Is it possible to *algorithmically* make a face?

I heard about Thanos’ face in that Avengers movie being generated by AI - I thought that was MIND-BOGGLING!! :slight_smile: I’d really love to find out more about it - how was it done and could this be done in Blender? (I’m not at all knowledgable about AI but I AM reading up! :slight_smile: ). What books/webpages are there on this subject (what do I google for?)

Thanks.

search… generative adversarial networks
or search your question … Is it possible to algorithmically make a face?

or try it HERE

I’m not talking about generating a 3D model from a photo, I’m talking about the computer coming up with a new design for a face - although that COULD be from scanning hundreds of human face photos, ie. already EXISTING ones. Isn’t that possible?? That’s what they DID for Thanos, right? ie. it’s not a REAL guy?

For Thanos, they didn’t actually create his face using AI. Machine learning was used to get more accurate performance capture data from Josh Brolin’s face. They still modeled the face manually.

But yes, generating human faces is possible, given enough work and resources. Though rather than hundreds of photos, you’d probably need thousands of full, professionally made 3d scans, or else the output quality would be limited to what you can extract from a photo, that is to say, bad. If you wanted these faces to also be ready for animation, you’d also have to provide each of those thousands of scans with correct topology and a rig. And then you’d need access to enough computing power to crunch all that data - no home PC or amazon server will do, we’re talking mainframes. Perhaps some of these things can be had more cheaply in the future, but for now it’d be a big and expensive project.

That is if you want to go the AI route. A less fancy solution using good old human ingenuity is also possible, and in fact used for random NPCs in many games like Fallout, Dragon Age, etc., to a varying degree of success.

1 Like

Search the keywords I gave you…and YES they use the same process, using different faces and scanning each and generating a NEW person.

As I said…Just type your original question into Google! Then follow the links!

Er - no, I was thinking: what about sites like free3d.com , you know, STOCK models?
Does software to do such a thing already exist?

Btw, if I DID want to scan, would the ToF sensor on a Samsung Galaxy Note10+ be capable of it? If you have any idea…?

I don’t know what you mean by stock models or ‘such a thing’.

You can use any scans that you want. But by the nature of AI, the output can never be better than the training data. So if you use poor training data, you’ll get poor output.

1 Like

Just go to that site. It’ll become clear.

You can generate believable faces with smart applications and algorithms or with insanely complex simulations. The human face is not a blob of some kind, there are underlying structures, rhythms and proportions and all of this is well understood. However data training based AI takes a totally different approach than what I am talking about here.

OK, let me ask a slightly different question, because I’m new to all this. Is it possible to make a face doing a combination of art n code? Like, some Python scripting, and some “fiddling with the shape in the viewport”?? If you could get the code to generate “something”, and then “fiddle with it” by hand - that would be Pretty Cool, no? To ME at least…

You could try something like MakeHuman

Yes, but just using pure math doesn’t look right. This thread is a really good example to me of how math and algorithms don’t work 100%. This model is “mathematically perfect”, but it looks wrong. I’ve never in my life actually seen a female that has proportions like this, even though it does follow “all the mathematical rules and ratios.” Humans CAN be described by rules, but we’re not precision, factory-milled, algorithmic creatures. Hence why they used Josh Brolin’s face as a base for Thanos. You could generate the perfect human face, but it wouldn’t look natural, because there isn’t a perfect human face in reality.

1 Like

Who says that math can only archieve “perfect” and symmetrical geometry? You can also add imperfections and varations with math.

I’m using it right now! :slight_smile:

I just took a look, and I don’t follow - the artist in that thread modelled the face, right? Just like, by hand, no algorithms or anything…? Or did I read something wrong? How WOULD you do it mathematically? (if you so chose to).

I meant that they are using mathematical proportions instead of a real-life reference. They did the modelling by hand, but because it’s 100% math, it could be scripted instead. I linked it as an example of why mathematically based modeling (or generation) doesn’t really work… the proportions might be “perfect”, but they don’t look anything near realistic.

To summarize, it wasn’t an example of mathematical generation, just an example of why mathematical precision isn’t always the best option.

Your best option is to use a Fourrier series in 3d Space… But that’s rather a complicated matter, and the amount of frequencies that you’d need to make it work, will be by far, bigger (a lot bigger) than just having control vertices (a base mesh)…
So it’s pretty inpractical to look for such an approach.