ProgrammingCan this be done?Posted:

Txby
  • TTG Senior
Status: Offline
Joined: Jun 29, 20136 Year Member
Posts: 1,254
Reputation Power: 61
I have a new business concept which relies on an app that can virtually create a photo-realistic avatar of someone.

Can this be done?

Thanks

(sorry if this is in the wrong forum, didn't know which would be most suitable
#2. Posted:
TTG
  • Winter 2017
Status: Offline
Joined: Mar 10, 20145 Year Member
Posts: 3,838
Reputation Power: 5315
Motto: Corona V2 RGH Owner, NiNJA User, Anime Is Life.
Yes this is definitely possible, it has been implemented already. VR has been doing this for quite some time, but from a business perspective this is something good to create and even have people invest in.
#3. Posted:
CriticaI
  • Supporter
Status: Offline
Joined: Nov 05, 20135 Year Member
Posts: 2,673
Reputation Power: 154
this is possible for sure, but it is a lot easier said than done. You would basically need to create some kind of program, possibly one that uses machine learning, that detects common facial features. You would then map those facial features to a list of premade images and you would stitch everything together.

I wish you the best of luck. There is a lot of money being thrown at projects like this.
#4. Posted:
iyop45
  • Prospect
Status: Offline
Joined: Apr 15, 20118 Year Member
Posts: 614
Reputation Power: 83
CriticaI wrotethis is possible for sure, but it is a lot easier said than done. You would basically need to create some kind of program, possibly one that uses machine learning, that detects common facial features. You would then map those facial features to a list of premade images and you would stitch everything together.

I wish you the best of luck. There is a lot of money being thrown at projects like this.


The way you describe is a lot easier said than done. Stitching everything together while still maintaining a coherent and realistic face is very hard. Unfortunately, humans are really good at spotting fake faces as part of our evolution - even small discrepancies in facial features can make them look weird. Naturally, the problem is fundamentally hard and it would need machine learning, like you said, to impose some sort of inductive bias.

I'm not quite sure I fully get the question though. If you want a photo-realistic picture of someone (for an avatar), just take a real picture. As for generative models, there has been a gold rush in research into GANs recently. The empirical results for generating fake realistic faces is very impressive. They end up looking very good because of the very large human datasets. You can always generate new specific faces by just identifying correlations in the modes of the latent space. From this maybe you could find a cool way to generate faces similar to a specific person.

The question is really not clear enough, I'm not sure if you meant to generate a 3D model of somebody, or to just generate a 2D avatar picture (which is what I assumed).

openai.com/blog/glow/
arxiv.org/abs/1809.11096
Users browsing this topic: None
Jump to:


RECENT POSTS

HOT TOPICS