Skip to content
Branch: master
Find file History
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.


SVG VAE is a Tensor2Tensor problem for generating font SVGs.

This is your main access point for the model described in [A Learned Representation for Scalable Vector Graphics] (

To run any of the below commands, you first need to install Magenta as described here

Sample from a pretrained model

You can download model checkpoints from here where you'll find a tarball with the models used in the paper, as well as trained on the externally-released data described below.

After you compile the dataset (see instructions below), you can use the given colab (coming soon) to sample from the model. Simply modify the variables and run all cells:

problem = 'glyph_azzn_problem'
data_dir = '/path/to/glyphazzn_final_t2t_dataset/'
model_name = 'svg_decoder'  # or image_vae
hparam_set = 'svg_decoder'  # or image_vae
hparams = ('vae_ckpt_dir=/path/to/saved_models/image_vae,' +
ckpt_dir = '/path/to/saved_models/svg_decoder'

Train your own

Training your own model consists of two main steps:

  1. Data generation / preprocessing
  2. Training

These two steps are performed by the t2t_datagen and t2t_trainer scripts, respectively.

Data generation / preprocessing

In order to re-create the dataset, which we refer to here as glyphazzn (glyphs A-z, 0-9), you first need to collect the fonts available in the urls listed in this file.

Store these in a parquetio database with columns:

{'uni': int64,  # unicode value of this glyph
 'width': int64,  # width of this glyph's viewport (provided by fontforge)
 'vwidth': int64,  # vertical width of this glyph's viewport
 'sfd': binary/str,  # glyph, converted to .sfd format, with a single SplineSet
 'id': binary/str,  # id of this glyph
 'binary_fp': binary/str}  # font identifier (provided in glyphazzn_urls.txt)

Your end result should be a directory with a bunch of files like /path/to/glyphs-parquetio-00xxx-of-0xxxx This database is the raw data, which our will process.

To process the raw data, build datagen_beam.par, and run it in a compute cluster with [beam] (, such as [Google Cloud Dataflow] (

datagen_beam.par --logtostderr \

This will create dataset and stats files in TFRecord format. The stats files has a mean and stdev of the vectorized paths in each glyph. They will be used to normalize input to the model at train time.

These files are almost ready for use, but first we have to run t2t_datageen. This will create a few extra fields that required by tensor2tensor for training. To do this, add /path/to/tfrecord_dataset/ and /path/to/tfrecord_stats to the top of, like so:

RAW_STAT_FILE = '/path/to/tfrecord_stats-00000-of-00001'
RAW_DATA_FILES = '/path/to/tfrecord_data*'

Then run t2t_datagen:

t2t_datagen --logtostderr \
  --data_dir /path/to/glyphazzn_final_t2t_dataset \
  --tmp_dir /tmp/t2t_datagen \
  --problem glyph_azzn_problem

You're done re-creating the dataset, and ready to train the models.


To train the VAE, run:

t2t_trainer --logtostderr \
  --problem glyph_azzn_problem \
  --data_dir /path/to/glyphazzn_final_t2t_dataset \
  --output_dir /path/to/saved_models/image_vae \
  --model image_vae \
  --hparams_set image_vae \
  --train_steps 100000

After the vae is done training, train the SVG decoder like so:

t2t_trainer --logtostderr \
  --problem glyph_azzn_problem \
  --data_dir /path/to/glyphazzn_final_t2t_dataset/ \
  --output_dir /path/to/saved_models/svg_decoder \
  --model svg_decoder \
  --hparams_set svg_decoder \
  --train_steps 300000 \

Note that if you train the vae with different hparams, you must also set vae_hparams. If you change the problem, you must set vae_problem.

You can’t perform that action at this time.