Skip to content

[NeurIPS 2020] Official code for the paper "DeepSVG: A Hierarchical Generative Network for Vector Graphics Animation". Includes a PyTorch library for deep learning with SVG data.

License

Notifications You must be signed in to change notification settings

alexandre01/deepsvg

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

31 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

logo

Stars Licence AI Unfolded Website

Introduction

Sponsored by LingoSub: Learn languages by watching videos with AI-powered translations

and ThumbnailsPro: Instant AI-generated Thumbnails, for videos that get clicks.

This is the official code for the paper "DeepSVG: A Hierarchical Generative Network for Vector Graphics Animation". Please refer to section below for Citation details.

AI Unfolded

  • 1min video:

DeepSVG video


This repository includes:

  • The training code to reproduce our Hierarchical Generative Network: DeepSVG.
  • A library for deep learning with SVG data, including export functionality to differentiable PyTorch tensors.
  • The SVG-Icons8 dataset.
  • A Graphical user interface showing a demo of DeepSVG for vector graphics animation.

Updates

  • December 2020: Added raw SVG dataloader (see Dataloader section).

  • September 2020: Accepted to NeurIPS2020 🎉

  • July 2020: Added pretrained models and notebook for Font generation.

Installation

Start by cloning this Git repository:

git clone https://github.com/alexandre01/deepsvg.git
cd deepsvg

Create a new conda environment (Python 3.7):

conda create -n deepsvg python=3.7
conda activate deepsvg

And install the dependencies:

pip install -r requirements.txt

Please refer to cairosvg's documentation for additional requirements of CairoSVG. For example:

  • on Ubuntu: sudo apt-get install libcairo2-dev.
  • on macOS: brew install cairo libffi.

Tested environments

  • Ubuntu 18.04, CUDA 10.1
  • macOS 10.13.6, CUDA 10.1, PyTorch installed from source

Dataset

icons_dataset Download the dataset using the script provided in dataset/ by running:

cd dataset/
bash download.sh

If this is not working for you, download the dataset manually from Google Drive, place the files in the dataset folder, and unzip (this may take a few minutes).

By default, the dataset will be saved with the following tree structure:

deepsvg
 └─dataset/
    ├── icons_meta.csv
    └── icons_tensor/

NOTE: The icons_tensor/ folder contains the 100k icons in pre-augmented PyTorch tensor format, which enables to easily reproduce our work. For full flexbility and more research freedom, we however recommend downloading the original SVG icons from icons8, for which you will need a paid plan. Instructions to download the dataset from source are coming soon.

To download the Font-dataset, we recommend following SVG-VAE's instructions: https://github.com/magenta/magenta/tree/master/magenta/models/svg_vae. For demo purposes, we also release a mini version of the dataset. To download it, run:

cd dataset/
bash download_fonts.sh

Or use these links:

Dataloader

To process a custom dataset of SVGs, use the SVGDataset dataloader. To preprocess them on the fly, you can set already_preprocessed to False, but we recommend preprocessing them before training for better I/O performance.

To do so, use the dataset/preprocess.py script:

python -m dataset.preprocess --data_folder dataset/svgs/ --output_folder dataset/svgs_simplified/ --output_meta_file dataset/svg_meta.csv

This will preprocess all input svgs in a multi-threaded way and generate a meta data file, for easier training filtering.

Then modify the training configuration by providing the correct dataloader module, data folder and meta data file:

cfg.dataloader_module = "deepsvg.svg_dataset"
cfg.data_dir = "./dataset/svgs_simplified/"
cfg.meta_filepath = "./dataset/svg_meta.csv"

Deep learning SVG library

DeepSVG has been developed along with a library for deep learning with SVG data. The main features are:

  • Parsing of SVG files.
  • Conversion of basic shapes and commands to the subset m/l/c/z.
  • Path simplification, using Ramer-Douglas-Peucker and Philip J. Schneider algorithms.
  • Data augmentation: translation, scaling, rotation of SVGs.
  • Conversion to PyTorch tensor format.
  • Draw utilities, including visualization of control points and exporting to GIF animations.

The notebook notebooks/svglib.ipynb provides a walk-trough of the deepsvg.svglib library. Here's a little sample code showing the flexibility of our library:

from deepsvg.svglib.svg import SVG
from deepsvg.svglib.geom import Point, Angle

icon = SVG.load_svg("docs/imgs/dolphin.svg").normalize()
icon.simplify_heuristic()                                 # path simplifcation
icon.zoom(0.75).translate(Point(0, 5)).rotate(Angle(15))  # scale, translate, rotate
icon.draw()

dolphin_png

And making a GIF of the SVG is as easy as:

icon.animate()

dolphin_animate

Differentiable SVG shape optimization

Similarly to PyTorch3D, differentiable operations can be performed on a SVGTensor, enabling to deform a circle to match an arbitrary target via gradient descent. Interestingly, using a lower number of Bézier commands in the initial circle (n) creates somehow artistic approximations of the target shape. See notebooks/svgtensor.ipynb.

n 4 8 16 32
optimization dolpin_optim_4 dolpin_optim_8 dolpin_optim_16 dolpin_optim_32

Graphical User Interface (experimental)

While developing DeepSVG, we have also built a Graphical User Interface (GUI) for easier visualization of our model and as a tool to easily create 2D animations. The code, available under deepsvg/gui, is written with Kivy and the UI style is inspired from the design tool Figma.

DISCLAIMER: this GUI has been developed for demo purposes mainly and features one would expect from a vector graphics editor (like rescaling) will be added in the near future. For more flexibility, we recommend creating animations programmatically using the template code provided in notebooks/animation.ipynb.

gui

Shortcuts:

  • H: hand tool
  • P: pen tool
  • Ctrl/⌘ Cmd P: pencil tool
  • K: make keyframe
  • spacebar: play/pause
  • Ctrl/⌘ Cmd E: export animation to GIF (file is saved by default in ./gui_data)

Training

Start a training by running the following command.

python -m deepsvg.train --config-module configs.deepsvg.hierarchical_ordered

The (optional) --log-dir argument lets you choose the directory where model checkpoints and tensorboard logs will be saved.

Inference (interpolations)

Download pretrained models by running:

cd pretrained/
bash download.sh

If this doesn't work, you can download them manually from Google Drive and place them in the pretrained folder.

We provide sample code in notebooks/interpolate.ipynb to perform interpolation between pairs of SVG icons and notebooks/latent_ops.ipynb for word2vec-like operations in the SVG latent space, as shown in the experiments of our paper.

Notebooks

Description Link to notebook
SVGLib walk-through svglib.ipynb
Differentiable SVGTensor optimization svgtensor.ipynb
DeepSVG interpolation between pairs of icons interpolation.ipynb
DeepSVG latent space operations latent_ops.ipynb
DeepSVG animation between user-drawn images animation.ipynb
DeepSVG generation and interpolation of Fonts fonts.ipynb

Citation

If you find this code useful in your research, please cite:

@misc{carlier2020deepsvg,
    title={DeepSVG: A Hierarchical Generative Network for Vector Graphics Animation},
    author={Alexandre Carlier and Martin Danelljan and Alexandre Alahi and Radu Timofte},
    year={2020},
    eprint={2007.11301},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

Contributing

Contributions are welcome! Feel free to submit a pull request if you have found a bug or developed a feature that may be useful for others. Also, don't hesitate to contact me for any further question related to this code repository.

Licence

This code is released under the MIT licence.

About

[NeurIPS 2020] Official code for the paper "DeepSVG: A Hierarchical Generative Network for Vector Graphics Animation". Includes a PyTorch library for deep learning with SVG data.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published