Multi Style Transfer

PyTorch implementation of A LEARNED REPRESENTATION FOR ARTISTIC STYLE, Using UNet as the image transform network.
Trained on coco dataset(2017).

The Mandalorian Concept Art Style Transfer

This model used for creating The Mandalorian concept art
Content Image

Style Images

Output

Style Combination

Styles 1 and 5	Styles 2 and 16

TODO

• Re-implement Johnson et al work.
• Add conditional instance normalization.
• Add UNet.
• Use Mandalorian concept art as style.
• Train on larger image size, to get better spatial information.
• Use transfer learning on WikiArt for style image grouping, see here.
• Add style weights.
• Add how to use.

References

Papers

A Neural Algorithm of Artistic Style (Leon A. Gatys, Alexander S. Ecker, Matthias Bethge)
Perceptual Losses for Real-Time Style Transfer and Super-Resolution (Justin Johnson, Alexandre Alahi, and Li Fei-Fei)
Instance Normalization: The Missing Ingredient for Fast Stylization (Dmitry Ulyanov, Andrea Vedaldi, Victor Lempitsky)
A LEARNED REPRESENTATION FOR ARTISTIC STYLE (Vincent Dumoulin & Jonathon Shlens & Manjunath Kudlur)

Implementations

• Google Magenta implementation (magenta: image-stylization)
• Nice repo that was followed