This page contains pre-trained models and examples from my implementation of style transfer algorithms. Everything here is based on the method described by Justin Johnson, Alexandre Alahi, and Fei-Fei Li in Perceptual Losses for Real-Time Style Transfer and Super-Resolution. My implementation has some modifications, the details of which can be found on the GitHub page. There, you can also find instructions for training models, and using them to style images.
With a trained model, you can process images super fast - all it takes is a single pass through a convolutional neural network. Of course this needs a GPU which means you can’t use very large images due to the limited memory on GPUs. You can style much larger images on a CPU but it will take a lot of time and memory. Styling this photo (4032x3024) took more than 5 minutes, and about 120 gigabytes of memory; but it makes for a pretty cool wallpaper.
The remaining images on this page are not so huge, and were processed in about 100 milliseconds. The original images are shown below.
Next are these images modified by various styles. For each style, the trained model file is provided, and the first image in each section is the style image used for the model. Unless specified otherwise, default values were used for training arguments.
[model]
For this model, the style weight was set to 0.0001.
[model]
For this model, the style size was set to 512.
[model]
[model]
For this model, the content weight was set to 10.
[model]
[model]
For this model, the content weight was set to 5, the style weight was set to 0.0001, and the style size was set to 400.
[model]
[model]
For this model, the content weight was set to 10.
[model]
[model]