Skip to main content

A PyTorch implementation of Neural Style Transfer (NST)

Project description

Build Status


A PyTorch-based Python implementation of Neural Style Transfer [1].


  • Support for saving intermediate images during optimization
  • An option for preserving colors from the content image
  • Multiple-device computation (--supplemental-device)
  • Style transfers utilizing multiple style images



  • Python 3.6 or greater


$ pip3 install pastiche


$ pip3 install --upgrade pastiche


The program is intended to be used from the command line.

The general command line usage is shown below.

$ pastiche --content CONTENT --styles STYLE [STYLE ...] --output OUTPUT

CONTENT is the path to the content image, STYLE is the path to the style image, and OUTPUT is the path to save the synthesized pastiche PNG file.

If the launcher script was not installed within a directory on your PATH, pastiche can be launched by passing its module name to Python.

$ python3 -m pastiche --content CONTENT --styles STYLE [STYLE ...] --output OUTPUT

There are various options, including but not limited to:

  • Devices (GPU, CPU, and/or a multi-device assortment)
  • Number of optimization iterations
  • VGG layers to utilize
  • Loss function term weights

For the full list of options and the corresponding documentation, see the source code or use --help.

$ pastiche --help


The image above was generated by applying the style from Vincent van Gogh's The Starry Night to a photo I took in Boston in 2015. The high-resolution image was generated incrementally, with increasing resolution, using the coarse-to-fine approach described in [2]. Example commands are shown below. Depending on GPU memory availability, the commands may necessitate execution partially on a CPU (e.g., --device cuda:0 --supplemental-device cuda:1 10 --supplemental-device cpu 20 would configure GPU 0 for layers 0 through 9, GPU 1 for layers 10 through 19, and the CPU for layers 20 through 36).

$ pastiche                            \
    --device cuda                     \
    --num-steps 2000                  \
    --content boston.jpg              \
    --styles vangogh_starry_night.jpg \
    --output pastiche0.png

$ pastiche                            \
    --device cuda                     \
    --size 1024                       \
    --num-steps 1000                  \
    --init pastiche0.png              \
    --content boston.jpg              \
    --styles vangogh_starry_night.jpg \
    --output pastiche1.png

# Split the computation across a GPU (layers 0 through 5 and 10 through 19), another GPU (layers 6
# through 9), and a CPU (layers 20 through 36). This device strategy is for the purpose of
# illustration. Tuning would be required for an actual device setup.
$ pastiche                            \
    --device cuda:0                   \
    --supplemental-device cuda:1 6    \
    --supplemental-device cuda:0 10   \
    --supplemental-device cpu 20      \
    --info-step 10                    \
    --size 2048                       \
    --num-steps 500                   \
    --init pastiche1.png              \
    --content boston.jpg              \
    --styles vangogh_starry_night.jpg \
    --output pastiche2.png

# Split the computation across a GPU (layers 0 through 3) and CPU (layers 4 through 36). This
# device strategy is for the purpose of illustration. Tuning would be required for an actual
# device setup.
$ pastiche                            \
    --device cuda                     \
    --supplemental-device cpu 4       \
    --info-step 1                     \
    --size 4096                       \
    --num-steps 100                   \
    --init pastiche2.png              \
    --content boston.jpg              \
    --styles vangogh_starry_night.jpg \
    --output pastiche3.png

$ convert pastiche3.png pastiche.jpg  # requires ImageMagick

The --preserve-color option can be used to retain colors from the content image. The image below was generated using the same commands as above (up to --size 2048), with the addition of --preserve-color.


The source code has an MIT License.



[1] Gatys, Leon A., Alexander S. Ecker, and Matthias Bethge. "A Neural Algorithm of Artistic Style." ArXiv:1508.06576 [Cs, q-Bio], August 26, 2015.

[2] Gatys, Leon A., Alexander S. Ecker, Matthias Bethge, Aaron Hertzmann, and Eli Shechtman. "Controlling Perceptual Factors in Neural Style Transfer." ArXiv:1611.07865 [Cs], November 23, 2016.

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

pastiche-1.4.0.tar.gz (18.5 MB view hashes)

Uploaded source

Built Distribution

pastiche-1.4.0-py3-none-any.whl (18.5 MB view hashes)

Uploaded py3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page