Neural multistyle transfer

An generalization and implementation of neural style for images and videos in PyTorch. Highly recommended to have a GPU to speed up the optimization process. A 512x512 image typically takes approximately 20 seconds on a 2070. The algorithm operates on videos frame-by-frame, so it can take hours depending on the resolution and size of the video.

Running for images and videos

python3 neural multistyle.py --content <content file> --styles <style file> --output <output file>

Additional arguments

Arguments	Description
content	The image or video to retain content from.
styles	The images to transfer style from.
output	Destination of output. Defaults to output.jpg or output.mp4
epochs	The number of epochs to run the LBFGS optimization algorithm for. Defaults to 301.
content_weight	A hyperparameter scalar for the importance of retaining the contents of the content image. Defaults to 1.
style_weights	A hyperparameter scalar for the importance of retaining the styles of the style images. Defaults to 1e6.
style_weights	A list of hyperparameter scalars for the importance of each style image. The number of scalars needs to be equal to the number of style images and they must sum to 1. Default is [1/n] where n is the number of style images.
verbose	Flag for printing the loss value at every iteration. Defaults to False

Examples

Neural style transfer on an image

Content Image	Style Image	Output Image	Progression

Neural multi-style transfer on an image

The video linearly transitions from style image to another.

Content Image	Style Image 1	Style Image 2	Output Image

Neural style transfer on a video

Stylized video	Source video

Neural multi-style transfer on a video

Multi-style transfer on a video starting with one style image and transitioning to the other. Frames in between are a convex sum of the two styles.

Stylized video	Source video