23. October 2022
First demo in cables.gl. Uses the first version of Stable Diffusion by StabilityAI. Won the Demo competition at TRSAC 2022
We accidentally a full demo /o\ Cables is just too good <3 Almost all images in this demo were AI generated using Stable Diffusion. Tweaking prompts is fun, so we made a demo about it.
Effects & Cable Wrangling: dojoe & draugven Music and SFX: SunSpire GFX & Stable Diffusion: D.Fox
Made with cables.gl, which we are glad exists. Thanks a lot to pandur, steam and kb for cabling support <3
What is this?
This demo is intended to be a throwback and hommage to demos from the late 90s and early 2000s as well as a showcase for artifically generated assets in a demoscene context, which was done by using Stable Diffusion 1.4.
So - instead of talking about the merits and pitfalls of this emerging technology, we “made a demo about it” - as one should.
We started this spontaneously about three weeks before the party and didn’t have any previous experience using cables.gl.
Artificially generated images?
In addition to the use of the basic Stable Diffusion model, we experimented with training the model to generate images of ourselves by using Textual Inversion. This turned out to be a time intensive, iterative process and it took us more than a week to produce the first usable results. Overall, training our four faces took close to 100 hours of calculations, most of which were done on the Google Colab infrastructure with high-end GPUs.
What turned out to be close to impossible was the creation of a repeatable, coherent art style. Over three weeks, we generated more than 15.000 images from which > 99% were unusable or unsuitable for what we imagined. The calculation of the image took about one minute each, resulting in a computing time of about 250 hours on an Apple M1 Pro Max.