I'm particularly impressed by the way Sora generates crashing waves here. Truly breathtaking stuff.
I fully realize there's no meaningful sense in which these models simulate physics, but it's simultaneously very difficult to deny that their output is a strong approximation of reality (at least visually).
I wonder if anyone seriously starts to model physics by throwing huge amounts of AI compute at AI models, rather than attempting to model reality.
While it is extremely impressive tech, it's somehow a worse demo than the original previews from a few months ago. You can see that most of the common issues are still present too.
There's also the fact that Luma, Runaway, and Pika are available right now while Sora isn't, even though Sora seems to be a step above them, at least from the cherry picked videos they've shown.
I'm actually pretty shocked they released this, it really suffers in quite a few places. The technology is clearly incredible, but I do wonder if they're trying to do too much too soon with this. Are the real applications much more targetted augmentation of existing video rather than these massive generations from scratch.
Transitions are amazing: the lady turning into a jellyfish turned my brain inside out!
There are signs that this is generative ML (legs on the cloud changing positions, weird placement of signs on the car wash), but this is impressively close to perfect.
Would be interesting to see how much compute power this video has required...
It's not safety testing or the election that is keeping sora out of your hands it's the uncanny valley. Why interrupt your competitors when they are making the mistake of publishing repulsive things with their name on it and producing B-grade content that will make your curated stuff look AAA and save the day.
The aloof cloud surfer scene helps suspend some disbelief as your eyes are diverted trying to work out what liquid cloud interactions should be. It isn't really workable unless all your output revolves around brainbusting novel physics. Maybe this works for music videos until it gets overplayed.
The main way things get considered highly produced enough to break boredom is to stretch a trend a little beyond expectation in a pleasing way. Think about how fight scenes have changed considerably with no technological changes but always along a trend. It's tough to define a trend and how acceptable a stretch is.
It takes more than melding scenes into each other to make a final product, there is audience back and forth to consider at every level that perhaps AI alone cannot keep up with.
It is very nauseating for me, just like previous Sora videos. I have watched it from start to end, don't want to see it again. The high speed parts going through the ocean were specially sickening.
Kling (the Chinese Sora basically) is publicly accessible and I don't get that feeling from Kling's videos.
I don’t understand how this is actually an improvement over other ai generated video models/systems.
I mean you get the same weird floaty movement they all have, the same kind of glossy sheen (I don’t know how to describe it exactly but all AI images look like they have like vaseline spread over them whenever they go for photorealistic). Extra limbs, too many fingers, impossible geometry, clothing stretching way too far down the body to make any sense.
The first Sora demo was way more impressive. This honestly seems like a step backwards. Like it got worst and isn’t any better than freely available tools people use to make YouTube videos today.
11 comments
[ 2.8 ms ] story [ 40.4 ms ] threadI fully realize there's no meaningful sense in which these models simulate physics, but it's simultaneously very difficult to deny that their output is a strong approximation of reality (at least visually).
I wonder if anyone seriously starts to model physics by throwing huge amounts of AI compute at AI models, rather than attempting to model reality.
https://www.youtube.com/watch?v=iOWamCtnwTc
There's also the fact that Luma, Runaway, and Pika are available right now while Sora isn't, even though Sora seems to be a step above them, at least from the cherry picked videos they've shown.
There are signs that this is generative ML (legs on the cloud changing positions, weird placement of signs on the car wash), but this is impressively close to perfect.
Would be interesting to see how much compute power this video has required...
The aloof cloud surfer scene helps suspend some disbelief as your eyes are diverted trying to work out what liquid cloud interactions should be. It isn't really workable unless all your output revolves around brainbusting novel physics. Maybe this works for music videos until it gets overplayed.
The main way things get considered highly produced enough to break boredom is to stretch a trend a little beyond expectation in a pleasing way. Think about how fight scenes have changed considerably with no technological changes but always along a trend. It's tough to define a trend and how acceptable a stretch is.
It takes more than melding scenes into each other to make a final product, there is audience back and forth to consider at every level that perhaps AI alone cannot keep up with.
Kling (the Chinese Sora basically) is publicly accessible and I don't get that feeling from Kling's videos.
I mean you get the same weird floaty movement they all have, the same kind of glossy sheen (I don’t know how to describe it exactly but all AI images look like they have like vaseline spread over them whenever they go for photorealistic). Extra limbs, too many fingers, impossible geometry, clothing stretching way too far down the body to make any sense.
The first Sora demo was way more impressive. This honestly seems like a step backwards. Like it got worst and isn’t any better than freely available tools people use to make YouTube videos today.