From February 28th until March 4th a workshop on image-based creativity using artificial intelligence was held in the Site-specific Art and Scenography department, taught by Chris Hales a PhD professor from United Kingdom.

Students were firstly introduced to the threats and opportunities posed by the rapid current rise in popularity of machine learning generative models and the terminology and various techniques were explained, with particular reference to “latent space”. An image-generating neural network is trained via a lengthy process to create its own imagery based on the initial set of images with which it has learnt, and this potentiality to create near-infinite outputs, many of which are unexpected, is known as latent space.

At the second session, students were shown how to use Runway ML software to go “spacewalking” through a machine learning model and were able to produce various still images and animated sequences. The third session investigated how to fine-tune the training of a classifier model so that it could function as part of an interactive installation using a webcam or video camera. Subsequently, students were shown an alternative to Runway ML which offers greater flexibility and functionality and were able to generate morphed animations using two neural networks trained by Chris Hales: one model was trained for 100 hours on 850 photos of the Hill of Crosses; the other, “MittenGAN”, was trained over ten days on 850 photos of Baltic mitten designs. On the final day of the course students experimented with four activities: finding the latent space representation of their own face within the StyleGAN2 neural network; making animated 3D sequences out of a single photograph using a MiDAS model; creating audio-responsive spacewalk animations; and generating imagery and animated sequences from text sentences using CLIP and DALL-E models.

A second part of this course will take place in the early autumn. What will follow next is for students to create and train their own neural networks based on their own photographs or scanned artwork.