Download Pack
This pack contains 93 VJ loops (44 GB)
https://www.patreon.com/posts/97674512
Behind the Scenes
Strike a pose! A while back I stumbled across the StyleGAN-Human codebase and was starting to figure out how to get it running. But then I realized that the base tech was StyleGAN2/3 and that they had shared their trained models! So first I explored the StyleGAN2 model and rendered out 10,000 seed images. From there I curated some seeds while being mindful to include all races and genders, rendered out the latent walk videos, and used Topaz to uprez them. This model wasn't trained on people with heavier body types which is a bummer, but that ironically tracks with fashion culture as a whole. I also rendered out a few videos using the flesh digressions script and that makes some really bizarre abstractions of the StyleGAN-Human model. Doing a bunch of various compositing experiments in After Effects was quite fun.
Having just rendered out the 10,000 seed images, I had another idea. Why not take a few of the seed images of these fashion models, inject it into Stable Diffusion via IMG2IMG, and create some extraterrestrials wearing alien fashion. With some tweaks to the text prompt I got it to consistently output images of this type and rendered out 14,205 images. From there I trained StyleGAN2 until 1204kimg and got some decent results. The issue was that the alien images were a bit too diverse in their wild fashion styles and so StyleGAN2 was having trouble converging on all of those patterns. But with some heavy seed curation I was able to make it work.
While exploring the StyleGAN-Human codebase I saw that they also had a StyleGAN3 trained model but at a lower resolution. So I wondered what the StyleGAN3 model would look like using the internal representations tool which shows the typically hidden layers of the neural network. I selected a bunch of random seeds, selected 3 channels of a few different hidden layers, and rendered out a bunch of videos. Such wild results! From there I did some color correction and extended the sides of each video.
I also did some experiments using AnimateDiff to create videos of diamonds floating in space and dollar bills rising in smoke, but the quality was lacking and so I ultimately decided to scrap it. But that did give me the idea of opening up Maya and animating a dollar bill floating around in air currents. Then I wondered what would happen if I duplicated the animation 10 times and offset it in time, which looked better than I expected. That prompted me to then add a twist deformer for another iteration. The slitscan technique was ripe for these animations and I love the liquid distortions that it adds. Anything for fashion.
Discussion (1)