Flexible Stroke Control in Fast Style Transfer

This repository contains the PyTorch code for our paper "Controlling Strokes in Fast Neural Style Transfer using Content Transforms", TVCJ 2022.

Make sure to pull the repository with git-lfs to retrieve the models.

Adjustable NST network

To run the adjustable model run the notebook notebooks/adjustable.ipynb, the contained interactive widget can be used to pick model variants, styles, and adjust stroke settings.

Reversible Content Transformations

To test reversible content transformations, run the notebook notebooks/reversible_warping.ipynb, or use apply_reversible.py to create animated GIFs using content transformations, such as swirl, rotation or warping. Furthermore reversible_edit/warp_gui.py contains a GUI for adjusting strokes using reversible local deformations (i.e. thin spline warping).

Training

The introduced adjustable nst network can be trained using python adjustable_upscaleNst train, it requires the ms_coco dataset, the args class contains possible configuration choices. The code in adaptiveStrokeNet.py is a pytorch re-implementation of "Stroke Controllable Fast Style Transfer", Jing et al., 2018.