Yesterady, NVidia has released the official StyleGAN3 (aka Alias-Free StyleGAN) codebase on Github.
It was last night. This morning, i could not find back the original link for the demo colab i played with this night.
I did not test it this morning. I assume it is still working. Here copy snapshot i took,
https://colab.research.google.com/drive/11J8whOt4o52e3bzFU4ttCb1-BLwOfiG6?usp=sharing
Since a couple of hours, different accounts are now releasing videos and pictures they just made with the codebase and adding also their own features. I guess a couple of new Colab will flourish very soon...
The resulting networks match the FID of StyleGAN2 but differ dramatically in their internal representations, and they are fully equivariant to translation and rotation even at subpixel scales. Our results pave the way for generative models better suited for video and animation.
Features are :
- Alias-free generator architecture and training configurations (stylegan3-t, stylegan3-r).
- Tools for interactive visualization (visualizer.py), spectral analysis (avg_spectra.py), and video generation (gen_video.py).
- Equivariance metrics (eqt50k_int, eqt50k_frac, eqr50k).
- General improvements: reduced memory usage, slightly faster training, bug fixes.
Source : https://github.com/NVlabs/stylegan3
A couple of pictures generated with the colab, without any pre or post processing here below :
Check the net with #stylegan3
Stay tuned ! This field of research is moving quite fast !
My previous posts are here
It's like "Yay, a new toy" and "Damn, I'm already overwhelmed with the current amount of AIs what don't understand properly yet", at same time hehe. So I agree with all of you, this stuff is running faster than us.
Thanks a lot for the mention and to share this with us. Outstanding results... wow!
do you have some tips for a newbie on how to start?
do I need to be a coder?
Hello @woelfchen, You don't have to be a coder but there is a small learning curve. You need a gmail account, then you will get automatically a free access to google Colab. This where you can run these "notebooks", usually, just some clicks and some field input and you will get results.
You may use some posts from our community members that gives tips, links and notes about thoses different tools, most of them are present @jotakrevs @lavista @kaliyuga and last post @castleberry,
check also the community C/The Latent Space, you may also use some online web demo see my own post https://beta.peakd.com/hive-158694/@dbddv01/my-creative-process-with-ai-generative-tools-for-drafting-the-broozel-city-guide
Enjoy.
Mit freundlichen grüssen.
Dankeschön :)
I will try looking into it :)
This field is moving faster than we can keep up I guess and the images generated look fantastic!
Thanks for the post and mentions too :)
Jotakrevs now has more work to do! But i think he will soil himself when he sees how those painting style outputs look. I agree with Lavista... this stuff is moving faster than we (i) can keep up! hahahhaa
Haha no more work, this thing works like a charm by itself lol