1. The document discusses several deep learning methods for line art colorization including pix2pix, style2paints V3, and Tag2pix.
2. Style2paints V3 uses a two-stage UNet with content loss, adversarial loss, and positive regulation loss to first generate a draft colorization and then refine it.
3. Tag2pix uses color invariant and variant tags along with a UNet, squeeze-and-excitation network, and discriminators to colorize line art conditioned on text tags.
6. What is Line Art Colorization ?
Brown_hair
Pink_dress
or or
7. Line Art Colorization Method
1.
a.
- PaintsChainer
- Two-stage Sketch Colorization (style2paints v3)[1]
- User-Guided Deep Anime Line Art Colorization with Conditional Adversarial Networks[2]
2.
a.
- Tag2Pix[3]
3.
a.
- Style Transfer for Anime Sketches with Enhanced Residual U-net and Auxiliary Classifier GAN
(style2paints v1)[4]
26. My Approach: Architecture
Line Art
Reference
Content
Encoder
Style
Encoder
AdaIN ResBlocks Decoder
Colored
Art
Discriminator
Content Encoder Style Encoder
AdaIN ResBlocks
x: Output of content encoder
y: Output of style encoder
27. Summary
1. pix2pix
a. UNet + Discriminator
b.
2. : style2paints V3
a.
b.
3. : Tag2pix
a.
4. : style2paints V1
a. Vgg19 4096 Discriminator 4096
b. AdaIN
32. [1] Lvmin Zhang, et al., “Two-stage Sketch Colorization”. SIGGRAPH ASIA 2018
[2] Yuanzheng Ci, et al., “User-Guided Deep Anime Line Art Colorization with Conditional
Adversarial Networks”. ACM Multimedia Conference 2018
[3] Hyunsu Kim, et al., “Tag2Pix: Line Art Colorization Using Text Tag With SECat and Changing
Loss”. ICCV2019
[4] Lvmin Zhang, et al., “Style Transfer for Anime Sketches with Enhanced Residual U-net and
Auxiliary Classifier GAN”. ACPR2017
[5] Phillip Isola, et al., “Image-to-Image Translation with Conditional Adversarial Nets”. CVPR2017
[6] Jie Hu, et al., “Squeeze-and-Excitation Networks”. CVPR2018
[7] Ming-Yu Liu, et al., “Few-Shot Unsupervised Image-to-Image Translation”. ICCV2019
[8] Illyasviel, “sketchKeras”. https://github.com/lllyasviel/sketchKeras
[9] Edgar Simo-Serra, et al., “Learning to Simplify: Fully Convolutional Networks for Rough Sketch
Cleanup” SIGGRAPH2016
[10] Holger Winnemoeller, et al., “XDoG: An eXtended difference-of-Gaussians compendium
including advanced image stylization” Computer & Graphics 36(6):740-753 2012