Nvidia GANverse3D "inflates flat images into realistic 3D models"
Quote:
In a (video) demo, Nvidia shows off a playable recreation of the famous Knight Rider KITT car.
Read more.
Re: Nvidia GANverse3D "inflates flat images into realistic 3D models"
Re: Nvidia GANverse3D "inflates flat images into realistic 3D models"
The 3D KITT looks like they've turned the brightness so far down there's no detail to miss.
In short; thanks Nvidia, I hate it.
Re: Nvidia GANverse3D "inflates flat images into realistic 3D models"
well as a guy that can not draw perspective / 3D it might be interesting.
But yeah look like ray tracing,,,,,, it will probably be better in the 2.0 version next year.
Re: Nvidia GANverse3D "inflates flat images into realistic 3D models"
Quote:
Originally Posted by
Gentle Viking
well as a guy that can not draw perspective / 3D it might be interesting.
But yeah look like ray tracing,,,,,, it will probably be better in the 2.0 version next year.
Indeed, it's something that'll be better with more machine learning and refinement. Interesting and could be great for indie games makers.
If you can't tell, the origional comment was written when my meds had worn off and now I'm all drugged up again and happy.
Re: Nvidia GANverse3D "inflates flat images into realistic 3D models"
Very suspect... you show it the front of a car and it 'knows' what the (out of sight) rear light cluster looks like?
This sets false and really dubious expectations.
Re: Nvidia GANverse3D "inflates flat images into realistic 3D models"
Quote:
Originally Posted by
KultiVator
Very suspect... you show it the front of a car and it 'knows' what the (out of sight) rear light cluster looks like?
This sets false really dubious and expectations.
To be fair, if you're making a model for a game and it doesn't have to be a proper replica, machine learning could do this to an acceptable standard. You're not going to get the back correct, but a large enough dataset could certainly get results suited to the "make 3D model from 2D photo in a single press" market.
Re: Nvidia GANverse3D "inflates flat images into realistic 3D models"
Quote:
Originally Posted by
philehidiot
To be fair, if you're making a model for a game and it doesn't have to be a proper replica, machine learning could do this to an acceptable standard. You're not going to get the back correct, but a large enough dataset could certainly get results suited to the "make 3D model from 2D photo in a single press" market.
Yeah - I wasn't clear in my comment... the video was pretty sparse and made the whole thing seem 'automagic' rather than emphasising the reality of the training needed to get the AI to be able to even recognise a particular object in a photo.
There's definitely uses for this kind of technology - just the real capabilities of it need a lot of ground work to prepare them for a specific task.
Re: Nvidia GANverse3D "inflates flat images into realistic 3D models"
Quote:
Originally Posted by
KultiVator
Yeah - I wasn't clear in my comment... the video was pretty sparse and made the whole thing seem 'automagic' rather than emphasising the reality of the training needed to get the AI to be able to even recognise a particular object in a photo.
There's definitely uses for this kind of technology - just the real capabilities of it need a lot of ground work to prepare them for a specific task.
Part of me wonders if they'll just give it a massive internet trawling dataset to make it automagic?
Re: Nvidia GANverse3D "inflates flat images into realistic 3D models"
Didn't look useable. If they could give it 50 images of kitt they had just scraped off the web that came with different lighting/backgrounds/etc and it was able to use them all to produce a near perfect 3d image then I'd be more impressed.
Even then it would just be a static 3d model, so even better if they could show it the images and then get the AI to watch the backlog of knight rider episodes and it comes up with a perfect 3d kitt with full animation.
Re: Nvidia GANverse3D "inflates flat images into realistic 3D models"
Quote:
Originally Posted by
Dribble
Didn't look useable. If they could give it 50 images of kitt they had just scraped off the web that came with different lighting/backgrounds/etc and it was able to use them all to produce a near perfect 3d image then I'd be more impressed.
Even then it would just be a static 3d model, so even better if they could show it the images and then get the AI to watch the backlog of knight rider episodes and it comes up with a perfect 3d kitt with full animation.
I suspect it's the sort of thing that'll find most use as placeholders for game / graphics development. Something you can use to put a proper 3D model in whilst someone is busy making the actual model without any major time / effort input. But, I also have no idea how games are designed, so that may be total bull.
