Daggerfall Workshop Forums

Is that based on the NVIDIA one?

Yup! It's called StyleGAN 2.

What if we train one with the Daggerfall portraits we already have? That might yield some interesting results.

It's possible, but it requires a MASSIVE amount of computational power. However, I read an article recently that an algorithm has been developed that allows a CPU to massively outperform GPU training, so if that becomes built into the machine learning frameworks, AI technology will be much closer to being fully democratized rather than being a tool for just the rich who can afford multple Tesla V100 GPUs (which are geared towards ML applications and are very very expensive). Until then, we only have the traditional training route...

So yeah, if you have access to a bank of high performance GPUs (like multiple 1080TIs), then we could train a neural network to transfer the style of Daggerfall onto generated photo-realistic faces. But it currently takes a lot of money/time to train the model first.

MasonFace wrote: ↑Tue Mar 10, 2020 3:45 pm Yup! It's called StyleGAN 2.

It's possible, but it requires a MASSIVE amount of computational power. However, I read an article recently that an algorithm has been developed that allows a CPU to massively outperform GPU training, so if that becomes built into the machine learning frameworks, AI technology will be much closer to being fully democratized rather than being a tool for just the rich who can afford multple Tesla V100 GPUs (which are geared towards ML applications and are very very expensive). Until then, we only have the traditional training route...

So yeah, if you have access to a bank of high performance GPUs (like multiple 1080TIs), then we could train a neural network to transfer the style of Daggerfall onto generated photo-realistic faces. But it currently takes a lot of money/time to train the model first.

Thanks, I wish I had access to a bank like that I'm setting up a Docker installation to run StyleGAN 2 atm but I guess training it will be out of the question then unless I want to wait a 100 years for it to finish :lol I'll play around with it though and see what packages are available, maybe something that suits Daggerfall has already been made.

. . . maybe something that suits Daggerfall has already been made.

I hadn't even considered that; that's a good point.


Also, you may be interested in the domain of style transfer in machine learning. You could use a photo of a person, then use an upscaled Daggerfall portrait for it to transfer the style from.

MasonFace wrote: ↑Tue Mar 10, 2020 5:15 pm I hadn't even considered that; that's a good point.

Also, you may be interested in the domain of style transfer in machine learning. You could use a photo of a person, then use an upscaled Daggerfall portrait for it to transfer the style from.

Thank you. I had a quick look at the article, a bit too complex for me at the moment but I'll bookmark it because it might do the trick once I'm more comfortable with this stuff. I understand what it does but code-wise this is all new to me so I'd better start slow.

My connection is pretty shitty atm so it looks I won't have the Docker set up until tomorrow I'll let it run overnight. While the thispersondoesn'texist website worked fine, I'd prefer to generate multiple images in one go. I assume the site probably has some 'abuse' rules in place that will lock you out when you keep refreshing at a high rate.

All in all this does look like fun stuff, even if I don't produce anything useful with it.

I was thinking about ai upscaling. You produced some great results on the character portraits Mason. Those which appear in conversations. Could that pipeline produce some nice heads as well?

King of Worms wrote: ↑Tue Mar 10, 2020 5:59 pm I was thinking about ai upscaling. You produced some great results on the character portraits Mason. Those which appear in conversations. Could that pipeline produce some nice heads as well?

I'm afraid not. I haven't tried, but I recall the portrait faces are a bit higher resolution than the player heads. I have a feeling that the player heads would end up pretty badly disfigured after AI upscaling. Also, I didn't have to worry about alpha on the character portraits, which as you know can be a huge hassle.

I would try it, but I've screwed up my upscaling ML environment when I installed/upgraded all my Python stuff to work on Voiced Greetings using Real-Time Voice Cloning. I should definitely start using a virtual environment like how communityus is doing....

U are right, the heads are lower res than the portraits, and it indeed would be hard/impossible to do it right with the current tech.

I think the alpha could be dealt with, if the star constellations are right and I do it at the full moon while sacrificing virgins. Also , its "just" cca 140 files or so... so I could do it by hand, if necessary.

But yea, probably nothing I should hope for. Yes, the upscalling environment is so fragile, I was lucky I made it work now... not even sure how, because it wasnt working few months back when I tried.