Sketchy Unet

The model demo running on Huggingface Spaces

I wanted a fast way to go from an image to something like a rough charcoal sketch. This would be the first step in a longer pipeline that would later add detail and colour, so all it has to do is give a starting point with the right sort of proportions.

Finding a dataset

I found a small dataset that seemed like a good starting point (originally created in ‘APDrawingGAN: Generating Artistic Portrait Drawings From Face Photos With Hierarchical GANs‘ by Ran Yi, Yong-Jin Liu, Yu-Kun Lai, Paul L. Rosin). It’s quick to download, and (with a little datablock wrangling) easy enough to load with fastai. See the notebook for details.

Training the model

I chose to model this as an image-to-image task, and used fastai’s unet_learner function to create a U-net style network based on a Resnet34 backbone. Starting with 128px images and then moving up to 224px, the model is trained to minimise the MSE between the output and the reference sketch. In about 3 minutes (!!) we end up with a model that is doing pretty much exactly what I want:

Images (left), artist’s sketch (center), model outputs (right)

Sharing a Demo

I’ve been playing around with HuggingFace Spaces recently, and this model was a great candidate for a simple demo that should run reasonably fast even on a CPU (like those provided by Spaces). At the end of the training notebook you can see the gradio interface code. Very user-friendly for these quick demos! The trained model was uploaded to huggingface as well, and they somehow detected that my code was downloading it because it shows up as a ‘linked model’ from the space.

It’s neat that I can so easily share everything related to a mini-project like this for others to follow along. The colab notebook provides a free cloud environment to replicate training, the model is hosted by someone with lots of bandwidth and is easy to download, and the demo needs no technical skills and lets anyone try it out in seconds. Hooray for fastai, gradio, huggingface and so many others who work so hard to make our lives easy 🙂

Update: What’s this for?

Waterface demo:

I used this model to ‘sketchify’ images before loading them into an imstack and optimising that to match a CLOOB prompt like ‘A charcoal and watercolor sketch of a person’. After a few steps the result looks pretty OR more likely a little creepy. Ah, the power of AI 🙂 Try it out here.


Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s