This collection of articles shares what I've learned while working with the StyleGAN3-fun repo. Originally written as notes for myself, it contains lessons learned from various experiments and projects. It's a bit scattered but contains some useful tidbits.
- Big Picture of Working with StyleGAN2+3
- My Experience with Training StyleGAN2+3
- Discussion with ChatGPT about StyleGAN2 attributes
- Script Shortcuts for StyleGAN2+3
- Internal Representation Mode for StyleGAN3
- Installation Instructions (CUDA 11.1)
- Installation Instructions (CUDA 11.3)
- Hack Fix for Training on Multiple GPUs
I always include technical notes with my VJ pack releases. The links listed below contain extra bits of experience related to StyleGAN2+3, particularly with using Stable Diffusion or Flux to generate image datasets for training StyleGAN2+3.
- PACK ⬕ Wildstyle Graf
- PACK ⬕ Fashion Pose
- PACK ⬕ Mask Hypno
- PACK ⬕ Chrome Contort
- PACK ⬕ Cyborg Fomo
- PACK ⬕ Robotics Foundry
- PACK ⬕ Graffiti Reset
- PACK ⬕ Body Body Body
- PACK ⬕ Alien Guest
- PACK ⬕ Nature Artificial
- PACK ⬕ Internal Rep
- PACK ⬕ Nanotech Mutations
- PACK ⬕ Cloud Computing
- PACK ⬕ Mask Oblivion
- PACK ⬕ Orangutan VR
- PACK ⬕ Machine Stakeout
- PACK ⬕ Emoji Wreck
- PACK ⬕ Machine Landscapes
The projects I'm most proud of are the StyleGAN2 models that I trained for the Graffiti Reset and Wildstyle Graf VJ packs. To encourage experimentation and creative exploration, I've made the PKL model files available for download using the cc-by-sa-4.0 license.
- GraffitiReset_StyleGAN2_512x512_71600ImagesDataset_Gamma10_MirrorX_FineTunedUntil2340kimg.pkl
- WildstyleGraf_StyleGAN2_512x512_7256ImagesDataset_Gamma10_MirrorX-Y_FineTunedUntil9024kimg.pkl
