Nvidia’s “Kitchen” Keynote was faked (rendered) and nobody noticed
blogs.nvidia.com
blogs.nvidia.com
https://news.ycombinator.com/newsguidelines.html
Clickbait title, it was very clear it was rendered at the time and there was no attempt to hide it.
Please flag the post if it doesn’t fit the guidelines, I didn’t mean it to be clickbaity, my apologies.
I didn't see the full thing, but the posture and lighting in the clips in the making-of don't look like they'd have convinced anyone? (See YouTube video avnigo links).
Is this just marketing saying water isn't wet or am I missing something?
But he was in the kitchen at the beginning of the 2020 keynote: https://youtu.be/bOf2S7OzFEg?list=PLZHnYvH1qtOZ2BSwG4CHmKSVH...
You can spot straight away that he looks like he's in a Zoom meeting with a fake background, he has a green halo behind his hair, and after he says "Our first kitchen keynote." he moves his shoulders and you can see the background behind his shoulders move. (Edit: what the hell, the moving background is visible in the 1080p version, but not really in the 2160p one! So I guess that's YouTube's compressor's fault.)
Since they already built a very detailed model of his kitchen, it would've been more interesting if they had a multi camera setup, or even a camera that moves around, like in the BBC studio https://www.bbc.com/sport/av/olympics/58113457
And the "deepfaked" Jensen seems to me is also very easy to spot...
Or... was he in his real kitchen in 2020, and I'm just seeing things? At the start of his talk in 2021 the kitchen is very out of focus and blurred.
Short version: Upload at 4K, 2mbps x frame rate(30fps x 2mbps = 60mbps), H.265
The kitchen in the background doesn't match the one being disassembled, you see items in a different place and far less realistic lighting in the 3D modeled one. So I think his background is real/a picture, and then they simply cut to the fake one (the version of the 2021 keynote that I see on YouTube is incredibly blurry even at the highest resolution of 1080p that I can select, so if they deepfaked the speaker, I suspect they were hiding that).
Anyway, back to the kitchen... as they said on "The Wire" follow the spatulas. On camera-left there are red, orange and pink spatulas (https://youtu.be/eAn_oiZwUXA?t=1079), they even move (https://youtu.be/eAn_oiZwUXA?t=1990).. presumably inconsistency across multiple takes and editing. Also note what's on the counter to his right. At the point they switch to the virtual kitchen (https://youtu.be/eAn_oiZwUXA?t=6470), we get the "unnecessary blurry zoom effect" (used in TV to save interpolation on eg the living room -> house -> city -> country -> planet zoom out).. now the items to his right are different (there was no American flag in that white dish just now) - there's less items, and the spatulas... well even the modellers refused to make so many, dropping the orange, red, and, pink ones from the container.
Even in simple text, I can relate to you my relief at finding a news article on HN which was posted by my stealth marketing team on behalf of Nvidia today, or I can describe the tiny white kitten curled up on my chest as I type this from the couch I found on the street.
In a way, that expression will create an effect similar to the Nvidia CG project for a lot of people.
Meanwhile, I have no tiny white kitten, I'm not laying on a couch, and I don't work on a stealth marketing team. But some will go as far as visualizing what I must look like, or waxing sentimental about kittens.
Some others would even doubt the retraction of those statements, merely because of a constant inner perceptive mandate which causes them to work extra hard at never getting caught closing off possibilities.
> https://www.youtube.com/watch?v=gT59YOdch8M
> Typically this type of project would take a team months to complete and weeks to render.
If it's naïevely rendered, 66 seconds times 30fps is 1800 frames, 10 minutes rendering per frame, 12 days. This scene is very simple, so I think 10 minutes is a bit much, but in any case it's not an unreasonable statement.
(The scene is so simple I feel that with PBR and some fancy postprocessing you could come very close in real time even without their RTX tech.)
I haven't seen this one and while I am sure many have missed, there's no way "nobody noticed". Of course, this is direct PR material straight from Nvidia. Such material is hard for me to read because at 50 percent is junk words and sentences
What's the closest example in mainstream media of cg human that [almost] passes?
The cinematics in videogames likewise can be pretty convincing, even when the gameplay itself isn’t so much.
Edit:
And then there’s AI-generated faces like these: https://thispersondoesnotexist.com/
No digitally rendered realistic human has ever made it through the uncanny valley. The uncanny valley is to be avoided at all costs, but many digital animators refuse to learn.
Gollum from LOTR worked well because it didn't matter that he was in the uncanny valley. He's supposed to creep you out, it suited his character.
Blade Runner comes awfully close. Compare the original movie footage with the replicated 3D version and it's almost impossible to tell.
https://www.youtube.com/watch?v=nQcxEo5MUc8
I'd say, rendering is basically solved. The problem now is animations. Very hard to replicate all micro expressions and imperfections we have come to accept from humans. Even humans with motion capture rigs seem to be insufficient.
That said, what about the John Connor death in Terminator: Dark Fate? What's uncanny about that?
https://www.youtube.com/watch?v=49YHDbdRP0A
Behind the scenes:
I'm not at all saying this isn't all incredible progress, but it seems like it's almost asymptotic to creating a convincing on screen effect?
Thanks all for the links and suggestions for scenes to check out.
I'd have expected rendering and modelling to attract quite different sets of people.
In general though, I agree it's a waste of time. I'd much rather play Super Hot, for instance.
Although I do play Dwarf Fortress occasionally, so I might not be representative of the average demographic...
Based on the current state of games and pay-to-win mechanics I really wish you were, might see an improvement :)
I have a photogrammetry scan of some guys room as my VR home on Steam and it's genuinely eery how I have to do a double take sometimes.
From Our Kitchen to Yours: NVIDIA Omniverse Changes the Way Industries Collaborate