Posted on

Style transfer changelog

Staying on top of the latest developments in AI is hard. We are constantly exploring new approaches and testing current best practices so that our models remain state of the art.

  • v1.1 – released August 24, 2017
  • v1.0 – released July 18, 2017

Version 1.1

Enhancements for iOS 11 beta 7

While reviewing our style transfer approach, we noticed distinct borders around the edges of the output as well as checkerboard artifacts. We have made enhancements to the model architecture to reduces these visual defects and released new versions of all our style transfer models.

Architecture improvements to the model:

  • Use reflection padding
  • Replace the deconvolution layers for cleaner upsampling
  • Removed workaround for CoreML bug in earlier iOS betas


Version 1.0

Initial release for iOS 11 beta 3

Transfer styles from famous paintings onto images and videos. Our pre-trained models transfer a specific style onto any content of yours. Fast performance allows preview transformations on live video frames as well as images.

 

Posted on

Resources

Style Transfer

  • Gatys, L.A., Ecker, A.S., Bethge, M.: A neural algorithm of artistic style. arXiv:1508.06576 (2015)
  • Ulyanov,D., Vedaldi, A., Lempitsky, V.: Instance normalization: the missing ingredient for fast stylization. arXiv:1607.08022 (2016)
  • Johnson, J., Alahi, A., and Li, F.: Perceptual losses for real-time style transfer and super- resolution. arXiv:1603.08155 (2016)
  • Engstrom, L.: Fast style transfer: https://github.com/lengstrom/fast-style-transfer
  • Changelog for version history.

Classification

Detection