reflection of a reflection of a reflection

“reflection of a reflection of a reflection” (roaroar in short), is a synthographic motion picture generated with Stable Diffusion.

A story about soups, identity and a misaligned creature, roaroar presents the journey of ⅄ finding their identity in a reality mediated by generative machine learning models. Inspired by the 2013 blog post “I, We, Waluigi: a Post-Modern analysis of Waluigi” by Franck Ribery, the film reflects on what it means to only exist in reference to others. ⅄, just like generative machine learning models, bears no relation to reality whatsoever.

The film was created through a novel technique explored by mixing Stable Diffusion, 3D animation and crowd-sourced datasets. With the goal of highlighting the medium’s limitations, roaroar keeps the aesthetic of synthetically generated imagery while still being enough coherent and intentional to work as a story. Each character had their own fine-tuned Stable Diffusion model, which were trained with 34 fanart images each. The dataset imagery was created by humans with the intention of compiling a collection of the same character in very different styles. This allowed the consequent models to recreate the four characters consistently, while still being stylistically flexible.

As generative machine learning models are a disruptive technology in the field of digital media, I approached the project as transparent as possible. Aside of the behind the scenes video, showcasing the workflow, I open-sourced the fine-tuned models, the original frames before going through Stable Diffusion, and all the parameters and prompts for each scene online, for everyone to download.