isoscelesVJ
patreon
PACK ⬕ Orangutan VR
🕑 Added 2022-10-01 14:58:59 +0000 UTCTORRENT DOWNLOAD
- This pack contains 36 VJ loops (42 GB)
Sometimes when I see a person completely immersed in tech, I remember that our not so distant ancestor is the monkey.
For a while now I've been yearning for some type of AI tool to be released that would enable me to create my own images and then use it as a dataset for training StyleGAN2 or 3. So I was quite excited when I heard about DallE2, applied, and got access to the beta program. After experimenting with various text prompts I finally nailed down a mainstay and then generated 542 images. It was slow going doing it all manually to generate and save 4 at a time, but this was before Stable Diffusion was even in beta. It's wild how fast these tools are maturing. And it's bending my mind that I can use DallE2 to extrapolate new images and then have StyleGAN2 interpolate on top of it. The future for artists is looking so weird.
I was doing some exploring of the various attributes of StyleGAN3 and researching what exactly FreezeD does. Which led me to stumble across this paper explaining that you can freeze the lower levels of the discriminator to improve the fine-tune training process and help to avoid mode collapse in small datasets. I hadn't seen anyone really talking about this online and so I tried it out freezing the first 4 layers.
The results were surprising in that it definitely helped with overfitting (aka mode collapse) along with the huge bonus that it cut the training time in half! Which makes sense as only 4 of the 8 layers of the neural network are being retrained. As I understand it the first 4 layers of the network are quite small in resolution and so retraining them will not change the visuals significantly, so they are simply frozen. Overall better results, with less self repetition, and trains in half the time... Hell yeah! This is a vital finding and I'll need to revisit a few prior trainings where I had to carefully curate and organize the seeds to hide the model being overfit, particularly the Machine Graffiti pack.
After exporting some videos from StyleGAN2, it was fun to first apply dd_GlitchAssist and then composite on top of that using PixelEncoder and Pixel Sorter plugins in After Effects. And of course some slitscan liquidness. Want a banana?
