Skip to content

Code for `DeepWriting: Making Digital Ink Editable via Deep Generative Modeling` paper


Notifications You must be signed in to change notification settings


Repository files navigation


Code for DeepWriting: Making Digital Ink Editable via Deep Generative Modeling paper.

Watch the video

Implementation of conditional variational RNNs (C-VRNN) fow handwriting synthesis, generation, beautification and editing tasks.


(November 2021) We released json files includeing timestamp information and stroke-based segmentations.

We collected data from 94 authors by using IAMOnDB corpus. After discarding noisy samples of IAMOnDB, we compiled a dataset of 294 authors, fully segmented. You can download our preprocessed data. For the raw dataset, license and other information, please visit our project page.

The dataset is in compressed .npz format. Strokes, labels and statistics can be accessed by using <key, value> pairs. We split handwriting samples by end-of-characters such that each sample consists of ~300 strokes. Samples are then moved to the origin and represented by using the offset values between consecutive strokes.

For example, you can visualize validation samples with indices 1, 5 and 20 by running

python -D ./data/deepwriting_validation.npz -O ./data_images -S 1 5 20

If you use our data, we kindly ask you to cite our work, and also fill IAMOnDB's registration form and follow their citation requirements since our dataset extends IAMOnDB.

Training Model

  1. Training details and hyper-parameters are defined in
  2. Download training dataset and copy into data folder. Otherwise, don't forget to update training_data and validation_data entries in
  3. Set PYTHONPATH to include source.
  1. Run training
python -S <path_to_save_experiment>
  1. If you want to continue training
python -S <path_to_save_experiment> -M <model_folder_name>

Evaluating Model

  1. You can either train a model or download our pretrained model.
  2. You can run
python -S <path_to_save_experiment> -M <model_folder_name> -QL
  1. Evaluation options are defined in globally.


  1. numpy
  2. Tensorflow 1.2+ (not sure if earlier versions work.)
  3. matplotlib
  4. OpenCV (pip install opencv-python is enough)
  5. svgwrite
  6. scikit
  7. sklearn


We are not planning to release demo interface.


If you use this code or dataset in your research, please cite us as follows:

	author = {Aksan, Emre and Pece, Fabrizio and Hilliges, Otmar},
	title = {{DeepWriting: Making Digital Ink Editable via Deep Generative Modeling}},
	booktitle = {SIGCHI Conference on Human Factors in Computing Systems},
	series = {CHI '18},
	year = {2018},
	location = {Montr{\'}eal, Canada},
	publisher = {ACM},
	address = {New York, NY, USA},


No releases published


No packages published