As described in the original paper on neural style transfer (https://arxiv.org/pdf/1508.06576.pdf), the basic idea is to use neural networks (specifically CNNs) to generate content from one image in the style of another, resulting in strange and fantasic works of art. I wanted to try this out for myself to understand it a bit better.
Using a pre-trained VGG-19 net, these were some of the experiments I ran:
- Varying image size.
- Varying learning rate.
- Relative importance of style and content losses.
- Higher weight to style loss componenets from latter layers vs. higher weight to earlier layers.
- Use content loss from shallow vs. deep layers.
- Adam vs. LBFGS optimiser.
- Initialise generated image to content image or style image rather than a white noise image.
- Standardise images.
- Replace max pools with average pooling.
Here is an example. I wanted to generate the image of Tony Stark bravely sacrificing himself, in the style of the abstract space man below:
The result:
