Hello! Are the trained models available for download? about stylegan_reimplementation HOT 4 OPEN

Hi Shawn!
I'm planning on publishing a followup blog within a week that has more details about how to use the tool, along with better instructions for building it and a link to a pre-compiled version. I can certainly include the (unfortunately shoddy) FFHQ model. I'd also like to include the option to load a model trained with the official implementation because this one is kind of a mess.

My main concern with releasing the anime model is that many of the highest quality images it generates are NSFW. I believe that's because clothes have a lot of variance which is difficult for the model to capture, so the fewer clothes in the image the better it tends to be. I'm a bit concerned about the risk of having the blog series closely associated with NSFW image generation as that could make people (including myself) uncomfortable. However I know that I'd certainly want access to the full tool + model if I was reading the blog, so I'll see if I can come up with a way to distribute it while mitigating that risk.

I haven't previously been active on twitter, but I do have an account (@nooch_moocher) and I'd be happy to respond to questions or suggestions/criticisms there.

Thanks,
Nolan

from stylegan_reimplementation.

By the way, are you on Twitter? There's an active discussion thread about your project here: https://twitter.com/jonathanfly/status/1188551728880336896 posted by @jonathanfly.

I'm @theshawwn, fwiw.

from stylegan_reimplementation.

FYI I found a compromise I'm satisfied with: if I cut off the bottom half of the learned input before the first layer I end up with square images without a reduction in quality for the trained model. This makes the risk of NSFW images much lower. I'm going to do a bit more testing tomorrow, and then I'll share a compiled version of the tool and the model file with anyone interested. However I'm working on another blog post that should be finished in a week or so to give a tutorial, and without that it might not be clear how to do a lot of things.

from stylegan_reimplementation.

Added link to a compiled version of tool + graph in nolan-dev/GANInterface#1 (comment)

from stylegan_reimplementation.