MrDeepFakes Forums

Some content may not be available to Guests. Consider registering an account to enjoy unrestricted access to guides, support and tools

  • We are looking for community members who are intested in helping out. See our HELP WANTED post.

found an AWESOME new tool

tania01

DF Admirer
Verified Video Creator
everyone has faced the problem of not having enough aligned photos of some kind of angel. i found this awesome website that can turn a single picture to a 3d model. it's not THAT great. but if it can be refined to work with DFL it would be incredible. currently it uses 1 picture, if we could give it all the pictures we have that are close to the angle we are missing, it could produce a much better result. i hope @iperov and @dpfks is willing to check it out. it could potentially revolutionize deepfakes!
the link is: http://cvl-demos.cs.nott.ac.uk/vrn/

you can this out too if you want: https://all3dp.com/2/3d-print-from-photo-how-to-do-it/ 
https://www.reallusion.com/crazytalk/features.html
 

dpfks

DF Enthusiast
Staff member
Administrator
Verified Video Creator
I've known about this already.

If you want to do tests, compare adding this vs. taking weird angles from data_dst and copying those to data_src.
 

tania01

DF Admirer
Verified Video Creator
@"dpfks" do you think it could be refined to work with DFL since both programs use machine learning. do you see any potential in this? i think you might have already tried this, if you have, which do you prefer? this or copying faces between folders?
 

dpfks

DF Enthusiast
Staff member
Administrator
Verified Video Creator
Personally I don't think the results of these sites are good enough to be placed in my facesets.
 
D

deleted1

Guest
Sometime I use Facegen Artis Pro, the result is better..... but it still looks unnatural
 

Pocketspeed

DF Admirer
Verified Video Creator
When a 3D model is created from a single image, machine learning is used to extrapolate and predict data from a single source. So there will be variances in how that data is constructed for the final result. Multiple images can help alleviate this, but at this time, cannot cover all the data necessary for complete accuracy.

So what I mean is this: you are essentially talking about artificial supplementation of data. So when you train a model, AI is making certain guesses and predictions based on the faces that it has available. If you add artificially produced faces (artificially produced data) into the model, you increase the probability of inaccuracy in the final results.

I'm not saying that it can't work. I know some creators in the past have used these very same methods to expand their facesets, and get angles that are otherwise unavailable. But as far as implementation in a DeepFake application, that's something the dev(s) will need to decide. My guess is that it probably won't happen. The 3D modelling code will need to be open-source, and even if it is, then the complexity of making this a feature of a DeepFake application will be an issue. Basically, the return isn't worth the investment.

I'm not a dev, though, so who knows?

If you use this to supplement your faces, keep in mind that the more artificial data you feed into training, the more likely that your final results will not look correct. So less is better.
 

HarryPalmer

DF Pleb
Verified Video Creator
I use 3D models from Poser and other sources to get hard-to-find angles. They work tolerably well, usually.
 
Top