A DVERSARIAL V IDEO G ENERATION

ON C OMPLEX DATASETS

Aidan Clark Jeff Donahue

DeepMind DeepMind
London, UK London, UK
[email protected] [email protected]

Karen Simonyan
DeepMind
arXiv:1907.06571v2 [cs.CV] 25 Sep 2019

London, UK
[email protected]

A BSTRACT

Generative models of natural images have progressed towards high fidelity samples
by the strong leveraging of scale. We attempt to carry this success to the field of
video modeling by showing that large Generative Adversarial Networks trained on
the complex Kinetics-600 dataset are able to produce video samples of substantially
higher complexity and fidelity than previous work. Our proposed model, Dual
Video Discriminator GAN (DVD-GAN), scales to longer and higher resolution
videos by leveraging a computationally efficient decomposition of its discriminator.
We evaluate on the related tasks of video synthesis and video prediction, and
achieve new state-of-the-art Fréchet Inception Distance for prediction for Kinetics-
600, as well as state-of-the-art Inception Score for synthesis on the UCF-101
dataset, alongside establishing a strong baseline for synthesis on Kinetics-600.

1 I NTRODUCTION

Time

Figure 1: Selected frames from videos generated by a DVD-GAN trained on Kinetics-600 at 256×256,
128 × 128, and 64 × 64 resolutions (top to bottom).

Modern deep generative models can produce realistic natural images when trained on high-resolution
and diverse datasets (Brock et al., 2019; Karras et al., 2018; Kingma & Dhariwal, 2018; Menick &
Kalchbrenner, 2019; Razavi et al., 2019). Generation of natural video is an obvious further challenge

1
Figure 2: Generated video samples with interesting behavior. In raster-scan order: a) On-screen
generated text with further lines appearing.. b) Zooming in on an object. c) Colored detail from a pen
being left on paper. d) A generated camera change and return.

for generative modeling, but one that is plagued by increased data complexity and computational
requirements. For this reason, much prior work on video generation has revolved around relatively
simple datasets, or tasks where strong temporal conditioning information is available.
We focus on the tasks of video synthesis and video prediction (defined in Section 2.1), and aim
to extend the strong results of generative image models to the video domain. Building upon the
state-of-the-art BigGAN architecture (Brock et al., 2019), we introduce an efficient spatio-temporal
decomposition of the discriminator which allows us to train on Kinetics-600 – a complex dataset of
natural videos an order of magnitude larger than other commonly used datasets. The resulting model,
Dual Video Discriminator GAN (DVD-GAN), is able to generate temporally coherent, high-resolution
videos of relatively high fidelity (Figure 1).
Our contributions are as follows:

• We propose DVD-GAN – a scalable generative model of natural video which produces

high-quality samples at resolutions up to 256 × 256 and lengths up to 48 frames.
• We achieve state of the art for video synthesis on UCF-101 and prediction on Kinetics-600.
• We establish class-conditional video synthesis on Kinetics-600 as a new benchmark for
generative video modeling, and report DVD-GAN results as a strong baseline.

2 BACKGROUND

2.1 V IDEO S YNTHESIS AND P REDICTION

The exact formulation of the video generation task can differ in the type of conditioning signal
provided. At one extreme lies unconditional video synthesis where the task is to generate any video
following the training distribution. Another extreme is occupied by strongly-conditioned models,
including generation conditioned on another video for content transfer (Bansal et al., 2018; Zhou
et al., 2019), per-frame segmentation masks (Wang et al., 2018), or pose information (Walker et al.,
2017; Villegas et al., 2017; Yang et al., 2018). In the middle ground there are tasks which are more
structured than unconditional generation, and yet are more challenging from a modeling perspective
than strongly-conditional generation (which gets a lot of information about the generated video
through its input). The objective of class-conditional video synthesis is to generate a video of a
given category (e.g., “riding a bike”) while future video prediction is concerned with generation of
continuing video given initial frames. These problems differ in several aspects, but share a common
requirement of needing to generate realistic temporal dynamics, and in this work we focus on these
two problems.

2.2 G ENERATIVE A DVERSARIAL N ETWORKS

Generative Adversarial Networks (GANs) (Goodfellow et al., 2014) are a class of generative models
defined by a minimax game between a Discriminator D and a Generator G. The original objective
was proposed by Goodfellow et al. (2014), and many improvements have since been suggested,
mostly targeting improved training stability (Arjovsky et al., 2017; Zhang et al., 2018; Brock et al.,
2019; Gulrajani et al., 2017; Miyato et al., 2018). We use the hinge formulation of the objective (Lim

2
& Ye, 2017; Brock et al., 2019) which is optimized by gradient descent (ρ is the elementwise ReLU
function):

h i h i h i
D: min E ρ(1 − D(x)) + E ρ(1 + D(G(z))) , G: max E D(G(z)) .
D x∼data(x) z∼p(z) G z∼p(z)

GANs have well-known limitations including a tendency towards limited diversity in generated
samples (a phenomenon known as mode collapse) and the difficulty of quantitative evaluation due
to the lack of an explicit likelihood measure over the data. Despite these downsides, GANs have
produced some of the highest fidelity samples across many visual domains (Karras et al., 2018; Brock
et al., 2019).

2.3 K INETICS -600

Kinetics is a large dataset of 10-second high-resolution YouTube clips (Kay et al., 2017; DeepMind,
2018) originally created for the task of human action recognition. We use the second iteration of
the dataset, Kinetics-600 (Carreira et al., 2018), which consists of 600 classes with at least 600
videos per class for a total of around 500,000 videos.1 Kinetics videos are diverse and unconstrained,
which allows us to train large models without being concerned with the overfitting that occurs on
small datasets with fixed objects interacting in specified ways (Ebert et al., 2017; Blank et al., 2005).
Among prior work, the closest dataset (in terms of subject and complexity) which is consistently used
is UCF-101 (Soomro et al., 2012). We focus on Kinetics-600 because of its larger size (almost 50x
more videos than UCF-101) and its increased diversity (600 instead of 101 classes – not to mention
increased intra-class diversity). Nevertheless for comparison with prior art we train on UCF-101
and achieve a state-of-the-art Inception Score there. Kinetics contains many artifacts expected from
YouTube, including cuts (as in Figure 2d), title screens and visual effects. Except when specifically
described, we choose frames with stride 2 (meaning we skip every other frame). This allows us to
generate videos with more complexity without incurring higher computational cost.
To the best of our knowledge we are the first to consider generative modelling of the entirety of
the Kinetics video dataset2 , although a small subset of Kinetics consisting of 4,000 selected and
stabilized videos (via a SIFT + RANSAC procedure) has been used in at least two prior papers (Li
et al., 2018; Balaji et al., 2018). Due to the heavy pre-processing and stabilization present, as well
as the sizable reduction in dataset size (two orders of magnitude) we do not consider these datasets
comparable to the full Kinetics-600 dataset.

2.4 E VALUATION M ETRICS

Designing metrics for measuring the quality of generative models (GANs in particular) is an active
area of research (Sajjadi et al., 2018; Barratt & Sharma, 2018). In this work we report the two most
commonly used metrics, Inception Score (IS) (Salimans et al. (2016)) and Fréchet Inception Distance
(FID) (Heusel et al., 2017). The standard instantiation of these metrics is intended for generative
image models, and uses an Inception model (Szegedy et al., 2016) for image classification or feature
extraction. For videos, we use the publicly available Inflated 3D Convnet (I3D) network trained
on Kinetics-600 (Carreira & Zisserman, 2017). Our Fréchet Inception Distance is therefore very
similar to the Fréchet Video Distance (FVD) (Unterthiner et al., 2018), although our implementation
is different and more aligned with the original FID metric.3

3 D UAL V IDEO D ISCRIMINATOR GAN

Our primary contribution is Dual Video Discriminator GAN (DVD-GAN), a generative video model
of complex human actions built upon the state-of-the-art BigGAN architecture (Brock et al., 2019)
while introducing scalable, video-specific generator and discriminator architectures. An overview of
1
Kinetics is occasionally pruned and so we cannot give an exact size.
2
In parallel with the concurrent work of Weissenborn et al. (2019).
3
We use ‘avgpool’ features (rather than logits) by default, our I3D model is trained on Kinetics-600 (rather
than Kinetics-400), and we pre-calculate ground-truth statistics on the entire training set.

3
 Generator Discriminators
Frame T Per Frame
G
ResNet 3 D
Block x2 ResNet ℒS
Block
x Depth Times
G
1 ResNet 3
Block x2

D
G ResNet ℒT
ResNet 3 Block
Frame 0 Block x2
1 x Depth Times

x Depth Times

Class Conditional Batch Norm Feature Concatenation Standard Gaussian Noise

ReLU Activation k Convolution with Kernel Size k Convolutional Gated Recurrent Unit

Bilinear x2 Upsampling Average Pool x2 Downsampling One-Hot Class Indicator

G ResNet Block D ResNet Block

Figure 3: Simplified architecture
1
diagram of G (left) and DS /DT (right).1 More details in B.2.

3 3 3 3
the DVD-GAN architecture is given in Figure 3 and a detailed description is in Appendix B.2. Unlike
some of the prior work, our generator contains no explicit priors for foreground, background or
motion (optical flow); instead, we rely on a high-capacity neural network to learn this in a data-driven
manner. While DVD-GAN contains sequential components (RNNs), it is not autoregressive in time
or in space. In other words, the pixels of each frame do not directly depend on other pixels in the
video, as would be the case for auto-regressive models or models generating one frame at a time.
Generating long and high resolution videos is a heavy computational challenge: individual samples
from Kinetics-600 (just 10 seconds long) contain upwards of 16 million pixels which need to be
generated in a consistent fashion. This is a particular challenge to the discriminator. For example, a
generated video might contain an object which leaves the field of view and incorrectly returns with a
different color. Here, the ability to determine this video is generated is only possible by comparing
two different spatial locations across two (potentially distant) frames. Given a video with length
T , height H, and width W , discriminators that process the entire video would have to process all
H × W × T pixels – limiting the size of the model and the size of the videos being generated.

3.1 D UAL D ISCRIMINATORS

DVD-GAN tackles this scale problem by using two discriminators: a Spatial Discriminator DS and a
Temporal Discriminator DT . DS critiques single frame content and structure by randomly sampling
k full-resolution frames and judging them individually. We use k = 8 and discuss this choice in
Section 4.3. DS ’s final score is the sum of the per-frame scores. The temporal discriminator DT must
provide G with the learning signal to generate movement (something not evaluated by DS ). To make
the model scalable, we apply a spatial downsampling function φ(·) to the whole video and feed its
output to DT . We choose φ to be 2 × 2 average pooling, and discuss alternatives in Section 4.3. This
results in an architecture where the discriminators do not process the entire video’s worth of pixels,
since DS processes only k × H × W pixels and DT only T × H W
2 × 2 . For a 48 frame video at
128 × 128 resolution, this reduces the number of pixels to process per video from 786432 to 327680:
a 58% reduction. Despite this decomposition, the discriminator objective is still able to penalize
almost all inconsistencies which would be penalized by a discriminator judging the entire video. DT
judges any temporal discrepancies across the entire length of the video, and DS can judge any high
resolution details. The only detail the DVD-GAN discriminator objective is unable to reflect is the
temporal evolution of pixels within a 2 × 2 window. We have however not noticed this affecting
the generated samples in practice. DVD-GAN’s DS is similar to the per-frame discriminator DI in
MoCoGAN (Tulyakov et al., 2018). However MoCoGAN’s analog of DT looks at full resolution
videos, whereas DS is the only source of learning signal for high-resolution details in DVD-GAN.
For this reason, DS is essential when φ is not the identity, unlike in MoCoGAN where the additional
per-frame discriminator is less crucial.

4
Figure 4: Each row is the first frame of 15 videos from a random class, all from the same check-
point. The classes are: cooking scallops, changing wheel (not on bike), calculating, dribbling
basketball.

3.2 R ELATED W ORK

Generative video modeling is a widely explored problem which includes work on VAEs (Babaeizadeh
et al., 2018; Denton & Fergus, 2018; Lee et al., 2018; Hsieh et al., 2018), auto-regressive mod-
els (Ranzato et al., 2014; Srivastava et al., 2015; Kalchbrenner et al., 2017; Weissenborn et al., 2019),
normalizing flows (Kumar et al., 2019), and GANs (Mathieu et al., 2015; Vondrick et al., 2016; Saito
et al., 2017; Saito & Saito, 2018). Much prior work considers decompositions which model the
texture and spatial consistency of objects separately from their temporal dynamics. One approach is
to split G into foreground and background models (Vondrick et al., 2016; Spampinato et al., 2018),
while another considers explicit or implicit optical flow in either G or D (Saito et al., 2017; Ohnishi
et al., 2018). Similar to DVD-GAN, MoCoGAN (Tulyakov et al., 2018) discriminates individual
frames in addition to a discriminator which operates on fixed-length K-frame slices of the whole
video (where K < T ). Though this potentially reduces the number of pixels to discriminate to
(H × W ) + (K × H × W ), Tulyakov et al. (2018) describes discriminating sliding windows, which
increases the total number of pixels. Other models follow this approach by discriminating groups of
frames (Xie et al., 2018; Sun et al., 2018; Balaji et al., 2018).
TGANv2 (Saito & Saito, 2018) proposes “adaptive batch reduction” for efficient training, an operation
which randomly samples subsets of videos within a batch and temporal subwindows within each
video. This operation is applied throughout TGANv2’s G, with heads projecting intermediate feature
maps directly to pixel space before applying batch reduction, and corresponding discriminators
evaluating these lower resolution intermediate outputs. An effect of this choice is that TGANv2
discriminators only evaluate full-length videos at very low resolution. We show in Figure 6 that a
similar reduction in DVD-GAN’s resolution when judging full videos leads to a loss in performance.
We expect further reduction (towards the resolution at which TGANv2 evaluates the entire length of
video) to lead to further degradation of DVD-GAN’s quality. Furthermore, this method is not easily
adapted towards models with large batch sizes divided across a number of accelerators, with only a
small batch size per replica.

4 E XPERIMENTS AND A NALYSIS

A detailed description of our training setup is in Appendix B.3. Each DVD-GAN was trained on
TPU pods (Google, 2018) using between 32 and 512 replicas with an Adam (Kingma & Ba, 2014)
optimizer. Video Synthesis models are trained for around 300,000 learning steps, whilst Video
Prediction models are trained for up to 1,000,000 steps. Most models took between 12 and 96 hours
to train.

4.1 C LASS -C ONDITIONAL V IDEO S YNTHESIS

Our primary results concern the problem of Class-Conditional Video Synthesis. We provide our
results for the UCF-101 and Kinetics-600 datasets. With Kinetics-600 emerging as a new benchmark
for generative video modelling, our results establish a strong baseline for future work.

5
Figure 5: All 48 frames (in raster-scan order) from a 64 × 64 sample from watermelon cutting class.

4.1.1 K INETICS -600 R ESULTS

Table 1: FID/IS for DVD-GAN on Kinetics-600 Video Synthesis. We

present the scores of the model taken at the point in training when
the best FID was attained. The "No Truncation" columns contain the
scores obtained without the truncation trick. The "With Truncation"
columns contain the scores obtained at the truncation level which
results in the best Inception Score.

(# Frames / Resolution) No Truncation With Truncation

FID (↓) IS (↑) FID (↓) IS (↑)
12/64 × 64 0.85 53.81 7.13 187.23
12/128 × 128 1.16 77.45 13.04 246.18
12/256 × 256 2.05 62.78 10.17 162.44
48/64 × 64 13.75 104.09 47.86 264.12
48/128 × 128 28.44 81.41 45.79 188.32

In Table 1 we show the main result of this paper: benchmarks for Video Synthesis on Kinetics-600.
We consider a range of resolutions and video lengths, and measure Inception Score and Fréchet
Inception Distance (FID) for each (as described in Section 2.4). We further measure each model
along a truncation curve, which we carry out by calculating FID and IS statistics while varying
the standard deviation of the latent vectors between 0 and 1. There is no prior work with which to
quantitatively compare these results (for comparative experiments see Section 4.1.2 and Section 4.2.1),
but we believe these samples to show a level of fidelity not yet achieved in datasets as complex
as Kinetics-600 (see samples from each row in Appendix E.1). Because all videos are resized for
the I3D network (to 224 × 224), it is meaningful to compare metrics across equal length videos
at different resolutions. Neither IS nor FID are comparable across videos of different lengths, and
should be treated as separate metrics.
Generating longer and larger videos is a more challenging modeling problem, which is conveyed
by the metrics (in particular, comparing 12-frame videos across 64 × 64, 128 × 128 and 256 × 256
resolutions). Nevertheless, DVD-GAN is able to generate plausible videos at all resolutions and with
length spanning up to 4 seconds (48 frames). As can be seen in Appendix E.1, smaller videos display
high quality textures, object composition and movement. At higher resolutions, generating coherent
objects becomes more difficult (movement consists of a much larger number of pixels), but high-level
details of the generated scenes are still extremely coherent, and textures (even complicated ones like
a forest backdrop in Figure 1a) are generated well. It is further worth noting that the 48-frame models
do not see more high resolution frames than the 12-frame model (due to the fixed choice of k = 8
described in Section 3.1), yet nevertheless learn to generate high resolution images.

4.1.2 V IDEO S YNTHESIS ON UCF-101

We further verify our results by testing the same model on UCF-101 (Soomro et al., 2012), a smaller
dataset of 13,320 videos of human actions across 101 classes that has previously been used for video
synthesis and prediction (Saito et al., 2017; Saito & Saito, 2018; Tulyakov et al., 2018). Our model
produces samples with an IS of 32.97, significantly outperforming the state of the art (see Table 2 for
quantitative comparison and Appendix C.1 for more details).

6
 Table 2: IS on UCF-101 with C3D. Table 3: FVD on BAIR.

Method IS (↑) Method FVD (↓)

VGAN (Vondrick et al., 2016) 8.31 ± .09 SVP-FP 315.5
TGAN (Saito et al., 2017) 11.85 ± .07 CDNA 296.5
MoCoGAN (Tulyakov et al., 2018) 12.42 ± .03 SV2P 262.5
ProgressiveVGAN (Acharya et al., 2018) 14.56 ± .05 SAVP 116.4
TGANv2 (Saito & Saito, 2018) 24.34 ± .35 DVD-GAN-FP (ours) 109.8
DVD-GAN (ours) 32.97 ± 1.7 Video Transformer 94 ± 2

Table 4: DVD-GAN-FP’s FVD scores on Video Prediction for 16 frames of Kinetics-600

without frame skipping. The final row represents a Video Synthesis model generating 16 frames.

Method Training Set FVD (↓) Test Set FVD (↓)

Video Transformer (Weissenborn et al., 2019) - 170 ± 5
DVD-GAN-FP 68.66 ± 0.78 69.15 ± 1.16
DVD-GAN 32.3 ± 0.82 31.1 ± 0.56

4.2 F UTURE V IDEO P REDICTION

Future Video Prediction is the problem of generating a sequence of frames which directly follow
from one (or a number) of initial conditioning frames. Both this and video synthesis require G to
learn to produce realistic scenes and temporal dynamics, however video prediction further requires G
to analyze the conditioning frames and discover elements in the scene which will evolve over time. In
this section, we use the Fréchet Video Distance exactly as Unterthiner et al. (2018): using the logits
of an I3D network trained on Kinetics-400 as features. This allows for direct comparison to prior
work. Our model, DVD-GAN-FP (Frame Prediction), is slightly modified to facilitate the changed
problem, and details of these changes are given in Appendix B.4.

4.2.1 F RAME -C ONDITIONAL K INETICS

For direct comparison with concurrent work on autoregressive video models (Weissenborn et al.,
2019) we consider the generation of 11 frames of Kinetics-600 at 64 × 64 resolution conditioned on
5 frames, where the videos for training are not taken with any frame skipping. We show results for all
these cases in Table 4. Our frame-conditional model DVD-GAN-FP outperforms the prior work on
frame-conditional prediction for Kinetics. The final row labeled DVD-GAN corresponds to 16-frame
class-conditional Video Synthesis samples, generated without frame conditioning and without frame
skipping. The FVD of this video synthesis model is notably better.
On the one hand, we hypothesize that the synthesis model has an easier generative task: it can
choose to generate (relatively) simple samples for each class, rather than be forced to continue frames
taken from videos which are class outliers, or contain more complicated details. On the other hand,
a certain portion of the FID/FVD metric undoubtedly comes from the distribution of objects and
backgrounds present in the dataset, and so it seems that the prediction model should have a handicap
in the metric by being given the ground truth distribution of backgrounds and objects with which to
continue videos. The synthesis model’s improved performance on this task seems to indicate that the
advantage of being able to select videos to generate is greater than the advantage of having a ground
truth distribution of starting frames.

4.2.2 BAIR ROBOT P USHING

We further test future video prediction on the single-class BAIR Robot Pushing Dataset (Ebert et al.,
2017), a dataset of stationary videos of a robot arm moving around a set of changing objects. In
order for direct comparison with previous results reported in Unterthiner et al. (2018), we consider
generating 15 frames conditioned on a single starting frame. Like on prediction with Kinetics, we
report FVD exactly as in Unterthiner et al. (2018), with ground truth statistics and conditioning

7
Figure 6: The effect of φ in DT (left two) and k in DS (right two). FID is similar for any choice of φ,
while IS declines as downsampling increases. Increasing k improves both with diminishing returns.

frames taken from the 256-video dev set. Results are reported in Table 3. Scores are taken from
Unterthiner et al. (2018). DVD-GAN-FP outperforms all prior adversarial models trained on this
dataset, but performs slightly worse than Video Transformer, a concurrently developed autoregressive
model Weissenborn et al. (2019).

4.3 D UAL D ISCRIMINATOR I NPUT

We analyze several choices for k (the number of frames per sample in the input to DS ) and φ (the
downsampling function for DT ). We expect setting φ to the identity or k = T to result in the best
model, but we are interested in the maximally compressive k and φ that reduce discriminator input
size (and the amount of computation), while still producing a high quality generator. For φ, we
consider: 2 × 2 and 4 × 4 average pooling, the identity (no downsampling), as well as a φ which
takes a random half-sized crop of the input video (as in Saito & Saito (2018)). Results can be seen in
Figure 6. For each ablation, we train three identical DVD-GANs with different random initializations
on 12-frame clips of Kinetics-600 at 64 × 64 resolution for 100,000 steps. We report mean and
standard deviation (via the error bars) across each group for the whole training period. For k, we
consider 1, 2, 8 and 10 frames. We see diminishing effect as k increases, so settle on k = 8. We note
the substantially reduced IS of 4 × 4 downsampling as opposed to 2 × 2, and further note that taking
half-sized crops (which results in the same number of pixels input to DT as 2 × 2 pooling) is also
notably worse.

5 C ONCLUSION
We approached the challenging problem of modeling natural video by introducing a GAN capable of
capturing the complexity of a large video dataset. We showed that on UCF-101 and frame-conditional
Kinetics-600 it quantitatively achieves the new state of the art, alongside qualitatively producing
sample videos with high complexity and diversity. We further wish to emphasize the benefit of
training generative models on large and complex video datasets, such as Kinetics-600, and envisage
the strong baselines we established on this dataset with DVD-GAN will be used as a reference point
by the generative modeling community moving forward. While much remains to be done before
realistic videos can be consistently generated in an unconstrained setting, we believe DVD-GAN is a
step in that direction.

ACKNOWLEDGMENTS
We would like to thank Eric Noland and João Carreira for help with the Kinetics dataset and Marcin
Michalski and Karol Kurach for helping us acquire data and models for Fréchet Video Distance
comparison. We would further like to thank Sander Dieleman, Jacob Walker and Tim Harley for
useful discussions and feedback on the paper.

R EFERENCES
Dinesh Acharya, Zhiwu Huang, Danda Pani Paudel, and Luc Van Gool. Towards high resolution
video generation with progressive growing of sliced Wasserstein GANs. arXiv:1810.02419, 2018.
Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein GAN. arXiv:1701.07875, 2017.

8
Mohammad Babaeizadeh, Chelsea Finn, Dumitru Erhan, Roy H Campbell, and Sergey Levine.
Stochastic variational video prediction. ICLR, 2018.

Yogesh Balaji, Martin Renqiang Min, Bing Bai, Rama Chellappa, and Hans Peter Graf. Tfgan:
Improving conditioning for text-to-video synthesis. 2018.

Nicolas Ballas, Li Yao, Chris Pal, and Aaron Courville. Delving deeper into convolutional networks
for learning video representations. arXiv:1511.06432, 2015.

Aayush Bansal, Shugao Ma, Deva Ramanan, and Yaser Sheikh. Recycle-GAN: Unsupervised video
retargeting. In ECCV, September 2018.

Shane Barratt and Rishi Sharma. A note on the Inception Score. arXiv:1801.01973, 2018.

Moshe Blank, Lena Gorelick, Eli Shechtman, Michal Irani, and Ronen Basri. Actions as space-time
shapes. In ICCV, 2005.

Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity
natural image synthesis. ICLR, 2019.

Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the Kinetics
dataset. In ICCV, 2017.

Joao Carreira, Eric Noland, Andras Banki-Horvath, Chloe Hillier, and Andrew Zisserman. A short
note about Kinetics-600. arXiv:1808.01340, 2018.

Harm De Vries, Florian Strub, Jérémie Mary, Hugo Larochelle, Olivier Pietquin, and Aaron C
Courville. Modulating early visual processing by language. In NeurIPS, 2017.

DeepMind, 2018. Kinetics. https://deepmind.com/research/open-source/

open-source-datasets/kinetics/. Accessed: 2019.

Emily Denton and Rob Fergus. Stochastic video generation with a learned prior. ICML, 2018.

Vincent Dumoulin, Jonathon Shlens, and Manjunath Kudlur. A learned representation for artistic
style. In ICLR, 2017.

Frederik Ebert, Chelsea Finn, Alex X Lee, and Sergey Levine. Self-supervised visual planning with
temporal skip connections. arXiv:1710.05268, 2017.

Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair,
Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014.

Google, 2018. Cloud tpu. https://cloud.google.com/tpu/. Accessed: 2019.

Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville.
Improved training of Wasserstein GANs. In NeurIPS, 2017.

Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter.
GANs trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS,
2017.

Jun-Ting Hsieh, Bingbin Liu, De-An Huang, Li Fei-Fei, and Juan Carlos Niebles. Learning to
decompose and disentangle representations for video prediction. In NeurIPS. 2018.

Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by
reducing internal covariate shift. arXiv:1502.03167, 2015.

Nal Kalchbrenner, Aäron van den Oord, Karen Simonyan, Ivo Danihelka, Oriol Vinyals, Alex Graves,
and Koray Kavukcuoglu. Video pixel networks. In ICML, 2017.

Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative
adversarial networks. arXiv:1812.04948, 2018.

9
Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan,
Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman.
The Kinetics human action video dataset. arXiv:1705.06950, 2017.
Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv:1412.6980,
2014.
Durk P Kingma and Prafulla Dhariwal. Glow: Generative flow with invertible 1x1 convolutions. In
NeurIPS, 2018.
Manoj Kumar, Mohammad Babaeizadeh, Dumitru Erhan, Chelsea Finn, Sergey Levine, Laurent
Dinh, and Durk Kingma. VideoFlow: A flow-based generative model for video. arXiv:1903.01434,
2019.
Alex X Lee, Richard Zhang, Frederik Ebert, Pieter Abbeel, Chelsea Finn, and Sergey Levine.
Stochastic adversarial video prediction. arXiv:1804.01523, 2018.
Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv:1607.06450,
2016.
Yitong Li, Martin Renqiang Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video
generation from text. In Thirty-Second AAAI Conference on Artificial Intelligence, 2018.
Jae Hyun Lim and Jong Chul Ye. Geometric GAN. arXiv:1705.02894, 2017.
Michael Mathieu, Camille Couprie, and Yann LeCun. Deep multi-scale video prediction beyond
mean square error. arXiv:1511.05440, 2015.
Jacob Menick and Nal Kalchbrenner. Generating high fidelity images with subscale pixel networks
and multidimensional upscaling. In ICLR, 2019.
Takeru Miyato and Masanori Koyama. cGANs with projection discriminator. arXiv:1802.05637,
2018.
Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for
generative adversarial networks. arXiv:1802.05957, 2018.
Katsunori Ohnishi, Shohei Yamamoto, Yoshitaka Ushiku, and Tatsuya Harada. Hierarchical video
generation from orthogonal information: Optical flow and texture. In AAAI, 2018.
Marc’Aurelio Ranzato, Arthur Szlam, Joan Bruna, Michaël Mathieu, Ronan Collobert, and Sumit
Chopra. Video (language) modeling: a baseline for generative models of natural videos.
arXiv:1412.6604, 2014.
Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with
VQ-VAE-2. arXiv preprint arXiv:1906.00446, 2019.
Masaki Saito and Shunta Saito. TGANv2: Efficient training of large models for video generation
with multiple subsampling layers. arXiv:1811.09245, 2018.
Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular
value clipping. In ICCV, 2017.
Mehdi SM Sajjadi, Olivier Bachem, Mario Lucic, Olivier Bousquet, and Sylvain Gelly. Assessing
generative models via precision and recall. In NeurIPS, 2018.
Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen.
Improved techniques for training GANs. In NeurIPS, 2016.
Andrew M Saxe, James L McClelland, and Surya Ganguli. Exact solutions to the nonlinear dynamics
of learning in deep linear neural networks. arXiv:1312.6120, 2013.
Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions
classes from videos in the wild. arXiv:1212.0402, 2012.

10
Concetto Spampinato, Sergio Palazzo, P D’Oro, Francesca Murabito, Daniela Giordano, and M Shah.
VOS-GAN: Adversarial learning of visual-temporal dynamics for unsupervised dense prediction
in videos. arXiv:1803.09092, 2018.
Nitish Srivastava, Elman Mansimov, and Ruslan Salakhutdinov. Unsupervised learning of video
representations using LSTMs. In ICML, 2015.
Ximeng Sun, Huijuan Xu, and Kate Saenko. A two-stream variational adversarial network for video
generation. arXiv:1812.01037, 2018.
Ilya Sutskever, James Martens, and Geoffrey E Hinton. Generating text with recurrent neural networks.
In ICML, 2011.
Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Re-
thinking the Inception architecture for computer vision. In CVPR, 2016.
Seiya Tokui, Kenta Oono, Shohei Hido, and Justin Clayton. Chainer: a next-generation open source
framework for deep learning. In Workshop on Systems for ML and Open Source Software at
NeurIPS, 2015.
Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spa-
tiotemporal features with 3D convolutional networks. In ICCV, 2015.
Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. MoCoGAN: Decomposing motion
and content for video generation. In CVPR, 2018.
Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing
ingredient for fast stylization. arXiv:1607.08022, 2016.
Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski,
and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges.
arXiv:1812.01717, 2018.
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz
Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017.
Ruben Villegas, Jimei Yang, Yuliang Zou, Sungryull Sohn, Xunyu Lin, and Honglak Lee. Learning
to generate long-term future via hierarchical prediction. In ICML, 2017.
Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. In
NeurIPS, 2016.
Jacob Walker, Kenneth Marino, Abhinav Gupta, and Martial Hebert. The pose knows: Video
forecasting by generating pose futures. In ICCV, 2017.
Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan
Catanzaro. Video-to-video synthesis. arXiv:1808.06601, 2018.
Dirk Weissenborn, Oscar Täckström, and Jakob Uszkoreit. Scaling autoregressive video models.
arXiv:1906.02634, 2019.
Yuxin Wu and Kaiming He. Group normalization. In ECCV, 2018.
You Xie, Erik Franz, Mengyu Chu, and Nils Thuerey. tempoGAN: A temporally coherent, volumetric
GAN for super-resolution fluid flow. TOG, 2018.
Ceyuan Yang, Zhe Wang, Xinge Zhu, Chen Huang, Jianping Shi, and Dahua Lin. Pose guided human
video generation. In ECCV, 2018.
Han Zhang, Ian Goodfellow, Dimitris Metaxas, and Augustus Odena. Self-attention generative
adversarial networks. arXiv:1805.08318, 2018.
Yipin Zhou, Zhaowen Wang, Chen Fang, Trung Bui, and Tamara L Berg. Dance dance generation:
Motion transfer for internet videos. arXiv:1904.00129, 2019.

11
A S EPARABLE ATTENTION
A previous version of this draft contained Separable Attention, a module which allows self-attention
(Vaswani et al., 2017) to be applied to spatio-temporal features which are too large for the quadratic
memory cost of vanilla self-attention. Though DVD-GAN as proposed in this draft does not contain
this module, we include a formal definition to aid understanding.
Self-attention on a single batch element X of shape [N, C] (where N is the number of spatial
positions and C is the number of features per location) can be given as:
Self Attention(X) = sof tmax XQ(XK)T XV
 

where Q, K, V are parameters all of shape [C, C] and the softmax is taken over the final axis. Batched
self-attention is identical, except X has a leading batch axis and matrix multiplications are batched
(i.e. the XQ multiplies two tensors of shape [B, N, C] and [C, C] and results in shape [B, N, C]).
Separable Attention recognizes the natural decomposition of N = H × W × T by attending over each
axis separately and in order. That is, first each feature is replaced with the result of a self-attention
pass which only considers other features at the same H/W location (but across different frames),
then the result of that layer (which contains cross-temporal information) is processed by a second
self-attention layer which attends to features at different heights (but at the same width-point, and
at the same frame), and then finally one which attends over width. The Python pseudocode4 below
implements this module assuming that X is given with the interior axes already separated (i.e., X is
of shape [B, H, W, T, C]).
def self_attention(x, q, k, v):
xq, xk, xv = np.matmul(x, q), np.matmul(x, k), np.matmul(x, v)
qv_correlations = np.matmul(xq, np.transpose(xk))
return np.matmul(np.softmax(qv_correlations, axis=−1), xv)

def separable_attention(x, q1, k1, v1, q2, k2, v2, q3, k3, v3):
b, h, w, t, c = x.shape
# Apply attention over time.
x = np.reshape(x, [b*h*w, t, c])
x = self_attention(x, q1, k1, v1)
# Apply attention over height.
x = np.reshape(x, [b*w*t, h, c])
x = self_attention(x, q2, k2, v2)
# Apply attention over width.
x = np.reshape(x, [b*h*t, w, c])
x = self_attention(x, q3, k3, v3)
return x

2
Separable
 Attention crucially reduces the
2 2 2
 asymptotic memory cost from O((HW T ) ) to
max O(H W T ), O(HW T ), O(HW T ) while still allowing the result of the module to con-
tain features at each location accumulated from all other features at any spatio-temporal location.

B E XPERIMENT M ETHODOLOGY
B.1 DATASET P ROCESSING

For all datasets we randomly shuffle the training set for each model replica independently. Experi-
ments on the BAIR Robot Pushing dataset are conducted in the native resolution of 64 × 64, where
for UCF-101 we operate at a (downsampled) 128 × 128 resolution. This is done by a bilinear resize
such that the video’s smallest dimension is mapped to 128 pixels (maintaining aspect ratio). From this
we take a random 128-pixel crop along the other dimension. We use the same procedure to construct
datasets of different resolutions for Kinetics-600. All three datasets contain videos with more frames
than we generate, so we take a random sequence of consecutive frames from the resized output.
4
To be completely correct, a similar function operating on numpy arrays must properly transpose the axes
before each reshape to ensure data is formatted in the proper order after the reshape operation.

12
 Frame 0 Block
1
x2 x Depth Times

x Depth Times

Class Conditional Batch Norm Feature Concatenation Standard Gaussian Noise

ReLU Activation k Convolution with Kernel Size k Convolutional Gated Recurrent Unit

Bilinear x2 Upsampling Average Pool x2 Downsampling One-Hot Class Indicator

G ResNet Block D ResNet Block

1 1

3 3 3 3

Figure 7: The residual blocks for G and DS /DT . See Figure 3 for the icons and B.2 for more detail.

B.2 A RCHITECTURE D ESCRIPTION

Our model adopts many architectural choices from Brock et al. (2019) including our nomenclature
for describing network width, which is determined by the product of a channel multiplier ch with a
constant for each layer in the network. The layer-wise constants for G are [8, 8, 8, 4, 2] for 64 × 64
videos and [8, 8, 8, 4, 2, 1] for 128 × 128. The width of the i-th layer is given by the product of ch
and the i-th constant and all layers prior to the residual network in G use the initial layer’s multiplier
and we refer to the product of that and ch as ch0 . ch in DVD-GAN is 128 for videos with 64 × 64
resolution and 96 otherwise. The corresponding ch lists for both DT and DS are [2, 4, 8, 16, 16] for
64 × 64 resolution and [1, 2, 4, 8, 16, 16] for 128 × 128.
The input to G consists of a Gaussian latent noise z ∼ N (0, I) and a learned linear embedding
e(y) of the desired class y. Both inputs are 120-dimensional vectors. G starts by computing an
affine transformation of [z; e(y)] to a [4, 4, ch0 ]-shaped tensor (in Figure 3 this is represented as a
1 × 1 convolution). [z; e(y)] is used as the input to all class-conditional Batch Normalization layers
throughout G (the gray line in Figure 7).
This is then treated as the input (at each frame we would like to generate) to a Convolutional Gated
Recurrent Unit (Ballas et al., 2015; Sutskever et al., 2011) whose update rule for input xt and previous
output ht−1 is given by the following:

r = σ(Wr ?3 [ht−1 ; xt ] + br )
u = σ(Wu ?3 [ht−1 ; xt ] + bu )
c = ρ(Wc ?3 [xt ; r ht−1 ] + bc )
ht = u ht−1 + (1 − u) c

In these equations σ and ρ are the elementwise sigmoid and ReLU functions respectively, the ?n
operator represents a convolution with a kernel of size n × n, and the operator is an elementwise
multiplication. Brackets are used to represent a feature concatenation. This RNN is unrolled once per
frame. The output of this RNN is processed by two residual blocks (whose architecture is given by
Figure 7). The time dimension is combined with the batch dimension here, so each frame proceeds
through the blocks independently. The output of these blocks has width and height dimensions which
are doubled (we skip upsampling in the first block). This is repeated a number of times, with the
output of one RNN + residual group fed as the input to the next group, until the output tensors have
the desired spatial dimensions. We do not reduce over the time dimension when calculating Batch
Normalization statistics. This prevents the network from utilizing the Batch Normalization layers to
pass information between timesteps.
The spatial discriminator DS functions almost identically to BigGAN’s discriminator, though an
overview of the residual blocks is given in Figure 7 for completeness. A score is calculated for
each of the uniformly sampled k frames (we default to k = 8) and the DS output is the sum over
per-frame scores. The temporal discriminator DT has a similar architecture, but pre-processes the
real or generated video with a 2 × 2 average-pooling downsampling function φ. Furthermore, the
first two residual blocks of DT are 3-D, where every convolution is replaced with a 3-D convolution
with a kernel size of 3 × 3 × 3. The rest of the architecture follows BigGAN (Brock et al., 2019).

B.3 T RAINING D ETAILS

Sampling from DVD-GAN is very efficient, as the core of the generator architecture is a feed-forward
convolutional network: two 64 × 64 48-frame videos can be sampled in less than 150ms on a single

13
 Generator Discriminators
Frame T Per Frame
G
ResNet 3 D
Block x2 ResNet ℒS
Block
x Depth Times
G
1 ResNet 3
Block x2

D
G ResNet ℒT
ResNet 3 Block
Frame 0 Block x2
1 x Depth Times

x Depth Times

Per Frame
Conditoning Stack
D
3 ResNet
Block
x Depth Times

Figure 8: An architecture diagram describing the changes for the frame conditional model.

Class Conditional Batch Norm Feature Concatenation Standard Gaussian Noise

TPU core. The dual discriminator D is updated twice for every update of G (Heusel et al., 2017) and
we useReLU Activation
Spectral k Convolution
Normalization (Zhang et al., with
2018)Kernel Size k
for all weightConvolutional Gated Recurrent Unit
layers (approximated by the first
singular value) and orthogonal
Bilinear x2 Upsampling
initialization of weights (Saxe
Average Pool x2 Downsampling
et al., 2013). Sampling
One-Hot Class Indicator
is carried out
using the exponential moving average of G’s weights, which is accumulated with decay γ = 0.9999
starting after 20,000 training steps. The model is optimized using Adam (Kingma & Ba, 2014) with
batch size 512 and a learning rate of 1 · 10−4 and 5 · 10−4 for G and D respectively. Class conditioning
in D (Miyato & Koyama, 2018) is projection-based whereas G relies on class-conditional Batch
Normalization (Ioffe & Szegedy, 2015; De Vries et al., 2017; Dumoulin et al., 2017): equivalent to
standard Batch Normalization without a learned scale and offset, followed by an elementwise affine
transformation where each parameter is a function of the noise vector and class conditioning.

B.4 A RCHITECTURE E XTENSION TO V IDEO P REDICTION

In order to provide results on future video prediction problems we describe a simple modification to
DVD-GAN to facilitate the added conditioning. A diagram of the extended model is in Figure 8.
Given C conditioning frames, our modified DVD-GAN-FP passes each frame separately through a
deep residual network identical to DS . The (near) symmetric design of G and DS ’s residual blocks
mean that each output from a D-style residual block has a corresponding intermediate tensor in G
of the same spatial resolution. After each block the resulting features for each conditioning frame
are stacked in the channel dimension and passed through a 3 × 3 convolution and ReLU activation.
The resulting tensor is used as the initial state for the Convolutional GRU in the corresponding block
in G. Note that the frame conditioning stack reduces spatial resolution while G increases resolution.
Therefore the smallest features of the conditioning frames (which have been through the most layers)
are input earliest in G and the larger features (which have been through less processing) are input
to G towards the end. DT operates on the concatenation of the conditioning frames and the output
of G, meaning that it does not receive any extra information detailing that the first C frames are
special. However to reduce wasted computation we do not sample the first C frames for DS on real
or generated data. This technically means that DS will never see the first few frames from real videos
at full resolution, but this was not an issue in our experiments. Finally, our video prediction variant
does not condition on any class information, allowing us to directly compare with prior art. This is
achieved by settling the class id of all samples to 0.

C F URTHER E XPERIMENTS

C.1 UCF-101

UCF-101 (Soomro et al., 2012) is a dataset of 13,320 videos of human actions across 101 classes
that has previously been used for video synthesis and prediction (Saito et al., 2017; Saito & Saito,

14
Figure 9: The first frames of interpolations between UCF-101 samples. Each row is a separate
interpolation. Contrast with samples in Appendix E.2.

2018; Tulyakov et al., 2018). We report Inception Score (IS) calculated with a C3D network (Tran
et al., 2015) for quantitative comparison with prior work.5 Our model produces samples with an IS
of 32.97, significantly outperforming the state of the art (see Table 2). The DVD-GAN architecture
on UCF-101 is identical to the model used for Kinetics, and is trained on 16-frame 128 × 128 clips
from UCF-101.
However, it is worth mentioning that our improved score is, at least partially, due to memorization of
the training data. In Figure 9 we show interpolation samples from our best UCF-101 model. Like
interpolations in Appendix E.2, we sample 2 latents (left and rightmost columns) and show samples
from the linear interpolation in latent space along each row. Here we show 4 such interpolations (the
first frame from each video). Unlike Kinetics-600 interpolations, which smoothly transition from one
sample to the other, we see abrupt jumps in the latent space between highly distinct samples, and
little intra-video diversity between samples in each group. It can be further seen that some generated
samples highly correlate with samples from the training set.
We show this both as a failure of the Inception Score metric, the commonly reported value for class-
conditional video synthesis on UCF-101, but also as strong signal that UCF-101 is not a complex or
diverse enough dataset to facilitate interesting video generation. Each class is relatively small, and
reuse of clips from shared underlying videos means that the intra-class diversity can be restricted to
just a handful of videos per class. This suggests the need for larger, more diverse and challenging
datasets for generative video modelling, and we believe that Kinetics-600 provides a better benchmark
for this task.

D M ISCELLANEOUS E XPERIMENTS

Here we detail a number of modifications or miscellaneous results we experimented with which did
not produce a conclusive result.

• We experimented with several variations of normalization which do not require calculating

statistics over a batch of data. Group Normalization (Wu & He, 2018) performed best, almost
on a par with (but worse than) Batch Normalization. We further tried Layer Normalization
(Lei Ba et al., 2016), Instance Normalization (Ulyanov et al., 2016), and no normalization,
but found that these significantly underperformed Batch Normalization.
• We found that removing the final Batch Normalization in G, which occurs after the ResNet
and before the final convolution, caused a catastrophic failure in learning. Interestingly, just
removing the Batch Normalization layers within G’s residual blocks still led to good (though
slightly worse) generative models. In particular, variants without Batch Normalization in
the residual blocks often achieve significantly higher IS (up to 110.05 for 64 × 64 12 frame
samples – twice normal). But these models had substantially worse FID scores (1.22 for the
aforementioned model) – and produced qualitatively worse video samples.

5
We use the Chainer (Tokui et al., 2015) implementation of Inception Score for C3D available at
https://github.com/pfnet-research/tgan.

15
 • Early variants of DVD-GAN contained Batch Normalization which normalized over all
frames of all batch elements. This gave G an extra channel to convey information across time.
It took advantage of this, with the result being a model which required batch statistics in
order to produce good samples. We found that the version which normalizes over timesteps
independently worked just as well and without the dependence on statistics.
• Models based on the residual blocks of BigGAN-deep trained faster (in wall clock time)
but slower with regards to metrics, and struggled to reach the accuracy of models based on
BigGAN’s residual blocks.

E G ENERATED S AMPLES

It is difficult to accurately convey complicated generated video through still frames. Where provided,
we recommend readers view the generated videos themselves via the provided links. We refer to
videos within these batches by row/column number where the video in the 0th row and column is in
the top left corner.

E.1 S YNTHESIS S AMPLES

16
Figure 10: The first frames from a random batch of samples from DVD-GAN trained on 12
frames of 64 × 64 Kinetics-600. Full samples at https://drive.google.com/file/d/
155F1lkHA5fMAd7k4W3CQvTsi1eKQDhGb/view?usp=sharing.

Figure 11: The first frames from a random batch of samples from DVD-GAN trained on 48
frames of 64 × 64 Kinetics-600. Full samples at https://drive.google.com/file/d/
1FjOQYdUuxPXvS8yeOhXdPQMapUQaklLi/view?usp=sharing.

17
Figure 12: The first frames from a random batch of samples from DVD-GAN trained on 12
frames of 128 × 128 Kinetics-600. Full samples at https://drive.google.com/file/
d/165Yxuvvu3viOy-39LhhSDGtczbWphj_i/view?usp=sharing

18
Figure 13: The first frames from a random batch of samples from DVD-GAN trained on 48
frames of 128 × 128 Kinetics-600. Full samples at https://drive.google.com/file/
d/1P8SsWEGP6tEGPPNPH-iVycOlN6vpIgE8/view?usp=sharing. The sample in row
1, column 5 is a stereotypical example of a degenerate sample occasionally produced by DVD-GAN.

19
Figure 14: The first frames from a random batch of samples from DVD-GAN trained on 12
frames of 256 × 256 Kinetics-600. Full samples at https://drive.google.com/file/
d/1RGRVKCpVaG8z3p9GBCamRk4apiIR7jUc/view?usp=sharing.

E.2 I NTERPOLATION S AMPLES

We expect G to produce samples of higher quality from latents near the mean of the distribution
(zero). This is the idea behind the Truncation Trick (Brock et al., 2019). Like BigGAN, we find that
DVD-GAN is amenable to truncation. We also experiment with interpolations in the latent space
and in the class embedding. In both cases, interpolations are evidence that G has learned a relatively
smooth mapping from the latent space to real videos: this would be impossible for a network that has
only memorized the training data, or which is only capable of generating a few exemplars per class.
Note that while all latent vectors along an interpolation are valid (and therefore G should produce a
reasonable sample), at no point during training is G asked to generate a sample halfway between two
classes. Nevertheless G is able to interpolate between even very distinct classes.

20
Figure 15: An example intra-class interpolation. Each column is a separate video (the vertical axis is
the time dimension). The left and rightmost columns are randomly sampled latent vectors and are
generated under a shared class. Columns in between represent videos generated under the same class
across the linear interpolation between the two random samples. Note the smooth transition between
videos at all six timesteps displayed here.

Figure 16: An example of class interpolation. As before, each column is a sequence of timesteps of a
single video. Here, we sample a single latent vector, and the left and rightmost columns represent
generating a video of that latent under two different classes. Columns in between represent videos of
that same latent generated across an interpolation of the class embedding. Even though at no point
has DVD-GAN been trained on data under an interpolated class, it nevertheless produces reasonable
samples.

21