In 2022, during a COVID isolation period, I worked on CYBERFLESH. I started with the inference pipelines available on GitHub.com - borisdayma/dalle-mini using trained weights from HuggingFace.co - dalle-mini/dalle-mini. I generated 256x256 pixel images based on prompts exploring how a cyberpunk kitsch AI from a science-fiction world would abstract flesh and the concept of corporeity. I then applied a series of upscaling algorithms and further elaborated the images in Photoshop, standardizing the results and framing them within a specific aesthetic framework.
Check inference_pipeline.ipynb for implementation details
My aim with this project was to investigate how physical concepts, specifically flesh and corporeity, are encoded within the trained weights of a neural network. Using plain code and trained neural network weights, I explored the visual biases embedded in the weight files through the ironic lens of a cyberpunk kitsch vision (see Bias). I sought to identify the visual archetypes related to flesh and corporeity that are statistically generated from a image dataset, while engaging with the “online grotesque” and the cheeky cyberpunk stereotypical visuals characteristic of certain kitsch science-fiction.
