Scripts

These are scripts which can be used to enhance neural-style-pt's capabilities.

Implements seamless tiling.
See the included README for more info on the new parameters and how to use them.

Use the -use_hooks parameter to use forward hooks instead of layers to compute content and style loss. Everything else is exactly the same as the original.
This version of neural-style-pt is meant more as a proof of concept to show that hooks can be used instead of layers.

Implements histogram loss layers using custom CUDA code as a separate layer type alongside content and style loss layers.
See the included README for more info on the new parameters and how to use them.

Implements mean loss layers as a separate layer type alongside content and style loss layers. Also implements optional histogram matching preprocessing.
See the included README for more info on the new parameters and how to use them.

Implements histogram loss layers as a separate layer type alongside content and style loss layers.
See the included README for more info on the new parameters and how to use them.

This version of neural-style-pt comes with an easily modifiable loss module based on the content/style loss modules, that can be used to experiment.
This script won't do anything new unless you create the loss function for the test loss module. It's meant for modders, researchers, and those looking to experiment with new loss functions.

Changes the way image sizing works, based on: https://github.com/bmaltais/neural-style

Adds -deepdream_layers, -deepdream_weight, -deepdream_channels, -deepdream_channel_mode parameters.

Unless -deepdream_channels is used, the entire chosen layer will be used for deepdream.
-deepdream_channel_mode will attempt to automatically select the strongest channels of the first value given to -deepdream_channels, while ignoring the other values.
Allows for simultaneous DeepDream and style transfer.

Allows for accurate timing of how long neural-style-pt takes to run.
Elapsed time is printed every -print_iter times, and after the script has finished.

Implements -content_channels, -style_channels, -content_channel_mode, -style_channel_mode, and -specific_content_channels parameters. The content and style channel parameters let you pick how many channels you are going to strengthen or weaken. The channel model parameters let you choose between weakening channels, and strengthening channels. The specific content channels parameter lets you select specific content channels by their id, for strengthening or weakening.
Implements the features of neural_bestchannels.lua and more. More info on this script can be found here.

A modified version of neural_style.py that prints what it thinks the content image contains, and then what it thinks the image being stylized is.
Every time the loss values are printed to the terminal, what it thinks the current image being stylized is.
Label file is included in gist for default VGG-16, and VGG-19 models.

A set of multires/multiscale resolution scripts which can easily be used with neural-style-pt and neural-style.

The successor to the original Neural-Zoom.
Neural-Zoom let's you create an "infinite zoom" effect with Neural-Style, and other style transfer projects.

This project currently implements Scale Control, and Color Control (Luminance-Only style transfer, and color matching with histogram matching) in Neural-Style using external Python scripts. The external scripts allow for the features to be used on pretty much any other style transfer system.
The features come from the Controlling Perceptual Factors in Neural Style Transfer research paper.
The Neural-Tools wiki (includes extensive usage information and results for comparison with normal style transfer), can be found here.

This project features implementations of neural style transfer written for PyTorch's Captum library.

Provide feedback