Style Transfer

... for images.

Idea

components
- original image
- style image
- generated image (initialized as noise)
... so we change input instead of weights (they are frozen)

Loss Function

$\mathcal{L}{total}(G) = \alpha \mathcal{L}{content}(C, G) + \beta \mathcal{L}_{style}(S, G)$

combines content loss and style loss

Content Loss

$\mathcal{L}{content}(C, G) = \frac{1}{2} \sum{i, j} (F_{ij}^C - F_{ij}^G)^2$

basically, takes norm for every selected layers' outputs for content and generated images
Gram matrix, which is a matrix multiplication of the feature map with its transpose

References

yt tutorial