Image style transfer task aims to properly fuse a style image (e.g., color and texture patterns) into an image while the content details (e.g., shapes and objects) are effectively preserved. Recently, Neural Style Transfer has demonstrated the effectiveness of deep neural networks in style transfer by optimizing a deep neural network model in an online or offline way(Jing et al., 2019). The offline methods, typically a feed-forward network (Gatys et al., 2016), are much faster in inference as they require no optimization process after the learning of models. Specifically, arbitrary style transfer is a very challenging sub-task of offline style transfer, which aims to synthesize artistic images conditioned on arbitrary styles.
In arbitrary style transfer, the content-style feature transformation plays a vital role and a series of content-style transformation algorithms have been proposed in the literature. The widely-used normalization-based methods including instance normalization (IN) (Ulyanov et al., 2017), conditional IN (CIN) (Dumoulin et al., 2016), and adaptive IN (AdaIN) (Huang and Belongie, 2017) generally normalize the style of content features to style features. More recently, the whitening and coloring transformation (WCT) is employed to peel off and recover style information on content features. Although the whitening operation was designed to remove style information, it has side effect that content details are also removed unintentionally, leading to unpleasant synthesized results as shown in Fig. 1.
Inspired by the order statistics (David and Nagaraja, 2004)
that tells the order of random variables contains effective information, we introduce order statistics into content-style feature transformation. Specifically, we separate the content and style information into the order statistics and scalar values of features, respectively. Based on this, we propose a novel parameter-free content-style transformation algorithm, Style Projection, for real-time arbitrary image style transfer. In a simple manner, Style Projection reorders the style features according to the order of the content features, such that the correlation of feature values (shape, texture, etc.) is provided by the content features, while the scaling information (color patterns) is provided by the style features, thus enabling a reasonable content-style fusion.
We further present a real-time learning-based feed-forward model to leverage Style Projection for arbitrary style transfer. Our model works in an encoder-decoder fashion that stylizes a content image based on arbitrarily given style images. In addition to transferring style structures via Style Projection effectively, we observe that the semantics of a content image are also desired to be properly reconstructed in its corresponding synthesized image. Motivated by this, we propose to regularize the distribution distance of content images and synthesized images in feature space by minimizing the KL divergence, enabling a more distinct recovering of content details. Fig. 2 shows an example of our arbitrary image style transfer method. In experiments, we conduct extensive empirical studies including quantitative evaluation, qualitative analysis, ablation studies, and user study, to comprehensively validate the effectiveness and efficiency of our style transfer framework. The contributions of this paper are summarized as follows.
We present a parameter-free method, namely Style Projection, for fast yet effective content-style feature transformation.
We present a real-time feed-forward model for arbitrary style transfer, including a KL divergence loss for further matching the content semantics between input and output.
We demonstrate the effectiveness and efficiency of our proposed methods through extensive empirical studies.
2. Related Work
Image style transfer. Image style transfer (Li and Wand, 2016a, b; Dumoulin et al., 2016; Gatys et al., 2017; Li et al., 2017a; Luan et al., 2017; Chen et al., 2017; Ghiasi et al., 2017; Liao et al., 2017; Li et al., 2019)
is a long term open and challenging problem in computer vision. The key aspect of successfully transferring style to a synthesized image is how to fuse content and style features in an appropriate manner. This makes the problem a texture/color pattern transfer task in nature. Pioneered by(Gatys et al., 2016)
, deep learning based solutions have gained much interest. Gatys et al.(Gatys et al., 2016) for the first time adopted a convolutional network to extract features for both content and style images and match the statistics in feature space. The style transfer process starts from a random noise input and iteratively updates this input until its features match the content and style features simultaneously. Such an optimization process costs a long time to converge and limits its applications.
attempted to adopt a feed-forward neural network to replace the optimization process and achieved notable improvements on time efficiency, while the quality of synthesized images was also enhanced. For instance, Style-Swap(Chen and Schmidt, 2016) matches the statistics of content and style image patches and swaps the content patches with their closest-matching style patches. The Style Projection proposed in this paper falls into the category of fast and arbitrary style transfer. The normalization-based methods (Ulyanov et al., 2017; Dumoulin et al., 2016; Huang and Belongie, 2017) and WCT-based methods (Li et al., 2017b) achieve high running time efficiency such that they are widely used in various style transformation applications (Yin et al., 2019; Dumoulin et al., 2017; Park et al., 2019; Karras et al., 2019; An et al., 2019a, b) in practice. Recently, Li et al. (Li et al., 2019) proposed a light-weighted linear propagation method for image and video style transfer and obtained promising results.
Feature transformation. For arbitrary style transfer, content-style feature transformation algorithm is the critical component. Traditional approaches such as histogram matching (Heeger and Bergen, 1995), multi-resolution (De Bonet, 1997), and wavelet (Portilla and Simoncelli, 2000) transfer low-level statistics. However the semantics cannot be properly recovered in synthesized images. Recent advances focus on transformation algorithms that can be deployed with deep neural networks without making extra effort. WCT (Li et al., 2017b) is a typical such method which transforms content features using a whitening operation that makes the features uncorrelated, and then fuses content and style features by a coloring step which is actually the inverse of the whitening operation. Normalization-based methods, such as IN (Ulyanov et al., 2017), CIN (Dumoulin et al., 2017), and AdaIN (Huang and Belongie, 2017; Karras et al., 2019), conduct normalization on features for efficient content-style transformation. Since the normalization-based methods rely on the first-order statistics of content and style features, they are able to transfer the global style structures effectively whereas some style details may not be well captured.
In this paper, we present Style Projection to synthesize appealing stylized images in a real-time way by reordering the style features based on the order of content features. A related work is Deep Reshuffle (Gu et al., 2018) which learns a feature shuffling function under the supervision of an artifact patch-based loss. Deep Reshuffle is built upon a parametric and data-driven feature shuffling function such that it heavily relies on the distributions of training data. In addition, Deep Reshuffle involves an expensive optimization procedure in its image generation process. Our Style Projection works in a parameter-free manner, thus leading to a more efficient, robust, and generalized style transfer performance.
3.1. Arbitrary Style Transfer
The goal of arbitrary image style transfer is to stylize a content image conditioned on an arbitrary style image. To address this task, we introduce a learning-based feed-forward style transfer model. Following the standard image style transfer practice in previous literature (Huang and Belongie, 2017; Li et al., 2017b), we embed content and style images into content and style features with an image encoder . As discussed above, the critical component of our style transfer model is Style Projection algorithm, which is a parameter-free feature transformation approach for proper fusion of content and style features. Based on Style Projection, the content and style features are fused and then reconstructed as a new stylized image with an image decoder . Despite the simplicity of the model, it is effective and there is a leeway to incorporate useful tricks, for instance the semantic matching regularization of KL divergence, to further boost the style transfer performance. In the following sections, we first introduce Style Projection algorithm and then discuss the learning objectives of our style transfer model.
3.2. Style Projection Algorithm
In this work, we introduce Style Projection which is a simple, fast, yet effective algorithm for content-style fusion. Given a content feature map and a style feature map , where , , are channel, height, and width respectively, we firstly vectorize and across dimension and to obtain features and , where . Then we compute rankings for elements in and . Afterwards, we reorder by aligning each element to its corresponding same ranked element in . This actually reorganizes the style feature according to the sorting order of the content feature . Then we reshape the adjusted feature to get , which will be treated as the input of a decoder to generate a stylized image. The computing procedure of Style Projection is illustrated in Alg. 1. For clarification, a simple example of Style Projection is shown in Fig. 3.
Analysis on Style Projection. To better understand the effectiveness of Style Projection, we investigate the Gram matrix of images. The Gram matrix (Gatys et al., 2016) can be used to evaluate texture synthesis algorithms by measuring the texture correlation between images. (Gu et al., 2018) theoretically proves that the reshuffling of style features does not alter the Gram matrices of style features, i.e., the style information is well preserved after style feature reshuffling. On the other hand, Style Projection preserves the structure of content images since the features are reordered according to the order of content features. Through Style Projection, the correlation of feature values (shape, texture, etc.) from content images and the scaling information (color patterns) from style images are well fused in a parameter-free manner.
To verify this claim, we conduct an empirical study on feature shuffling mechanisms and the results are shown in Fig. 4. We observe that among the three methods, only Style Projection is capable of fusing content and style properly. ‘No reshuffling’ of style features shows an exact reconstruction of the original style image. ‘Random reshuffling’ shows a repetition of random style patterns, demonstrating that the style patterns can be preserved after feature reshuffling. Only Style Projection, which reorders the style features based on content features, shows reasonable image stylization. This study reveals the effectiveness of Style Projection as it is able to preserve both content and style information during feature transformation.
3.3. The Learning of Style Transfer Model
We illustrate our style transfer model in Fig. 5. The learning objective of our style transfer model is composed of three parts, including the style loss, the content perceptual loss, and the content KL divergence loss. The style loss is used to match the feature statistics between the style image and the stylized image as
denotes the mean and standard deviation, andis the intermediate output of the -th layer of encoder , and is the number of encoder layers. The content perceptual loss (Johnson et al., 2016) is used to minimize the pixel-wise feature distance between the content image and the stylized image as
Semantic matching as distribution learning. Most image style transfer methods suffer from non-natural image clues in stylized results. We conjecture this is partially due to the missing of semantic information such as brightness. Inspired by the insights provided by generative models (Goodfellow et al., 2014), we introduce a distribution matching objective, Kullback-Leibler (KL) divergence (Kullback and Leibler, 1951), into the style transfer framework to leverage more semantic information in content images, as
where is an input content image and is the stylized image produced by decoder . With the aid of KL divergence, we regularize to generate images that contain more semantics of content images.
Overall, the complete learning objective of our model is formulated as
where and are the loss weights of and , respectively. As Style Projection is a parameter-free method and the added KL divergence loss only brings negligible extra computing time, our style transfer approach is highly efficient.
4.1. Experimental Setup
We implement our style transfer model based on Pytorch framework(Paszke et al., 2017). The encoder network used in the model is the VGG-19 network (Simonyan and Zisserman, 2014)
pre-trained on ImageNet(Deng et al., 2009)
, and its weights are not updated during training. The decoder network is composed of nine ‘Padding-Conv-ReLU’ blocks, except the last block that has no ReLU layer. Three up-sampling layers are adopted right after the 1-st, 5-th, and 7-th block to restore the input image dimension successively, where the nearest neighbor interpolation is employed for up-sampling. We do not use normalization layer in our decoder network since it will hurt the diversity of synthesized images, as demonstrated in(Huang and Belongie, 2017).
|Pairwise Comparison (%)||Method Score (%)||
|AdaIN vs. SP||AdaIN vs. Full||SP vs. Full||AdaIN||SP||Full|
|Scene||36.36 / 63.64||28.79 / 71.21||40.15 / 59.85||32.58||51.89||65.53||68.94|
|Person||41.67 / 58.33||27.27 / 72.73||16.67 / 83.33||34.47||37.50||78.03||76.52|
|Total||39.02 / 60.98||28.03 / 71.97||28.41 / 71.59||33.52||44.70||71.78||72.73|
Datasets. We adopt the training set of MS-COCO dataset (Lin et al., 2014) as the content images and that of WikiArt dataset (Nichol, 2016) as the style images. In training, we resize all input images to the size of 512512 and randomly crop each image to 256256. All content and style images used for testing purpose are selected from the test set of the two datasets, and the test images are never observed by the model during training. Our encoder and decoder networks work in a fully-convolutional manner, and thus can be applied to images of arbitrary size in inferencec phase.
Learning details. We train our style transfer model for 160,000 iterations under the learning objective formulated in Eq. 4. We use an Adam optimizer (Kingma and Ba, 2014) with an initial learning rate of 1e-4 and a learning rate decay of 5e-5. Unless otherwise specified, the loss weights and are set to 10 and 2.5, respectively. Note that the loss weights are employed to balance the style transfer and content semantics preservation. More empirical studies on the KL divergence weight are illustrated in the following ablation studies. We use a batch size of 8 for training, and we observe that a larger batch size can only bring a marginal performance improvement.
4.2. Comparison with State-of-the-Arts
|CNN (Gatys et al., 2016)||0.90||2.51||3.41|
|Style-Swap (Chen and Schmidt, 2016)||4.87||2.56||7.43|
|WCT (Li et al., 2017b)||1.33||4.08||5.41|
|AdaIN (Huang and Belongie, 2017)||0.39||2.56||2.95|
|Linear Propagation (Li et al., 2019)||2.86||2.86||5.73|
|Style Projection + Skip||0.48||2.56||3.04|
|Style Projection + KL||0.38||2.25||2.63|
|Style Projection + Skip + KL||0.49||2.09||2.58|
Qualitative comparison. In Fig. 6, we show stylized images generated by different methods. The images generated by WCT have a low fidelity and we conjecture that this is due to the whitening operation which actually peels off some critical clues from the content images. AdaIN better reconstructs the content details, while there are still several ‘non-natural strokes’ which might be caused by only transferring the mean and standard deviation of the style features. For example, Brad Pitt’s eye (see Fig. 1
) and Lena’s face are ill-synthesized. The images produced by the Style Swap method are too dark probably due to the loss of semantics which is caused by its way of replacing content features. Linear propagation fails to recover Lena’s eyes and also fails to transfer the color of the style image into the synthesized result (e.g., Golden Gate). Our methods (last two columns) provide more pleasant results. Even using Style Projection independently, style information is properly transferred while main content is preserved. More importantly, Style Projection+Skip+KL method produces more appealing results where content details and semantics are better preserved.
Quantitative evaluation. In Table 2, we quantitatively evaluate the state-of-the-art arbitrary image style transfer methods, including Style-Swap (Chen and Schmidt, 2016), WCT (Li et al., 2017b), AdaIN (Huang and Belongie, 2017), Linear Propagation (Li et al., 2019), Style Projection (ours), and Style Projection+skip connections+KL loss (ours). We investigate the content, style and total loss in inference. Our Style Projection shows competitive results compared with the state-of-the-art style transfer modules with respect to both style and content loss. The Style Projection+Skip+KL method presents more superior results than the other methods and similar style loss as AdaIN, indicating that the modules proposed in this paper, including Style Projection and KL divergence loss, work well in arbitrary style transfer tasks. We also observe that there is a trade-off between style loss and content loss. Generally, Style Projection+KL and Style Projection+Skip+KL reach a good balance in such a trade-off.
4.3. User Study
We further conduct an user study on the images synthesized by three different methods, including AdaIN, our Style Projection, and our full model (Style Projection+Skip+KL). We use 8 content images and each is associated with 3 different style images, and thus build 24 pairs of synthesized images. Each pair is generated by two of the three involved approaches, and each approach is sampled with an equal probability. The original content and style images are also provided to the users who have no style transfer knowledge. Within each pair, the two synthesized images are randomly placed. The users are asked to choose one synthesized image from each pair under three criteria: (1) whether the image is sharp and clean while its content is close to that of the content image, (2) whether the image has a similar style as the style image, (3) which image they prefer if they want to buy.
We have collected feed-backs from 33 individuals and the results are summarized in Table 1. For all the pairs of images that are produced by AdaIN and Style Projection, more users prefer images synthesized by Style Projection. Moreover, most users favor the images generated by our full model when it is compared with AdaIN. In addition, the produced images by the full model also obtain more votes than those synthesized by Style Projection only. The “method scores” and “user consistency” rates indicate that most users all favor the synthesized images of our methods. Interestingly, for the person images, more users prefer our full model (78.03%), which indicates that fidelity is deserved to be well maintained in the synthesized person images (see Fig. 1).
|Deep Reshuffle (Gu et al., 2018)||learnable||114 seconds|
|Style Projection (Ours)||parameter-free||0.068 seconds|
A comparison of shuffling-based feature transformation methods, including Deep Feature Reshuffle(Gu et al., 2018) and our proposed Style Projection.
4.4. Study on Style Projection
Comparison between Deep Reshuffle and Style Projection. Both Deep Feature Reshuffle (Gu et al., 2018) and our Style Projection are reshuffling-based feature transformation methods. A comparison between DFS and Style Projection is illustrated in Fig. 7. The feature shuffling function in Deep Reshuffle is parametric and data-driven. Deep Reshuffle also includes an expensive optimization process in its framework, thus it is time-consuming in image generation. Our parameter-free Style Projection is able to give appealing stylized results with a high efficiency.
Trade-off between content and style in inference. We further investigate the way to control the level of stylization in Fig. LABEL:fig:tradeoff. There is a trade-off between content preservation and style transfer by adding a parameter to adjust the granularity of feature transformation. Specifically, the transformed feature , which is the input of the decoder, is combined as follows for inference.
Fig. LABEL:fig:tradeoff shows an example of image stylization with the same model but different in inference. All the images are clear in visual appearance and reasonable in stylization. The style is vastly transferred when equals 1, whereas the content image will be reconstructed with no change if is set to 0. The stylization level can be increased with the increment of . It is important to note that is not considered during training, and we fix when we compared our method with the other methods.
Style interpolation. For inference, our model is also able to stylize a content image with multiple style images. We obtain a combined feature for decoder by summing each weighted transformed feature as follows
where denotes the feature transformed from the -th style feature, the interpolation weight, and the number of style images. Fig. 9 shows the style interpolation results of our method. The content is a sailboat and the four corners are style images. As expected, the style is changing gradually while the content is reasonably preserved in synthesized images.
Spatial control. Fig. 10 shows an example of spatial control where different parts of an image are stylized based on different styles. We adopt a fore/background mask to the content image, and stylize the background with the first style image, and the foreground with the second style image. In the artistic image, the style of the fore/background can be easily distinguished.
4.5. Study on Skip Connection and KL Loss
In this paper, we propose to adopt Style Projection, content skip connections, and KL loss in the style transfer framework. To better understand the role of each module, we make ablation studies on these modules and the results are shown in Fig. 11. We have the following observations:
Without skip connections and KL loss, Style Projection module can decently transfer content details and style structure to a synthesized image (check SP).
Using either skip connections or KL divergence contributes to content detail reconstruction (check SP, SP+KL, and SP+Skip). Moreover, using three modules simultaneously gives a more appealing result (check SP+Skip+KL).
AdaIN (Huang and Belongie, 2017) can also be benefited from content skip connections and KL loss (check AdaIN+Skip+KL). Style Projection is generally competitive or better than AdaIN with the same add-ons.
The KL loss weight can be tuned conveniently to control the level of content details preservation. We further investigate its role by conducting an empirical study in Fig. 12. A local region of an image is denoted by a black box and the zoomed-in result is placed on the top-right corner of the image. As shown in Fig. 12, with the increment of the content details in the stylized image are better preserved. When is set to a very large value (i.e., ), the content details can be well reconstructed to the content image, however the stylization effect will be sacrificed slightly.
|Method||256256 pix||512512 pix|
|CNN (Gatys et al., 2016)||15.86||50.80|
|Style-Swap (Chen and Schmidt, 2016)||1.96||5.77|
|WCT (Li et al., 2017b)||0.689||0.997|
|AdaIN (Huang and Belongie, 2017)||0.036||0.064|
|Avatar-Net (Sheng et al., 2018)||0.248||0.356|
|Deep Reshuffle (Gu et al., 2018)||-||114|
|Ours||0.031 / 0.004||0.068 / 0.036|
|Ours+Skip+KL||0.031 / 0.005||0.069 / 0.036|
4.6. Real-Time Style Transfer
We demonstrate that our proposed method achieves real-time inference in Table 3. The inference time is measured with a NVIDIA GTX 1080Ti GPU on the testing set, averaging over the time of producing 220 synthesized images. As can be seen, our method has a similar inference time expense to the other real-time approaches such as AdaIN. We observe that the main computing cost of our method comes from the encoder and decoder network. The introduction of Style Projection and KL divergence loss brings negligible extra computing cost. The inference time without I/O process tells that our style transfer model and its upgraded version Ours+Skip+KL are very efficient especially on low-resolution images (256256).
In this paper, we have presented a real-time feed-forward model for arbitrary style transfer, including a parameter-free, fast, and universal content-style feature transformation module, Style Projection. We have also introduced the KL divergence loss into our style transfer model for a regularization of semantic consistency on content structures. We have conducted extensive experiments, including quantitative evaluation, qualitative analysis, and user study to validate the capacity and efficiency of our method on image style transfer tasks.
- An et al. (2019a) Jie An, Haoyi Xiong, Jiebo Luo, Jun Huan, and Jinwen Ma. 2019a. Fast Universal Style Transfer for Artistic and Photorealistic Rendering. arXiv preprint arXiv:1907.03118 (2019).
- An et al. (2019b) Jie An, Haoyi Xiong, Jinwen Ma, Jiebo Luo, and Jun Huan. 2019b. StyleNAS: An Empirical Study of Neural Architecture Search to Uncover Surprisingly Fast End-to-End Universal Style Transfer Networks. arXiv preprint arXiv:1906.02470 (2019).
- Chen et al. (2017) Dongdong Chen, Lu Yuan, Jing Liao, Nenghai Yu, and Gang Hua. 2017. Stylebank: An explicit representation for neural image style transfer. In CVPR. 1897–1906.
- Chen and Schmidt (2016) Tian Qi Chen and Mark Schmidt. 2016. Fast patch-based style transfer of arbitrary style. arXiv preprint arXiv:1612.04337 (2016).
- David and Nagaraja (2004) Herbert Aron David and Haikady Navada Nagaraja. 2004. Order statistics. Encyclopedia of Statistical Sciences (2004).
- De Bonet (1997) Jeremy S De Bonet. 1997. Multiresolution sampling procedure for analysis and synthesis of texture images. In Annual conference on computer graphics and interactive techniques. 361–368.
- Deng et al. (2009) Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. Imagenet: A large-scale hierarchical image database. In CVPR. 248–255.
- Dumoulin et al. (2016) Vincent Dumoulin, Jonathon Shlens, and Manjunath Kudlur. 2016. A learned representation for artistic style. arXiv preprint arXiv:1610.07629 (2016).
- Dumoulin et al. (2017) Vincent Dumoulin, Jonathon Shlens, and Manjunath Kudlur. 2017. A learned representation for artistic style. In ICLR.
et al. (2016)
Leon A Gatys, Alexander S
Ecker, and Matthias Bethge.
Image style transfer using convolutional neural networks. InCVPR. 2414–2423.
- Gatys et al. (2017) Leon A Gatys, Alexander S Ecker, Matthias Bethge, Aaron Hertzmann, and Eli Shechtman. 2017. Controlling perceptual factors in neural style transfer. In CVPR. 3985–3993.
- Ghiasi et al. (2017) Golnaz Ghiasi, Honglak Lee, Manjunath Kudlur, Vincent Dumoulin, and Jonathon Shlens. 2017. Exploring the structure of a real-time, arbitrary neural artistic stylization network. arXiv preprint arXiv:1705.06830 (2017).
- Goodfellow et al. (2014) Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative adversarial nets. In NIPS. 2672–2680.
- Gu et al. (2018) Shuyang Gu, Congliang Chen, Jing Liao, and Lu Yuan. 2018. Arbitrary Style Transfer with Deep Feature Reshuffle. In CVPR. 8222–8231.
- Heeger and Bergen (1995) David J Heeger and James R Bergen. 1995. Pyramid-based texture analysis/synthesis. In Annual conference on computer graphics and interactive techniques. 229–238.
- Huang and Belongie (2017) Xun Huang and Serge Belongie. 2017. Arbitrary style transfer in real-time with adaptive instance normalization. In CVPR. 1501–1510.
- Jing et al. (2019) Yongcheng Jing, Yezhou Yang, Zunlei Feng, Jingwen Ye, Yizhou Yu, and Mingli Song. 2019. Neural style transfer: A review. IEEE transactions on visualization and computer graphics (2019).
et al. (2016)
Justin Johnson, Alexandre
Alahi, and Li Fei-Fei. 2016.
Perceptual losses for real-time style transfer and super-resolution. InECCV. 694–711.
- Karras et al. (2019) Tero Karras, Samuli Laine, and Timo Aila. 2019. A Style-Based Generator Architecture for Generative Adversarial Networks. In CVPR. 4401–4410.
- Kingma and Ba (2014) Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014).
- Kullback and Leibler (1951) Solomon Kullback and Richard A Leibler. 1951. On information and sufficiency. The annals of mathematical statistics 22, 1 (1951), 79–86.
- Li and Wand (2016a) Chuan Li and Michael Wand. 2016a. Combining markov random fields and convolutional neural networks for image synthesis. In CVPR. 2479–2486.
- Li and Wand (2016b) Chuan Li and Michael Wand. 2016b. Precomputed real-time texture synthesis with markovian generative adversarial networks. In ECCV. 702–716.
et al. (2019)
Xueting Li, Sifei Liu,
Jan Kautz, and Ming-Hsuan Yang.
Learning Linear Transformations for Fast Image and Video Style Transfer. InCVPR. 3809–3817.
- Li et al. (2017a) Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. 2017a. Diversified texture synthesis with feed-forward networks. In CVPR. 3920–3928.
- Li et al. (2017b) Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. 2017b. Universal style transfer via feature transforms. In NIPS. 386–396.
- Liao et al. (2017) Jing Liao, Yuan Yao, Lu Yuan, Gang Hua, and Sing Bing Kang. 2017. Visual attribute transfer through deep image analogy. arXiv preprint arXiv:1705.01088 (2017).
- Lin et al. (2014) Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. 2014. Microsoft coco: Common objects in context. In ECCV. 740–755.
- Luan et al. (2017) Fujun Luan, Sylvain Paris, Eli Shechtman, and Kavita Bala. 2017. Deep photo style transfer. In CVPR. 4990–4998.
- Nichol (2016) K Nichol. 2016. Painter by numbers, wikiart. https://www.kaggle.com/c/painter-by-numbers.
- Park et al. (2019) Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. 2019. Semantic Image Synthesis with Spatially-Adaptive Normalization. In CVPR. 2337–2346.
- Paszke et al. (2017) Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. 2017. Automatic Differentiation in PyTorch. In NIPS Autodiff Workshop.
- Portilla and Simoncelli (2000) Javier Portilla and Eero P Simoncelli. 2000. A parametric texture model based on joint statistics of complex wavelet coefficients. IJCV 40, 1 (2000), 49–70.
- Sheng et al. (2018) Lu Sheng, Ziyi Lin, Jing Shao, and Xiaogang Wang. 2018. Avatar-net: Multi-scale zero-shot style transfer by feature decoration. In CVPR. 8242–8250.
- Simonyan and Zisserman (2014) Karen Simonyan and Andrew Zisserman. 2014. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014).
- Ulyanov et al. (2016) Dmitry Ulyanov, Vadim Lebedev, Andrea Vedaldi, and Victor S Lempitsky. 2016. Texture Networks: Feed-forward Synthesis of Textures and Stylized Images.. In ICML, Vol. 1. 4.
- Ulyanov et al. (2017) D. Ulyanov, A. Vedaldi, and V. Lempitsky. 2017. Improved texture networks: Maximizing quality and diversity in feed-forward stylization and texture synthesis. In CVPR. 6924–6932.
- Yin et al. (2019) Guojun Yin, Bin Liu, Lu Sheng, Nenghai Yu, Xiaogang Wang, and Jing Shao. 2019. Semantics Disentangling for Text-to-Image Generation. In CVPR. 2327–2336.