Sketch-RNN: A Neural Representation of Sketch Drawings

[David Ha]

Hi!

As you all know, most models these days deal with pixelated images (GANs, VAEs, PixelCNN and variants).  Even style transfer is based on ConvNet features operating in pixel space.  Wouldn't be interesting to have machine learning models generate high resolution SVG images as well?

Today we published the pre-print of our new paper, “A Neural Representation of Sketch Drawings”, that works with only with simple Vector Images (like polyline subset of SVG files). We present a recurrent neural network (sketch-rnn) generative model capable of producing sketches of everyday objects.  We would like to give you a sneak preview of the creative applications possible with this new algorithm.

This algorithm potentially has many creative applications. For instance, pattern designers can use sketch-rnn to generate a large number of similar, but unique designs for textile or wallpaper prints.

Creative designers can use sketch-rnn to come up with abstract designs.  For instance, we can feed sketches of four chairs into our model to produce four chair-like cats.  We can even explore the latent space of these vector images by using our model to interpolate between them, and view more possible designs.

Our model can also assist the creative process of an artist by suggesting alternative ways to finish an incomplete sketch. By suggesting multiple possible endings, we hope to offer a tool which expands (not diminishes!) the possible creative directions an artist might take. In this way, we view our model as a thesaurus for art, rather than an design generator. 

We plan on releasing the full sketch dataset, the code for sketch-rnn and pre-trained weights after tidying up some stuff. Then you'll be able to play with these models yourself.

Learn more about sketch-rnn by reading our paper, “A Neural Representation of Sketch Drawings”.

- David Ha, Doug Eck, and the Magenta Team.

[Adam Wang]

Nice work and beautiful paper on print, esp. for the tech paper. Waiting for your code. Assuming it's based on Tensorflow?

[Ben Morss]

I was really blown away by this - especially as the machine-generated sketches truly resemble what a human would draw.  Thanks for sharing!

I've been showing it off to people today around the office :)

--
Magenta project: magenta.tensorflow.org
To post to this group, send email to magenta...@tensorflow.org
To unsubscribe from this group, send email to magenta-discuss+unsubscribe@tensorflow.org
---
You received this message because you are subscribed to the Google Groups "Magenta Discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to magenta-discuss+unsubscribe@tensorflow.org.

[David Ha]

Thanks, we will be releasing the TensorFlow-based model.