ArcaneGAN by Alex Spirin

Photos

Videos

If you like what I'm doing you can:

• follow me on twitter
• check my collections at opensea
• tip me on patreon

Thank you for being awesome!

Changelog

• 2022-04-04 Added a new image2image model training url
• 2021-12-27 Added colab for videos
• 2021-12-25 ArcaneGAN v0.4 is live
• 2021-12-14 Added Video demo on huggingface
• 2021-12-12 ArcaneGAN v0.3 is live
• 2021-12-09 Thanks to ak92501 we now have a huggingface demo

ArcaneGAN v0.4

The main differences are:

• lighter styling (closer to original input)
• sharper result
• happier faces
• reduced childish eyes effect
• reduced stubble on feminine faces
• increased temporal stability on videos
• reduced mouth\teeth artifacts

Image samples

v0.3 vs v0.4

Video samples

https://user-images.githubusercontent.com/11751592/146966428-f4e27929-19dd-423f-a772-8aee709d2116.mp4

https://user-images.githubusercontent.com/11751592/146966462-6511998e-77f5-4fd2-8ad9-5709bf0cd172.mp4

ArcaneGAN v0.3

Videos processed by the huggingface video inference colab.

https://user-images.githubusercontent.com/11751592/145702737-c02b8b00-ad30-4358-98bf-97c8ad7fefdf.mp4

https://user-images.githubusercontent.com/11751592/145702740-afd3377d-d117-467d-96ca-045e25d85ac6.mp4

Image samples

Faces were enhanced via GPEN before applying the ArcaneGAN v0.3 filter.

ArcaneGAN v0.2

The release is here

Implementation Details

The model is a pytroch *.jit of a fastai v1 flavored u-net trained on a paired dataset, generated via a blended stylegan2.

• Model architecture: fastai v1 u-net
• Stylegan2 implementation used: stylegan3 repo
• Stylegan blending example: stylegan3 blending
• Paired image2image training: fastai v1 superres notebook