Local Loops


Local Loops is a project about generating text using a hyper-local dataset locked into a refeed cycle. In 2021 I hiked a lot, and most of it was adjacent to my neighbourhood or immediate location in Ontario. There were several paths I took on a regular basis through the different seasons and I found this activity was a real way to ground me at a time when a lot of other areas in my life were in a state of upheaval. On these hikes I took a lot of photos, amassing a dataset of well over 1000 images. For this project I used parts of that dataset to explore caption generation through online models and tools. Those captions were then seed prompts for GPT-Neo to generate as much text as possible, which I then hand edited into small vingettes. I didn't fine-tune the model because I wanted to explore where I could push it around to (much the same as the GPT-looper project). The results were interesting, but also much more coherent than GPT-2 in 2019 which gave them a different flavour.

After hand editing the vingettes I made a small tickertape zine, which I layed out in Affinity Publisher. I was originally going to work with my old printer code, but ran into issues hacking the bluetooth on the new printer, and found my older adafruit printers didn't offer the fidelity I wanted for the images. I also created 8 unique collages out of found items, and images that I included in the zines.

The end result is something more reflective of my earlier practice, when I was doing zines and printmaking. It was refreshing to deal with text, leading, and layout again, after many years of dealing with code and firmware. I also created a useful database of tools around NLP/Caption generation. Which, when formatted I will share here.

Made Possible by a Canada Council Explore and Create Grant

Check out the Blog Posts.

Visit the Flickr Album