I’ve was on the hunt for an open-source style transfer implementation able to produce the output I saw on the different papers, after trying several implementations and getting underwhelming outputs, I finally came across Katherine Crowson’s style-transfer-pytorch. And wow—the results blew me away!
What Makes This Implementation Stand Out
Katherine’s implementation based on the paper “A Neural Algorithm of Artistic Style” supports both CPUs and Nvidia GPUs, and even goes as far as producing high-resolution, print-ready stylizations. Some of the cool modifications from the original paper include:
- Using PyTorch pre-trained VGG-19 weights: A small change that makes a big difference.
- Improved padding: Changing the first layer’s padding mode to
'replicate'helps reduce edge artifacts. - Scaled pooling results: Ensuring the output magnitude remains consistent.
- Wasserstein-2 style loss: For a more refined style comparison.
- Exponential moving average: To reduce noise and improve the overall output.
- Multi-scale stylization: Stylizing images at progressively larger scales (each larger by a factor of √2).
Running on MPS (Mac) Without a Hitch
While testing on my Mac, I discovered that the code was defaulting to CPU—even though my system supports MPS. After a bit of digging, I found that in cli.py (line 216), the device selection wasn’t checking for MPS hardware. Here’s the quick fix I made:
XXXXXXXXXXXXX
With this small change, I was able to run the style transfer on my Mac’s GPU, and the performance was fantastic.
The Results Speak for Themselves
For my first test, I applied the Starry Night style to a photo of the Golden Gate Bridge. The transformation was nothing short of extraordinary. (See the screenshot below, captured with asitop, showing the process running on GPU!)
The image of the Golden Gate Bridge I used to apply the Starry Night style
And the result? Check by yourself!
Big Thanks
A huge shoutout to Katherine Crowson for her exceptional work on this project. If you’re into neural style transfer and want a reliable, high-quality implementation, I highly recommend giving this a try.
As always, you can download the code here. [Edit] I’ve pushed this fix to my fork—check it out if you’re interested!
Happy styling!