Maradonna - Comments

#559537 FM NewGAN Faces V1
Maradonna
13 years ago
5 months ago
52
@murrayg We are aware of this site. Issue is the limit of images. We are working with ten-thousands of images. Its just to costly.

As I said the project is alive, there is currently an issue with processing the generated images with Photoshop. The other project members are working on it, but we are a bit short on time right now. I really want this to be released latest with FM21 Beta.
#552987 FM NewGAN Faces V1
Maradonna
13 years ago
5 months ago
52
I can't give any specific date for the release. The project is still well alive though.
#545631 FM NewGAN Faces V1
Maradonna
13 years ago
5 months ago
52
Simultaneously besides improving the generated faces, I am working on an app that manages the config generation for your regens. I aim to make it cross-platform compatible (Linux/OSX/Win10). Also we teamed up with people from Zealand to improve and diversify our next facepack. Plus we have a better cleaning process, so we get rid of the "halo" most of the time completely.

We hope to release it soon, but it will still take some time. So stay tuned.
#541820 Newgen Replacement Tool
Maradonna
13 years ago
5 months ago
52
maybe put it on github or so?

I definetly interested. I am currently working with HRiddick on GAN generated faces and we are about to have a system that will be able to generate hundred of thousands new faces for potential newgens. I have a small python script that generates a config file based on a given csv with UIDs. But an all in one solution would be preferable and easier to use for the average user.

Or we use gitlab for a free private repo.

P.S: I would really like to make it platform independent, not just because I am on Linux myself, but also for the Mac users.
#541659 FM NewGAN Faces V1
Maradonna
13 years ago
5 months ago
52
There is a Python script to do it but as far as I know it is only available as part of paid-for content. Works really well though, almost no hiccups at all. Only thing I've noticed is that English players with obviously Asian names get assigned Northern European faces. I can PM you later if you're interested in more details...


definetly interested.
#541349 FM NewGAN Faces V1
Maradonna
13 years ago
5 months ago
52
I managed to remove the white halos by batch processing a simple script in Paintshop Pro utilizing the feathering tool. Combining this with ethnically specific automatic assignment and now every regen in my save has a proper face. =) Took some years to achieve but finally I have a game that SI should have made to begin with...


How did you do it? How accurate are your results?
#541093 FM NewGAN Faces V1
Maradonna
13 years ago
5 months ago
52
Excuse my French, but... Holy fuck, this is amazing. Currently using a 3D-face newgen pack, which works incredibly, but this is next-level. You mention that you could theoretically generate millions of them, but about 5% are bad. If you're willing, I'd gladly help you sort through the generated images, identify the bad ones, if it means a larger selection of faces. Obviously not millions, but from my experience, ~15.000 is where you reach a point that duplicates are so rare, you don't really notice them when you see them. Even half that would be great. Hit me up if it's something you'd be interested in, but naturally it's fair if it's too time-intensive, I don't know how long these take to generate :-)

Training takes the most time. Generating is moderate. The goal is to release a megapack in the end, so it could potentially be used for automatic regen replacement. Right now I am working with HRiddick to improve the cleaning of generated images.

These are amazing! We need more of these. Since there is no way to link these faces to skin colour of regens (as of now) anyway, you could just randomly assign those to newgen IDs like its done in the Zealand Regen Pack. These are much better faces than in that pack.

Yes of course you could use it in that way, but at first we focus on better image quality. As said above. Automatic regen replacement is the end goal. And the ethnicity option really just works if the guys who work on the live editors would help us out here. I looked into it, but this is not really my area of informatics im into. Unfortunately none of them have answered our messages, so i dont have much hope unfortunately.

First off, this is great! I've been using real faces to replace regens for 4 years now, and I was starting to struggle to find good faces for some ethnicities, that I hadn't used yet. So a theoretically endless supply of faces is perfect!

I also might have a solution for the first problem (being able to generate a specific hair color or ethnicity). Now I don't really know how the whole StyleGAN thing works (I looked at it briefly in the past, but I found it a bit to complicated), but could you in theory just use a bunch of images of for only example blond white guys to ensure you generate a blond white guy?
If that does work and you've not done it because it's too much work to sort through face packs and select all the blond guys, then maybe a facepack like this would help? It was made back when you just replaced regen hairstyles with faces so it's fully sorted by ethnicity, hair colour and hair length.

Again, no idea if that's an option and maybe you already thought of it yourself and it's too much work with the training of StyleGAN for a bunch of different smaller image sets or something else, but it might be useful.

Could be done, but training the StyleGAN2 from a pretrained network takes at least 8hrs. I think we can achieve great diversity with a well designed training set. I am already using stratified sampling, so i try to balance the different ethnicities out. The odd one might just not make it, because GANs sometimes don't reproduce super small minorities. E.g lets say there is a player with blue hair and just one in FM, im pretty confident the GAN would not produce any player with blue hair. Its just an outlier.

A future version will have the generated images pre-sorted by ethnicities. This will not be 100% accurate, especially for ethnicities that look very similar (e.g African and Carribean, or East Asian and South East Asian). But it should be good enough to aid you.

P.S: If we finish improving the image quality, there will be a staff pack as well.
#535199 FM NewGAN Faces V1
Maradonna
13 years ago
5 months ago
52
This is amazing, just gave all my youth intake a face, amazing for my youth only save, can't believe these aren't real people, extraordinary .
Can't thank you enough, they look great. My only gripe is that they do need cleaning up a bit if they're used on a dark skin as they have white ghosting around them but its not a deal breaker and can be done one by one as I use them.
Is there any chance you could do older ones for staff?
Thanks again


Im working on that white ghosting stuff, but it seems like there always will be some roughness to it. A staff pack will be possible, but at first I will try to sort some other issues out.
#534596 FM NewGAN Faces V1
Maradonna
13 years ago
5 months ago
52
https://i.imgur.com/J34IPxp.jpg
https://i.imgur.com/ksebWaR.jpg
This is a sample of newly generated faces from a machine learning algorithm.



FM NEWGAN FACES V1
  • What is this?
  • What are the limitations?
  • Why is SI not using this?
  • Can I download it?
What is This?

I trained a GAN (pretrained StyleGan2), which is a specific type of neural networks, to generate novel faces with a subset of the sortitoutsi-mega-cutout facepack. The goal of a GAN is to generate indistinguishable images compared to the input images. Therefore, it is an ideal approach to generate fotorealistic faces for newgen player inside Football Manager. More info on what a GAN is and what the application are [1].

What Are The Limitations?

First of all, we can not control the properties of a generated image, so we can not say "generate a face with blonde hair and brown eyes and give it a medium teint". This is huge in the context of Football Manager, because we need to generate a face that suits a specific newgen (e.g. african). Although there is hope, because current research tries to solve this problem and morph a image into a similar image that fulfills our desired properties [2]. Additionally, sometimes you will get garbage samples that look like an accident from Frankensteins lab. From my 2000 generated samples I had about 114 bad ones (5.7%). The hardware specification to run the training process are very high. I ran this on a Google Colab instance, this is googles cloud computing platform for AI research and makes use of high-end GPU units. Nvidia used for their pretrained network several high-end GPUs in parallel. One can not replicate this on hardware that the average user has at home. Further, even nvidia trained their networks for several days (afaik 16).

Why is SI not using this?

Beside, some of the limitations mentioned above, SI what need to sort some legal questions. First Nvidia released the StyleGan2 under a specific license that forbids usage in a commercial context. Therfore, they could train their network with a non-commercial dataset for ~200k faces. SI would need to clarify the usage of the GAN technology with Nvidia and have a dataset, which they are allowed to use. Since, Football Manager comes with some faces they might already have a dataset, but it needs to be big enough in size and have some diversity. For example you can not generate east asian faces when your set consists of caucasian faces. But the GAN doesn't care if a face is from Messi or some 4th division player, so this might be the angle.
As I mentioned above, computation is also an issue, I would suggest to figure out an agreement with nvidia to use their technology in exchange for some marketing exposure. I mean they released the noise cancellation for free, so to see some of their research applied in a market might have some considerable value for them. Maybe they could even help out with some of the limitations I mentioned.
Further, the images would ne to be generated in the cloud, it is impossible to even just generate one picture on a user spec computer. The amount of VRAM those networks need is insane and mostly above 10GB.

Can I download it?

Yes, here is a zip file with 1886 images. Technically I can generate millions of them but as mentioned above, about 5% are bad.

FAQ

placeholder


Cheers