24 Feb Neural Style Transfer
I’ve been working with neural net image processing for a few years now, trying to figure out how to make more powerful creative use of this class of tools, which using AI to recognize and develop patterns. Here is a favorite photo, that I took of a factory from the Willamette River at Oregon City, that I have styled with the neural filters that are built into Photoshop.
There are a number of online tools that take advantage of generative AI image processing algorithms, as well as a number of mobile apps. Wherever I start processing, I often end up taking it back through Photoshop for additional compositing and adjustments. I was excited then when Adobe decided to include some of these types of filters within Photoshop.
Many people are familiar with the strange and hallucinogenic images that come out the Deep Dream algorithm, which is trained to see things like animals and architecture in images where they may or may not be originally, and then to paint its own interpretation that focuses on the objects that it identifies.
While I’m a fan of Deep Dream, I’ve always been more interested in style transfer algorithms that allow me more deliberate control over the final output.
Besides Photoshop, I am a particular fan of Deep Dream Generator, which allows a user to explore a variety of ways to style their own images, including style transfer and Deep Dream.
The real nitty gritty though happens in Python language programming. Python code is compiled in Jupyter notebooks, which can be run either in a local Python environment, or someplace in the cloud, like Google Colab, which makes it easy for someone like myself to start digging in to the nuts and bolts of these processes.
Generally speaking, these algorithms all fall under the umbrella of TensorFlow, which is an open source machine learning platform developed by Google. TensorFlow features a number of pre-trained models which users can employ to process their own images. A great way to explore and run the code is through tutorials, where you can check out the basic Style Transfer and Deep Dream methods of processing as well as others, such as Pix2Pix.
If all you need is a fun and easy way to play with a neural network, check out Wombo, which will let you make a picture from as little as a word.