Fast Patch-based Style Transfer of Arbitrary Style

Paper: https://arxiv.org/abs/1612.04337

Code is written in Torch. CUDA and CPU modes are available.

Examples

(3x3 Patches) Content - w/ Starry Night - w/ Small Worlds I

with AvgPooling - using Inverse Network - using Inverse Network

(w/ Composition X) Original - 5x5 Patch - 9x9 Patch - 15x15 Patch

(w/ La Muse) Original - 3x3 Patch - 5x5 Patch - 9x9 Patch

Download Pretrained VGG-19

git clone https://github.com/rtqichen/style-swap
cd style-swap/models
sh download_models.sh
cd ..

Usage

Stylizing a single image:

th style-swap.lua --content images/content/bike.jpg --style images/style/starry_night.jpg

More options:

th style-swap.lua --help

eg. increase --patchSize for more abstract stylization

th style-swap.lua --content images/content/brad_pitt.jpg --style images/style/la_muse.jpg --patchSize 7 --patchStride 3

eg. use --contentBatch to stylize all images in a directory.

th style-swap.lua --contentBatch images/content --style images/style/starry_night.jpg

Training an inverse network

Install nninit module:

luarocks install nninit

Train:

th train-vgg-decoder.lua --contentDir /path/to/dir --styleDir /path/to/dir

More options:

th train-vgg-decoder.lua --help

For training the network in our paper, we used images from MS COCO and the Painter by Numbers competition hosted by Kaggle. A trained network can be downloaded here.

Video

Frame-by-frame stylization can be done using the -contentBatch option.

An example script using ffmpeg to extract frames, stylize, and re-encode a video.

mkdir video_frames
ffmpeg -i /path/to/video -qscale:v 2 video_frames/video_%04d.jpg
th style-swap --contentBatch video_frames --style /path/to/style/file --save stylized_frames
ffmpeg -i stylized_frames/video_%04d_stylized.jpg -c:v libx264 -pix_fmt yuv420p stylized_video.mp4

Examples of stylized videos are placed in the videos folder. (Original video by TimeLapseHD.)

Reducing GPU Memory Usage

A few ways to reduce memory usage for style-swap.lua:

• Decrease --maxStyleSize and --maxContentSize. The latter changes the size of the resulting image.
• Increase --patchStride. This extracts less patches to use for style swap. Best to use a larger --patchSize to ensure the patches still overlap.
• Last resort: use CPU-only mode by specifying --cpu.