stylegan3-fun VS StyleGAN-nada

I rely on the StyleGAN3-fun since it's a bit more mature and the latent walk video render tool is much better than the one within the StyleGAN2 framework. But I train using the StyleGAN2 config since it's more forgiving when using tiny datasets in the 200-1000 image range.

I'm not OP, but if you want to play around with SG3, I'd recommend you to use this version of it.

You can stabilize it with the modified code of stylegan3-fun repo to be honest, then it looks just the same way it used on stylegan2-ada

Here's a paper about adding the ability to guide outputs with text a full year before Stable Diffusion was published.

StyleGAN-NADA: CLIP-Guided Domain Adaptation of Image Generators

project page: https://stylegan-nada.github.io/

Sooo I retrained the FFHQ model to be Danny using StyleGAN-NADA via this Colab notebook.

Haha thanks for trying it out :) It was actually really challenging to get it working (especially all in the browser without processing on a server). A lot of help came from stylegan-nada https://github.com/rinongal/StyleGAN-nada (and a custom lightweight model basically distilling pairs from it and ffhq).

As for usage on non-face images most of NVidia's pre-trained models were face based (animal, humans, and paintings). Which was the aim of releasing our WikiArt model so the community would have something that could generate a greater variety of images. However these models are still constrained to the dataset that they were trained on. So without some tricks you can't generate "novel" images (like mashups of different objects)

I think the directional loss concepts in https://github.com/rinongal/StyleGAN-nada have real potential for artistic work, as they can go beyond the filter and paint effects that traditional style transfer applies well, while maintaining recognisable equivalency between the resulting images.

5 Min Read | Paper | Project | Code