Meta’s new text-to-video AI generator is like DALL-E for video | Hacker News Reader

Meta’s new text-to-video AI generator is like DALL-E for video

theverge.com

103 pointshardmaru28 comments

That is extremely impressive progress. Lots of people were toying with using Stable Diffusion for video, but the results, while inspirational, were quite far away from looking like anything but a psychedelic trip. People were guessing we might see coherent video in a few years. Turns out it was a few weeks.

I think the Singularity is little more than a new age religion, but it’s moments like this that give me pause.

The actual page about 'make-a-video` is here:

https://makeavideo.studio/

...and has a lot more examples on it, and the paper.

I particularly liked the dancing bear thing https://makeavideo.studio/assets/similarvid1a.webp

Those videos of the astronaut looking at you and slowly moving toward you are fucking terrifying when you remember an AI created the whole thing

I watched it, I don't find it even remotely terrifying? It's a composition of a bunch of input from real photos?

I agree -- astronaut not terrifying. However, the hand on that teddy bear... :shiver:

I would not describe what Dall.E does as a comp of real photos, so I assume that's also not what is happening here.

It would be like saying an original piece of artwork from a human artist is just a comp of a bunch of input images they've seen. Technically true, but not usually how the word "composition" is used.

More like the re creation of images from the millions of learned patterns.

All of the examples look pretty bad to me for a "finished product" - maybe it's because the Stable Diffusion revolution has taken off so fast.

Looks messy, uncanny and low res.

Is there anything adjacent to Stable Diffusion in the works for video?

Dall-e 1 felt surprisingly similar just a few years ago. It's the first iteration and it will get better before we even know it.

Risky click there :)

The pace media generating AI is advancing at is insane. At this rate the tech is going to be mature before a sizeable part of the population even knows it's a thing.

I have a hunch this will become a widespread story to the general public when election season gets into full swing

I’ve always thought that if the AGI singularity was real, it would occur so quickly that no one would really know what happened. The last few years have only strengthened my view on that.

Things are indeed moving really fast now, and that's really exciting, but I'd be cautious about extrapolating that pace too far into the future. Nearly all of the recent image generating AIs are all based on a single breakthrough technique called image diffusion[1]. Once we start hitting the limits of diffusion models progress will likely start to plateau a bit, at least until the next major breakthrough.

[1]: https://www.assemblyai.com/blog/diffusion-models-for-machine...

There is something very 'meta' (the adjective) about Meta facilitating fake generated videos which will no doubt be propagated all over Meta's platform.

Outside of quality improvement, it would be interesting to have a way to persist a theme. For example, a way to save the teddy bear as a character that can be used in other videos. Or a way to save a version of “large medieval town with a grassy town square” as the set for a movie.

Research paper: https://makeavideo.studio/Make-A-Video.pdf

Examples: https://make-a-video.github.io/

Demo site: https://makeavideo.studio/

I am told live demo and open model are on the way.

What's with hacker news and having negative reactions to 100% of everything Meta does?

I don't agree with a lot of Meta's practices, but it's not like every single thing they do is purely evil... Every once in a while, something they do is cool and I think people are being dishonest if people don't recognize that.

It's like reading a news that a drug cartel built a hospital and then someone goes on commenting "what's with people having negative reactions to 100% drug cartel does? I don't agree with a lot of what drug cartels are doing, but it's not like every single thing they do is purely evil".

Come on, man.

I feel in a similar way.

Meta being evil to the point of profiting from political manipulation, doxxing, addiction, massive extraction of attention, and genocide, is already established.

In related news, in this time of severe drought, the Mexican cartels have expanded their business to stealing and selling water.

> it's not like every single thing they do is purely evil

No, but they've lost the benefit of doubt. So in the absence of evidence it isn't evil, they're presumed to be so. Facebook pays well. But that's the cost of that compensation, a cost borne in the courts of public opinion and law, as well as in politics.

Meta does cool things. But knowing that these cool things will be controlled by Meta is not a cause for celebration.

Most people are just meme posters.

Meta: This is evil. Zuck is a robot. Everyone laugh at the metaverse

Google: Wow when are they going to cancel this

Amazon: I've lost confidence in buying anything from amazon

Apple: Apple is too corporate since Cook took over, I hate the app store monopoly

These companies are functioning like whales with pet chihuahuas on cocaine with a trillion dollar war-chest. I might try to make a prompt and see what the AI gives me back.

That they still exist and making money hand over fist should be considered a wonderful, quirky oddity of human fatuousness. I don't think they should be protected.

>Most people are just meme posters.

>Meta: This is evil. Zuck is a robot. Everyone laugh at the metaverse

>Google: Wow when are they going to cancel this

>Amazon: I've lost confidence in buying anything from amazon

>Apple: Apple is too corporate since Cook took over, I hate the app store monopoly

Because it's all accurate*!

You can try and be all hand-wavy about it, but the fact of the matter is these companies do a lot of harm and I will take advantage of every opportunity to warn others. They've long since lost the benefit of the doubt.

* Excluding Apple. I don't have much experience with them or their platform(s) to make an informed argument either way.

why the non-sense? you didn't read the comments?

What are the next steps to look at from here?

- Creating a grammar to 'direct' the video

-- Blocklist for inappropriate words or phrasing through an Independent advisory body

- Optimising the output to synthesise a coherent video

- Add rudimentary sound output for the video

- Ultimately, the next frontier is building virtual simulations, ie. World Building <> Metaverse

Reply on news.ycombinator.com