Exemplar Geometry-Aware Neural Style Transfer (Part2)

Survey some Exemplar Geometry-Aware Neural Style Transfer (Continue)
type: tutoriallevel: medium

In previous blog, I have presented a survey on Geometry-Aware Neural Style Transfer (NST) methods and compare them with other standard NST methods. Furthermore, some weaknesses of current Geometry-Aware NST methods are also analyzed. In this blog, we continue to examine the effectiveness of style aware image translation in terms of different configs of hyper-parameters. Note that, all of the experiments are conducted based on the implementation of DST, NST, AdaIN, and FastDST.

IN this blog, all the images follow the structure: style image - content image with facial landmarks - style image with facial landmarks.

1. Result on various style images

In this experiment, I do style transfer using only one content image and various style images The style-weight is set to 0.5.

Fig-1

Figure 1: Different style transfer results from a static source image

2. Result on various content images

In this experiment, I do style transfer using only one style image and various content images The style-weight is set to 0.5.

Fig-2

Figure 2: Different style transfer results from different source image on a specific style image

Fig-3

Figure 3: Different style transfer results from different source image on a specific style image (Cont)

3. Result on various style-weights

In this experiment, I do style transfer using only one style image and onecontent image The style-weight is set in range: 5.0, 4.5, 4.0, 3.5, and so on to 0.5, 0.1 The lower style-weight, the more remaining content

Fig-4

Figure 4: Different style transfer results from different style weights

4. Demo

The implementation based on AAMS Colab Demo: https://colab.research.google.com/drive/1mGxE3ng8SCYunBpMmiHLA7tdnWhc7iVl?usp=sharing

5. Conclusion

Pros:

  • Faster than DST, original NST
  • Well keep semantic regions of content image (e.g. eyes, nose,...) (outperforms AdaIN, NST)
  • Do arbitrary style transfer (ASMAGAN can not)

Cons:

  • Slower than AdaIN, ASMAGAN
  • Sensitive with style image

Considered version: DST without deformation loss can improve performance and the sensitive property of AAMS. However, we must suffer from the low inference time.

6. Reference

"Attention-Aware Multi-Stroke Style Transfer - CVF Open Access." https://openaccess.thecvf.com/content_CVPR_2019/papers/Yao_Attention-Aware_Multi-Stroke_Style_Transfer_CVPR_2019_paper.pdf. Accessed 7 Oct. 2021.