BOVINE! (2024)
Karin & Shane Denson
Bovine! is a part of the GlitchesAreLikeWildAnimalsInLatentSpace! series of AI, generative video, and painting works. Inspired in equal parts by glitch-art vernaculars, the chronophotography of Eadweard Muybridge and Étienne-Jules Marey, the cut-up methods of Brion Gysin and William Burroughs, and generative practices from Oulipo to Brian Eno and beyond, our ongoing series GlitchesAreLikeWildAnimalsInLatentSpace! stages an encounter between human imagination and automated image-making.
The above video is a screen recording of a real-time, generative/combinatory video. There are currently two versions:
Bovine.app displays generative text over combinatory video, all composited in real time. It is mathematically possible but virtually impossible that the same combination of image, sound, and text will ever be repeated.
Bovine-Video-Only.app removes text and text-to-speech elements, and only features generative audiovideo, which is assembled randomly from five cut-up versions of a single video, composited together in real-time.
The underlying video was generated in part with RunwayML (https://runwayml.com). Karin’s glitch paintings (https://karindenson.com) were used to train a model for image generation.
Prompting the model with terms like “Glitches are like wild animals” (a phrase she has been working with for years, originally found in an online glitch tutorial, now offline), and trying to avoid the usual suspects (lions, tigers, zebras), produced a glitchy cow, which Karin painted with acrylic on canvas:
The painting was fed back into RunwayML as the seed for a video clip (using Gen-2 in spring/summer 2024), which was extended a number of times. The resulting video was glitched with databending methods (in Audacity). The soundtrack was produced by feeding a jpg of the original cow painting into Audacity as raw data, interpreted with the GSM codec. After audio and video were assembled, the glitchy video was played back and captured with VLC and Quicktime, each of which interpreted the video differently. The two versions were composited together, revealing delays, hesitations, and lack of synchronization.
The full video was then cropped to produce five different strips. The audio on each was positioned accordingly in stereo space (i.e. the left-most strip has its audio turned all the way to the left, the next one over is half-way from the left to the center, the middle one is in the center, etc.). The Max app chooses randomly from a set of predetermined start points where to play each strip of video, keeping the overall image more or less in sync.
Onscreen and spoken text is generated by a Markov model trained on Shane’s book Discorrelated Images (https://www.dukeupress.edu/discorrelated-images), the cover of which featured Karin’s original GlitchesAreLikeWildAnimals! painting.
Made with Max 8 (https://cycling74.com/products/max) on a 2023 Mac Studio (Mac 14,14, 24-core Apple M2 Ultra, 64 GB RAM) running macOS Sonoma (14.6.1). Generative text is produced with Pavel Janicki’s MaxAndP5js Bridge (https://www.paweljanicki.jp/projects_maxandp5js_en.html) to interface Max with the p5js (https://p5js.org) version of the RiTa tools for natural language and generative writing (https://rednoise.org/rita/). Jeremy Bernstein’s external Max object, shell 1.0b3 (https://github.com/jeremybernstein/shell/releases/tag/1.0b3), passes the text to the OS for text-to-speech.
