Making a 20 second AI generated video doesn't take twice as much energy as a 10 second video.
-
Making a 20 second AI generated video doesn't take twice as much energy as a 10 second video. It takes 4 times as much.
This is because the whole video needs to strive for coherence and the size of the data set that determines the next frame keeps expanding like a polynomial.
At least it's not exponential, right?
Could all of the compute in the world even now even make a two hour movie?
-
Making a 20 second AI generated video doesn't take twice as much energy as a 10 second video. It takes 4 times as much.
This is because the whole video needs to strive for coherence and the size of the data set that determines the next frame keeps expanding like a polynomial.
At least it's not exponential, right?
Could all of the compute in the world even now even make a two hour movie?
@futurebird I asked on Bluesky, and was told that gen AI has a problem with coherent physics, which is why it can't make complete films yet. (It may be able to do frame interpolation, but so far I haven't seen artists work with it, instead of managers try to use it to eliminate artists.)
-
Making a 20 second AI generated video doesn't take twice as much energy as a 10 second video. It takes 4 times as much.
This is because the whole video needs to strive for coherence and the size of the data set that determines the next frame keeps expanding like a polynomial.
At least it's not exponential, right?
Could all of the compute in the world even now even make a two hour movie?
This is the same problem I ran into with LLMs. An LLM can't process an entire novel with the care and attention to detail it might manage for a paragraph (and the quality of the paragraphs is already limited)
The exciting thing about machines is how they can automate large tasks. But, making "full sized" media seems to remain out of reach even many years into this exciting revolution.
-
@futurebird I asked on Bluesky, and was told that gen AI has a problem with coherent physics, which is why it can't make complete films yet. (It may be able to do frame interpolation, but so far I haven't seen artists work with it, instead of managers try to use it to eliminate artists.)
The various image generators now being promoted as "AI" do not have any internal representation of physics or geometry at all.
They can never make images or animations that respect those things except incidentally.
All they do is approximate "words to image cube" based on the data fed into them.
(Nor is making an animation that respects physics the same as making a film, of course.)
-
Making a 20 second AI generated video doesn't take twice as much energy as a 10 second video. It takes 4 times as much.
This is because the whole video needs to strive for coherence and the size of the data set that determines the next frame keeps expanding like a polynomial.
At least it's not exponential, right?
Could all of the compute in the world even now even make a two hour movie?
@futurebird actuly.. the resources it DOES take humans to make a movie I might watch 1 or 3 times boggles my mind. I've always thought those efforts shld go to making a school museum dance... kinda thing instead
-
The various image generators now being promoted as "AI" do not have any internal representation of physics or geometry at all.
They can never make images or animations that respect those things except incidentally.
All they do is approximate "words to image cube" based on the data fed into them.
(Nor is making an animation that respects physics the same as making a film, of course.)
And these people keep saying this part, the physics, the structure, the logic will be trivial to just tack on at any moment but it just keeps on NOT happening.
But every 101 level programming class teaches there are three kinds of programming errors:
1. syntax (won’t compile)
2. runtime (won’t run)
3. And? The hardest to fix? Logical.Wow. It can fix the first two. Wow. Amazing.
-
F myrmepropagandist shared this topic
