Machine Learning Primer

Adding Machine Learning to our Designer's Toolbelt

During just four days, thanks to the wonderful people at Taller Estampa, we managed to collect a dataset, cluster its contents, run and modify pre-trained deep image generation networks, explore the latent space within, collect a ton of inspiration (see the link list below!) and finally launch our own speculative project using machine learning. While for now this is simply a collection of interesting findings during the seminar, in the next days I will gradually add more explanatory text to it to shine a light into the methodology used. More soon.

Image classification network grouping by similarities: vertical structures (left), graffitis (right)

Different viewpoints but same relation, grouped in the dataset: a microscopic image of a bacteria culture (left image in the yellow circle), a satellite image of a Ukrainian river delta (right image).

Increasing the Truncation PSI value, decreasing the network's incentive to "play safe" (from top left to bottom right): 0,2 / 0,5 / 1 / 1,5 / 2 / 3

Finding myself in the latent space as the network tries to approximate data points with a given base image (left) and the my 'digital twin' in the network compared to the original portrait (right) after 1000 steps.

Messing around with my latent space pendant by changing the data point obviously responsible for the level of eye-openness (left) and finding myself in another dataset of historical images (right).

Trying abstract prompts for the image generation network: 'Heavenly Universe' (left), 'Squirrel of Doom' (right)

Turning my portrait into 'a bunny with a top hat' (within 250 steps).