z_{i}^{t},t,i,v,c)$ is the predicted noise of $n_{t}$. Then the CRA loss is calculated given the synthesized and real video frames using a video discriminator as depicted in Fig. 2.

In our case, we expect the discriminator to provide supervision on global coherence and local realism of synthesized video frames. To ensure the effectiveness of the CRA loss, we formulate it based on the following two guiding principles: (1) the real samples selected for the discriminator must exhibit the desired correlations, whereas the fake samples must deviate from the expected patterns and are often challenging to distinguish; (2) the real samples selected for the video auto-encoder (i.e. the generator) should serve as the fake samples for the discriminator for adversarial training. In particular, to make the discriminator aware of local realism, we select $\langle x_{i},x_{j}\rangle$ as the real sample and $\langle x_{i},\bar{x}_{j}\rangle$, $\langle\bar{x}_{i},x_{j}\rangle$, $\langle\bar{x}_{i},\bar{x}_{j}\rangle$ as fake samples according to the first principle. This is because the latter samples contain at least one synthesized video frame and violate the target pattern of realism. Furthermore, to ensure that the discriminator is aware of the global coherence, we utilize samples that violate temporal relationships like $\langle x_{j},x_{i}\rangle$ and $\langle\bar{x}_{j},\bar{x}_{i}\rangle$ as fake samples of the discriminator. Then, according to the second principle, {$\langle\bar{x}_{i},\bar{x}_{j}\rangle$, $\langle x_{i},\bar{x}_{j}\rangle$, $\langle\bar{x}_{i},x_{j}\rangle$} are chosen as real samples of the video auto-encoder, since these samples contain at least one synthesized video frames. The CRA loss can be formulated as $\mathcal{L}_{cra}^{D}$ for the discriminator and $\mathcal{L}_{cra}^{G}$ for the generator:

	$\displaystyle\mathcal{L}_{cra}^{D}=$	$\displaystyle log(1-\mathcal{D}(\langle x_{i},x_{j}\rangle))+log\mathcal{D}(\langle x_{i},\bar{x}_{j}\rangle)+log\mathcal{D}(\langle\bar{x}_{i},x_{j}\rangle)$		(10)
		$\displaystyle+log\mathcal{D}(\langle\bar{x}_{i},\bar{x}_{j}\rangle)+log\mathcal{D}(\langle x_{j},x_{i}\rangle)+log\mathcal{D}(\langle\bar{x}_{j},\bar{x}_{i}\rangle)$
	$\displaystyle\mathcal{L}_{cra}^{G}=$	$\displaystyle log(1-\mathcal{D}(\langle\bar{x}_{i},\bar{x}_{j}\rangle))+log(1-\mathcal{D}(\langle\bar{x}_{i},x_{j}\rangle))+log(1-\mathcal{D}(\langle x_{i},\bar{x}_{j}\rangle))$

As illustrated in Fig. 2, the discriminator is built on several spatio-temporal modules that consist of residual blocks and spatio-temporal full attentions. Considering that traditional attention is invariant to the order of input features, two position embeddings $e_{0}$ and $e_{1}$ are employed and added to the previous video frame $x_{i}$ and the subsequent video frame $x_{j}$ respectively with $0\leq i<j\leq F$. These position embeddings are randomly initialized and optimized at the same time with the discriminator.

Table 1 reports the results of our model trained on the Sky Time-lapse, TaiChi-HD, and UCF-101 datasets for 16-frame video generation in both unconditional and conditional settings. As shown in Table 1(a) and Table 1(b), our method achieves comparable performance with prior state-of-the-art models StyleGAN-V [16] and DIGAN [15] on the Sky Time-lapse and TaiChi-HD datasets, respectively. To comprehensively compare the performance on the more challenging UCF-101 dataset, we conduct experiments on both unconditional and conditional video generation at resolutions of $128^{2}$ and $256^{2}$. For unconditional video generation, we use the fixed textual prompt "A video of a person" as the video description, while for conditional video generation, we use the class name of each video as the video description. Our proposed method significantly outperforms previous state-of-the-art models, as shown in Table 1(a) and Table 1(b). The superior performance of our method can be attributed to two factors. Firstly, we explicitly separate the generation of global guidance from the synthesis of frame-wise local characteristics. As global features are 2D features with relatively small spatial resolution, our generative model can model them effectively and efficiently without imposing a high computational burden. Secondly, the synthesis of frame details can refer to the global feature, which provides global guidance such as scene layout, object appearance, and overall behaviours, making it easier for our method to achieve global-coherent and local-realistic video generation.

We quantitatively compare the generation speed among various video generation methods and report the results in Table 2. Our GLOBER method demonstrates remarkable efficiency in generating video frames, thanks to its utilization of the non-autoregression strategy. In contrast, prior methods such as VIDM and VDM, which follow the auto-regression strategy, maintain unchanged memory but require significantly more time for generation. VideoFusion also adopts a non-autoregressive strategy for generating multiple video frames. However, VideoFusion remains slower than our GLOBER, since VideoFusion directly models frame pixels, which brings huge computational burden when generating multiple video frames. TATS employs the interpolation strategy by first generating video keyframes and then interpolating 3 frames between adjacent keyframes. However, it involves autoregressive generationg of video keyframes and the interpolation process is not parallelly implemented, thus being slower than our GLOBER. Notably, GPU memory of VideoFusion and our GLOBER increases with the number of video frames due to the parallel synthesis of all video frames.