Famous artists are typically renowned for a particular artistic style, which takes years to develop. Even once perfected, a single piece of art can take days or even months to create. This motivates us to explore efficient computational strategies for creating artistic images. While there is a large classical literature on texture synthesis methods that create artwork from a blank canvas [8, 20, 22, 35], several recent approaches study the problem of transferring the desired style from one image onto the structural content of another image. This approach is known as artistic style transfer.
. However, recent methods that achieve impressive visual quality come from the use of convolutional neural networks (CNN) for feature extraction[9, 10, 11, 21]. The success of these methods has even created a market for mobile applications that can stylize user-provided images on demand [15, 29, 32].
Despite this renewed interest, the actual process of style transfer is based on solving a complex optimization procedure, which can take minutes on today’s hardware. A typical speedup solution is to train another neural network that approximates the optimum of the optimization in a single feed-forward pass [6, 17, 33, 34]. While much faster, existing works that use this approach sacrifice the versatility of being able to perform style transfer with any given style image, as the feed-forward network cannot generalize beyond its trained set of images. Due to this limitation, existing applications are either time-consuming or limited in the number of provided styles, depending on the method of style transfer.
In this work we propose a method that addresses these limitations: a new method for artistic style transfer that is efficient but is not limited to a finite set of styles. To accomplish this, we define a new optimization objective for style transfer that notably only depends on one layer of the CNN (as opposed to existing methods that use multiple layers). The new objective leads to visually-appealing results while this simple restriction allows us to use an “inverse network” to deterministically invert the activations from the stylized layer to yield the stylized image.
Section 2 reviews related work while Sections 3-4 describe our new optimization objective that combines style and content statistics in a single activation layer. We then propose a method of training a neural network that can invert the activations to yield an image. Section 5 presents experiments with the new method, showing it has desirable properties not found in existing formulations.
2 Related Work
Style Transfer as Optimization. Gatys et al. 
formulates style transfer as an optimization problem that combines texture synthesis with content reconstruction. Their formulation involves additive loss functions placed on multiple layers of a pretrained CNN, with some loss functions synthesizing the textures of the style image and some loss functions reconstructing the content image. Gradients are computed by backpropagation and gradient-based optimization is carried out to solve for the stylized image. An alternative approach uses patch-based similarity matching between the content and style images[9, 10, 21]. In particular, Li and Wand  constructs patch-based loss functions, where each synthetic patch has a nearest neighbour target patch that it must match. This type of patch-matching loss function is then combined additively with Gatys et al.’s formulation. While these approaches allow arbitrary style images, the optimization frameworks used by these methods makes it slow to generate the stylized image. This is particularly relevant for the case of video where we want to style a huge number of frames.
Feed-forward Style Networks As mentioned previously, it is possible to train a neural network that approximates the optimum of Gatys et al.’s loss function for one or more fixed styles [6, 17, 33, 34]. This yields a much faster method, but these methods need to be re-trained for each new style.
Style Transfer for Video. Ruder et al.  introduces a temporal loss function that, when used additively with Gatys et al.’s loss functions, can perform style transfer for videos with temporal consistency. Their loss function relies on optical flow algorithms for gluing the style in place across nearby frames. This results in an order of maginitude slowdown compared to frame-by-frame application of style transfer.
for the goal of visualization. Other works have trained inverse networks as part of an autoencoder[19, 27, 36]. To the best of our knowledge, all existing inverse networks are trained with a dataset of images and a loss function placed in RGB space.
In comparison to existing style transfer approaches, we propose a method for directly constructing the target activations for a single layer in a pretrained CNN. Like Li and Wand , we make use of a criteria for finding best matching patches in the activation space. However, our method is able to directly construct the entire activation target. This deterministic procedure allows us to easily adapt to video, without the need to rely on optical flow for fixing consistency issues. In addition to optimization, we propose a feed-forward style transfer procedure by inverting the pretrained CNN. Unlike existing feed-forward style transfer approaches, our method is not limited to specifically trained styles and can easily adapt to arbitary content and style images. Unlike existing CNN inversion methods, our method of training does not use a pixel-level loss, but instead uses a loss on the activations. By using a particular training setup (see Section 4.1), this inverse network is even able to invert activations that are outside the usual range of the CNN activiations.
3 A New Objective for Style Transfer
The main component of our style transfer method is a patch-based operation for constructing the target activations in a single layer, given the style and content images. We refer to this procedure as “swapping the style” of an image, as the content image is replaced patch-by-patch by the style image. We first present this operation at a high level, followed by more details on our implementation.
3.1 Style Swap
Let and denote the RGB representations of the content and style images (respectively), and let be the function represented by a fully convolutional part of a pretrained CNN that maps an image from RGB to some intermediate activation space. After computing the activations, and , the style swap procedure is as follows:
Extract a set of patches for both content and style activations, denoted by and , where and are the number of extracted patches. The extracted patches should have sufficient overlap, and contain all channels of the activations.
For each content activation patch, determine a closest-matching style patch based on the normalized cross-correlation measure,
Swap each content activation patch with its closest-matching style patch .
Reconstruct the complete content activations, which we denote by , by averaging overlapping areas that may have different values due to step 3.
This operation results in hidden activations corresponding to a single image with the structure of the content image, but with textures taken from the style image.
3.2 Parallelizable Implementation
To give an efficient implementation, we show that the entire style swap operation can be implemented as a network with three operations: (i) a 2D convolutional layer, (ii) a channel-wise argmax, and (iii) a 2D transposed convolutional layer. Implementation of style swap is then as simple as using existing efficient implementations of 2D convolutions and transposed convolutions111 The transposed convolution is also often referred to as a “fractionally-strided” convolution, a “backward” convolution, an “upconvolution”, or a ”deconvolution”.
The transposed convolution is also often referred to as a “fractionally-strided” convolution, a “backward” convolution, an “upconvolution”, or a ”deconvolution”..
To concisely describe the implementation, we re-index the content activation patches to explicitly denote spatial structure. In particular, we’ll let be the number of feature channels of , and let denote the patch where is the patch size.
Notice that the normalization term for content activation patches is constant with respect to the argmax operation, so (1) can be rewriten as
The lack of a normalization for the content activation patches simplifies computation and allows our use of 2D convolutional layers. The following three steps describe our implementation and are illustrated in Figure 2:
The tensorcan be computed by a single 2D convolution by using the normalized style activations patches as convolution filters and as input. The computed has spatial locations and feature channels. At each spatial location,
is a vector of cross-correlations between a content activation patch and all style activation patches.
To prepare for the 2D transposed convolution, we replace each vector by a one-hot vector corresponding to the best matching style activation patch.
The last operation for constructing is a 2D transposed convolution with as input and unnormalized style activation patches as filters. At each spatial location, only the best matching style activation patch is in the output, as the other patches are multiplied by zero.
Note that a transposed convolution will sum up the values from overlapping patches. In order to average these values, we perform an element-wise division on each spatial location of the output by the number of overlapping patches. Consequently, we do not need to impose that the argmax in (3) has a unique solution, as multiple argmax solutions can simply be interpreted as adding more overlapping patches.
, and also in RGB space. Due to the naming convention of VGG-19, “relu_1” refers to the first ReLU layer after the -th maxpooling layer. The style swap operation uses patches of size and stride , and then the RGB image is constructed using optimization.
3.3 Optimization Formulation
The pixel representation of the stylized image can be computed by placing a loss function on the activation space with target activations . Similar to prior works on style transfer [11, 21], we use the squared-error loss and define our optimization objective as
where we’ll say that the synthesized image is of dimension by by , is the Frobenius norm, and is the total variation regularization term widely used in image generation methods [1, 17, 26]. Because contains multiple maxpooling operations that downsample the image, we use this regularization as a natural image prior, obtaining spatially smoother results for the re-upsampled image. The total variation regularization is as follows:
Since the function is part of a pretrained CNN and is at least once subdifferentiable, (4) can be computed using standard subgradient-based optimization methods.
4 Inverse Network
Unfortunately, the cost of solving the optimization problem to compute the stylized image might be too high in applications such as video stylization. We can improve optimization speed by approximating the optimum using another neural network. Once trained, this network can then be used to produce stylized images much faster, and we will in particular train this network to have the versatility of being able to use new content and new style images.
The main purpose of our inverse network is to approximate an optimum of the loss function in (4) for any target activations. We therefore define the optimal inverse function as:
where represents a deterministic function and
is a random variable representing target activations. The total variation regularization term is added as a natural image prior similar to (4).
4.1 Training the Inverse Network
A couple problems arise due to the properties of the pretrained convolutional neural network.
Non-injective. The CNN defining contains convolutional, maxpooling, and ReLU layers. These functions are many-to-one, and thus do not have well-defined inverse functions. Akin to existing works that use inverse networks [4, 25, 36], we instead train an approximation to the inverse relation by a parametric neural network.
where denotes the parameters of the neural network and are activation features from a dataset of size . This objective function leads to unsupervised training of the neural network as the optimum of (4) does not need to be known. We place the description of our inverse network architecture in the appendix.
Non-surjective. The style swap operation produces target activations that may be outside the range of
due to the interpolation. This would mean that if the inverse network is only trained with real images then the inverse network may only be able to invert activations in the range of. Since we would like the inverse network to invert style swapped activations, we augment the training set to include these activations. More precisely, given a set of training images (and their corresponding activations), we augment this training set with style-swapped activations based on pairs of images.
4.2 Feedforward Style Transfer Procedure
Once trained, the inverse network can be used to replace the optimization procedure. Thus our proposed feedforward procedure consists of the following steps:
Compute and .
Obtain by style swapping.
Feed into a trained inverse network.
This procedure is illustrated in Figure 4. As described in Section 3.2, style swapping can be implemented as a (non-differentiable) convolutional neural network. As such, the entire feedforward procedure can be seen as a neural net with individually trained parts. Compared to existing feedforward approaches [6, 17, 33, 34], the biggest advantage of our feedforward procedure is the ability to use new style images with only a single trained inverse network.
In this section, we analyze properties of the proposed style transfer and inversion methods. We use the Torch7 framework  to implement our method222Code available at https://github.com/rtqichen/style-swap, and use existing open source implementations of prior works [16, 21, 30] for comparison.
5.1 Style Swap Results
Target Layer. The effects of style swapping in different layers of the VGG-19 network are shown in Figure 3. In this figure the RGB images are computed by optimization as described in Section 3. We see that while we can style swap directly in the RGB space, the result is nothing more than a recolor. As we choose a target layer that is deeper in the network, textures of the style image are more pronounced. We find that style swapping on the “relu3_1” layer provides the most visually pleasing results, while staying structurally consistent with the content. We restrict our method to the “relu3_1” layer in the following experiments and in the inverse network training. Qualitative results are shown in Figure 10.
Consistency. Our style swapping approach concatenates the content and style information into a single target feature vector, resulting in an easier optimization formulation compared to other approaches. As a result, we find that the optimization algorithm is able to reach the optimum of our formulation in less iterations than existing formulations while consistently reaching the same optimum. Figures 5 and 6 show the difference in optimization between our formulation and existing works under random initializations. Here we see that random initializations have almost no effect on the stylized result, indicating that we have far fewer local optima than other style transfer objectives.
Straightforward Adaptation to Video. This consistency property is advantageous when stylizing videos frame by frame. Frames that are the same will result in the same stylized result, while consecutive frames will be stylized in similar ways. As a result, our method is able to adapt to video without any explicit gluing procedure like optical flow . We place stylized videos in the code repository.
Simple Intuitive Tuning. A natural way to tune the degree of stylization (compared to preserving the content) in the proposed approach is to modify the patch size. Figure 7 qualitatively shows the relationship between patch size and the style-swapped result. As the patch size increases, more of the structure of the content image is lost and replaced by textures in the style image.
5.2 CNN Inversion
Here we describe our training of an inverse network that computes an approximate inverse function of the pretrained VGG-19 network . More specifically, we invert the truncated network from the input layer up to layer “relu3_1”. The network architecture is placed in the appendix.
Dataset. We train using the Microsoft COCO (MSCOCO) dataset  and a dataset of paintings sourced from wikiart.org and hosted by Kaggle . Each dataset has roughly natural images and paintings, respectively. Since typically the content images are natural images and style images are paintings, we combine the two datasets so that the network can learn to recreate the structure and texture of both categories of images. Additionally, the explicit categorization of natural image and painting gives respective content and style candidates for the augmentation described in Section 4.1.
Training. We resize each image to pixels (corresponding to activations of size
) and train for approximately 2 epochs over each dataset. Note that even though we restrict the size of our training images (and corresponding activations), the inverse network is fully convolutional and can be applied to arbitrary-sized activations after training.
We construct each minibatch by taking 2 activation samples from natural images and 2 samples from paintings. We augment the minibatch with 4 style-swapped activations using all pairs of natural images and paintings in the minibatch. We calculate subgradients using backpropagation on (7
) with the total variance regularization coefficient(the method is not particularly sensitive to this choice), and we update parameters of the network using the Adam optimizer  with a fixed learning rate of .
Result. Figure 8 shows quantitative approximation results using 2000 full-sized validation images from MSCOCO and 6 full-sized style images. Though only trained on images of size , we achieve reasonable results for arbitrary full-sized images. We additionally compare against an inverse network that has the same architecture but was not trained with the augmentation. As expected, the network that never sees style-swapped activations during training performs worse than the network with the augmented training set.
5.3 Computation Time
|Method||N. Iters.||Time/Iter. (s)||Total (s)|
|Gatys et al. ||500||0.1004||50.20|
|Li and Wand ||200||0.6293||125.86|
|Style Swap (Optim)||100||0.0466||4.66|
|Style Swap (InvNet)||1||1.2483||1.25|
on a GeForce GTX 980 Ti. Note that the number of iterations for optimization-based approaches should only be viewed as a very rough estimate.
Computation times of existing style transfer methods are listed in Table 1. Compared to optimization-based methods, our optimization formula is easier to solve and requires less time per iteration, likely due to only using one layer of the pretrained VGG-19 network. Other methods use multiple layers and also deeper layers than we do.
We show the percentage of computation time spent by different parts of our feedforward procedure in Figures 8(a) and 8(c). For any nontrivial image sizes, the style swap procedure requires much more time than the other neural networks. This is due to the style swap procedure containing two convolutional layers where the number of filters is the number of style patches. The number of patches increases linearly with the number of pixels of the image, with a constant that depends on the number of pooling layers and the stride at which the patches are extracted. Therefore, it is no surprise that style image size has the most effect on computation time (as shown in Figures 8(a) and 8(b)).
Interestingly, it seems that the computation time stops increasing at some point even when the content image size increases (Figure 8(d)), likely due to parallelism afforded by the implementation. This suggests that our procedure can handle large image sizes as long as the number of style patches is kept manageable. It may be desirable to perform clustering on the style patches to reduce the number of patches, or use alternative implementations such as fast approximate nearest neighbour search methods [12, 28].
We present a new CNN-based method of artistic style transfer that aims to be both fast and adaptable to arbitrary styles. Our method concatenates both content and style information into a single layer of the CNN, by swapping the textures of the content image with those of the style image. The simplistic nature of our method allows us to train an inverse network that can not only approximate the result in much less time, but can also generalize beyond its trained set of styles. Despite the one-layer restriction, our method can still produce visually pleasing images. Furthermore, the method has an intuitive tuning parameters (the patch size) and its consistency allows simple frame-by-frame application to videos.
While our feedforward method does not compete with the feedforward approach of Johnson et al.  in terms of speed, it should be noted that the biggest advantage of our method is the ability to stylize using new style images, whereas Johnson et al.’s method requires training a new neural network for each style image. We’ve shown that our inverse method can generalize beyond its training set (Figure 8), while another useful application is the ability to change the size of the style image without retraining the inverse network.
Some drawbacks of the proposed style swap procedure include lack of a global style measurement and lack of a similarity measure for neighboring patches, both in the spatial domain and the temporal domain. These simplifications sacrifice quality for efficiency, but can occassionally lead to a local flickering effect when applied to videos frame-by-frame. It may be desirable to look for ways to increase quality while keeping the efficient and versatile nature of our algorithm.
-  H. A. Aly and E. Dubois. Image up-sampling using total-variation regularization with a new observation model. IEEE Transactions on Image Processing, 14(10):1647–1659, 2005.
R. Collobert, K. Kavukcuoglu, and C. Farabet.
Torch7: A matlab-like environment for machine learning.In BigLearn, NIPS Workshop, 2011.
-  A. Dosovitskiy and T. Brox. Inverting convolutional networks with convolutional networks. CoRR, abs/1506.02753, 2015.
-  A. Dosovitskiy, J. Springenberg, M. Tatarchenko, and T. Brox. Learning to generate chairs, tables and cars with convolutional networks.
-  S. Y. Duck. Painter by numbers, wikiart.org. https://www.kaggle.com/c/painter-by-numbers, 2016.
-  V. Dumoulin, J. Shlens, and M. Kudlur. A learned representation for artistic style. CoRR, abs/1610.07629, 2016.
-  A. A. Efros and W. T. Freeman. Image quilting for texture synthesis and transfer. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 341–346. ACM, 2001.
-  A. A. Efros and T. K. Leung. Texture synthesis by non-parametric sampling. In Computer Vision, 1999. The Proceedings of the Seventh IEEE International Conference on, volume 2, pages 1033–1038. IEEE, 1999.
-  M. Elad and P. Milanfar. Style-transfer via texture-synthesis. arXiv preprint arXiv:1609.03057, 2016.
-  O. Frigo, N. Sabater, J. Delon, and P. Hellier. Split and match: Example-based adaptive patch sampling for unsupervised style transfer. 2016.
-  L. A. Gatys, A. S. Ecker, and M. Bethge. A neural algorithm of artistic style. CoRR, abs/1508.06576, 2015.
-  K. Hajebi, Y. Abbasi-Yadkori, H. Shahbazi, and H. Zhang. Fast approximate nearest-neighbor search with k-nearest neighbor graph.
-  A. Hertzmann. Paint By Relaxation. Proceedings Computer Graphics International (CGI), pages 47–54, 2001.
-  A. Hertzmann, C. E. Jacobs, N. Oliver, B. Curless, and D. H. Salesin. Image analogies. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 327–340. ACM, 2001.
-  A. Inc. Artify, 2016.
-  J. Johnson. neural-style. https://github.com/jcjohnson/neural-style, 2015.
J. Johnson, A. Alahi, and L. Fei-Fei.
Perceptual Losses for Real-Time Style Transfer and Super-Resolution.Arxiv, 2016.
-  D. Kingma and J. Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.
-  T. D. Kulkarni, W. F. Whitney, P. Kohli, and J. Tenenbaum. Deep convolutional inverse graphics network. In Advances in Neural Information Processing Systems, pages 2539–2547, 2015.
-  V. Kwatra, I. Essa, A. Bobick, and N. Kwatra. Texture optimization for example-based synthesis. ACM Transactions on Graphics (ToG), 24(3):795–802, 2005.
-  C. Li and M. Wand. Combining Markov Random Fields and Convolutional Neural Networks for Image Synthesis. Cvpr 2016, page 9, 2016.
-  L. Liang, C. Liu, Y.-Q. Xu, B. Guo, and H.-Y. Shum. Real-time texture synthesis by patch-based sampling. ACM Transactions on Graphics (ToG), 20(3):127–150, 2001.
-  T. Lin, M. Maire, S. J. Belongie, L. D. Bourdev, R. B. Girshick, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick. Microsoft COCO: common objects in context. CoRR, abs/1405.0312, 2014.
-  P. Litwinowicz. Processing Images and Video for an Impressionist Effect. Proc. SIGGRAPH, pages 407–414, 1997.
J. Long, E. Shelhamer, and T. Darrell.
Fully convolutional networks for semantic segmentation.
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3431–3440, 2015.
-  A. Mahendran and A. Vedaldi. Understanding deep image representations by inverting them. In 2015 IEEE conference on computer vision and pattern recognition (CVPR), pages 5188–5196. IEEE, 2015.
-  J. Masci, U. Meier, D. Cireşan, and J. Schmidhuber. Stacked convolutional auto-encoders for hierarchical feature extraction. In International Conference on Artificial Neural Networks, pages 52–59. Springer, 2011.
M. Muja and D. G. Lowe.
Scalable nearest neighbor algorithms for high dimensional data.IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(11):2227–2240, 2014.
-  I. Prisma Labs. Prisma. http://prisma-ai.com/, 2016.
-  M. Ruder, A. Dosovitskiy, and T. Brox. Artistic style transfer for videos. pages 1–14, 2016.
-  K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.
-  P. P. Studio. Picsart. https://picsart.com/, 2016.
-  D. Ulyanov, V. Lebedev, A. Vedaldi, and V. Lempitsky. Texture Networks: Feed-forward Synthesis of Textures and Stylized Images. CoRR, 2016.
-  D. Ulyanov, A. Vedaldi, and V. S. Lempitsky. Instance normalization: The missing ingredient for fast stylization. CoRR, abs/1607.08022, 2016.
-  L.-Y. Wei and M. Levoy. Fast texture synthesis using tree-structured vector quantization. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 479–488. ACM Press/Addison-Wesley Publishing Co., 2000.
-  M. D. Zeiler, D. Krishnan, G. W. Taylor, and R. Fergus. Deconvolutional networks. In Computer Vision and Pattern Recognition (CVPR), 2010 IEEE Conference on, pages 2528–2535. IEEE, 2010.
Inverse Network Architecture
The architecture of the truncated VGG-19 network used in the experiments is shown in Table A2, and the inverse network architecture is shown in Table A2. It is possible that better architectures achieve better results, as we did not try many different types of convolutional neural network architectures.
Convolutional layers use filter sizes of
, padding of, and stride of .
The rectified linear unit (ReLU) layer is an elementwise function.
The instance norm (InstanceNorm) layer standardizes each feature channel independently to have mean and a standard deviation of . This layer has shown impressive performance in image generation networks .
Maxpooling layers downsample by a factor of by using filter sizes of and stride of .
Nearest neighbor (NN) upsampling layers upsample by a factor of by using filter sizes of and stride of .
|Layer Type||Activation Dimensions|
|Layer Type||Activation Dimensions|