Live data from Hacker News

Meta Movie Gen

ai.meta.com

41–50 of 1001 posts

Re: Meta Movie Gen

#41
Incredible, simply incredible. You know a paper is seminal when all the methods seem obvious in hindsight! Though I’m not caught up on SOTA, so maybe some of this is obvious in normal-sight, too.

RIP Pika and ElevenLabs… tho I guess they always can offer convenience and top tier UX. Still, gotta imagine they’re panicking this morning!

  Upload an image of yourself and transform it into a personalized video. Movie Gen’s cutting-edge model lets you create personalized videos that preserve human identity and motion.
Given how effective the still images of Trump saving people in floodwater and fixing electrical poles have been despite being identifiable as AI if you look closely (or think…), this is going to be nuts. 16 seconds is more than enough to convince people, I’m guessing the average video watch time is much less than that on social media.

Also, YouTube shorts (and whatever Meta’s version is) is about to get even worse, yet also probably more addicting! It would be hard to explain to an alien why we got so unreasonably good at optimal content to keep people scrolling. Imagine an automated YouTube channel running 24/7 A/B experiments for some set of audiences…

Re: Meta Movie Gen

#44
post #34
post #4

Any chance of this being released open weights? Or is the risk of bad PR too high (especially near a US election)? It being 30B gives me hope.

That and the NFSW finetunes that will inevitably follow; unlike the text-gen finetunes these could really cause trouble with deepfakes.

Deepfakes are already a reality, the technology is already there and good enough for harm, the genie is not going back to the bottle.

In fact, the more realistic the deepfakes become, the less harmful actual revenge porn and stolen sex videos can be, because of plausible deniability.

Re: Meta Movie Gen

#45
post #27

Earlier quoted context omitted.

>All the vids have that instantly recognizable GenAI "sheen" That's something that can be fixed in a future release or you can fix it right now with some filters in post in your pipeline.

I think the big blind spot people have with these models is that the release pages only show just the AI output. But anyone competently using these AI tools will be using them in step X of a hundred step creative process. And it's only going to get worse as both the AI tools improve and people find better ways to integrate them into their workflow.

Yeah exactly. Video pipelines that go into productions we only see the end results of have a lot of steps to them beyond just the raw video output/capture. Even Netflix/Hollywood productions without VFX have a lot of retouching and post processing to them.

Re: Meta Movie Gen

#46

Website doesn't work on Firefox and videos don't play on Edge. They should consider asking the AI to make a correct website before having it make hippos swim.

I use Edge at work: the videos played without issue (version 129.0.2792.65 on Windows).

I use Firefox on my personal device: the website worked fine though took an extra "hiccup" to load compared to Edge (version 131.0 on Windows).

Re: Meta Movie Gen

#47

I find the edit video with text the most fascinating aspect. I can see this being used for indie films that doesn’t have a CGI budget. Like the scene with the movie theater, you can film them on lounge chairs first and then edit it to seem like a movie theater.

Why bother? Actors cost money and scheduling is difficult. Do the whole thing in AI - the model will be trained on better actors than your indie cast, anyway.

Re: Meta Movie Gen

#49
post #31

Website doesn't work on Firefox and videos don't play on Edge. They should consider asking the AI to make a correct website before having it make hippos swim.

The entire page load is completely broken on Edge for me. Bizarre

All works fine for me here in Edge, odd.

Re: Meta Movie Gen

#50

Earlier quoted context omitted.

5 years ago there were no AI videos. A bit over a year ago the best AI videos were hilarious hallucinations of Will Smith eating spaghetti. Today we have these realistic videos that are still in the uncanny valley. That's insane progress in the span of a year. Who knows what it will be like in another year. Let'em cook.

Disco Diffusion was a (bad) thing in 2021 that lead to the spaghetti video / weird Burger Kind Ads level quality. But it ran on consumer GPUs / in Jupyter notebook. 2 years ago we had decent video generation for clips 7 months ago we have Sora https://news.ycombinator.com/item?id=39393252 (still silence since then) With these things, like DALL-E 1 and GPT-3, the original release of the game changer often comes ca. 2…

What video generation was decent 2 years ago? Will smith eating spaghetti was barely coherent and clearly broken, and that was March 2023 (https://knowyourmeme.com/memes/ai-will-smith-eating-spaghett...).

And isn’t this model open source…? So we get access to it, like, momentarily? Or did I miss something?

Post reply on HN