Hacker News Reader

Comment by popcar2 | Hacker News Reader

Not nearly as impressive as Sora. Sora was impressive because the clips were long and had lots of rapid movement since video models tend to fall apart when the movement isn't easy to predict.

By comparison, the shots here are only a few seconds long and almost all look like slow motion or slow panning shots cherrypicked because they don't have that much movement. Compare that to Sora's videos of people walking in real speed.

The only shot they had that can compare was the cyberpunk video they linked to, and it looks crazy inconsistent. Real shame.

View on HN

> Not nearly as impressive as Sora. Sora was impressive because the clips were long and had lots of rapid movement

The most impressive Sora demo was heavily edited.

https://www.fxguide.com/fxfeatured/actually-using-sora/

Interesting to see that OpenAI was successful in creating their own reality distortion spells, just like Apple's reality distortion field which has fooled many of these commenters here.

It's quite early to race to the conclusion that one is better than the other when not only they are both unreleased, but especially when the demos can be edited, faked or altered to look great for optics and distortion.

EDIT: It appears there is at least one commenter who replied below that is upset with this fact above.

It is OK to cope, but the truth really doesn't care especially when the competition (Google) came out much stronger than expected with their announcements.

Well, as a counterpoint, Apple did become a $2 trillion dollar company...

Distortion is easiest when the products really work. :)

Apple got up to $3 trillion back in 2023.

Indeed, and they’re at 2.87T today… Built largely on differentiated high-margin products, which is not how I would describe OpenAI. I should clarify that I’m a fan of both companies, but the reality is that OpenAI’s business model depends on how well it can commoditize itself.

To Shy Kids credit they made it clear the Sora footage was heavily edited, but OpenAIs site still presents Air Head without that context.

https://www.youtube.com/watch?v=KFzXwBZgB88 (posted the day after the short debuted)

https://openai.com/index/sora-first-impressions (no mention of editing, nor do they link to the above making-of video)

There is now on that second link:

>The videos below were edited by the artists, who creatively integrated Sora into their work, and had the freedom to modify the content Sora generated.

Ha, here's an archive from yesterday for posterity.

https://web.archive.org/web/20240513050023/https://openai.co...

They also just added a link to the making-of video.

If you modified something because it got some attention on HN, at least have the guts to own up to it :/

That's hilarious. Your comment clearly got seen by someone.

I believe it was clear that Air Head was an edited video.

The intention wasn't to show "This is what Sora can generate from start to end" but rather "This is what a video production team can do with Sora instead of shooting their own raw footage."

Maybe not so obvious to others, but for me it was clear from how the other demo videos looked.

> Sora was impressive because the clips were long and had lots of rapid movement

Sora videos ran at 1 beat per second, so everything in the image moved at the same beat and often too slow or too fast to keep the pace.

It is very obvious when you inspect the images and notice that there are keyframes at every whole second mark and everything on the screen suddenly goes in their next animation step.

That really limits the kind of videos you can generate.

So it needs to learn how far each object can travel in 1sec at its natural speed?

It also needs to separate animation steps for different objects so that objects can keep different speeds. It isn't trivial at all to go from having a keyframe for the whole picture to having separate for separate parts, you need to retrain the whole thing from the ground up and the results will be way worse until you figure out a way to train that.

My point is that it isn't obvious at all that Soras way actually is closer to the end goal, it might look better today to have those 1 second beats for every video but where do you go from there?

The best case scenario would probably being able to generate "layers" at a time. That would give more creative control over the outcome, but I have no idea how you would do it.

Objectively speaking (if people would be honest with themselves), both are just decent at best.

I think comparing them now is probably not that useful outside of this AI hype train. Like comparing two children. A lot can happen.

The bigger message I am getting from this is it's clear OpenAI won't have a super AI monopoly.

Comparing two children is a good one. My girlfriend has taken to pointing out when I’m engaging in “punditry”. They're an engineer like I am and we talk about tech all the time, but sometimes I talk about which company is beating which company like it’s a football game, and they call me out for it.

Video models are interesting, and to some extent trying to imagine which company is gonna eat the other’s lunch is kind of interesting, but sometimes that’s all people are interested in and I can see my girlfriend's reasoning for being disinterested in such discussion.

Except that many of the people involved do think of it like a football game, and thus it actually is like one. Of course the researchers and engineers at both OpenAI and Google DeepMind have a sense of rivalry and strive to one up another. They definitely feel like they are in a competition.

> They definitely feel like they are in a competition.

Citation needed?

Although I did not work in AI, I did work at Google X robotics on a robot they often use for AI research.

Maybe some people felt like it was a competition, but I don’t have much reason to believe that feeling is common. AI researchers are literally in collaboration with other people in the field, publishing papers and reading the work of others to learn and build upon it.

> AI researchers are literally in collaboration with other people in the field, publishing papers and reading the work of others to learn and build upon it.

When OpenAI suddenly stopped publishing their stuff I bet that many researchers now started feeling like it started to be a competition.

OpenAI is no longer cooperating, they are just competing. They still haven't said anything about how gpt-4 works.

What would make this "Good?"

I’m fairly certain Google just has a big stack of these in storage but never released, or the moment someone pulls ahead it’s all hands on deck to make the same thing.

Also Sora demos had some really impressive generations featuring _people_. Here we hardly see any people which likely means exactly what you’d guess.

Has Gemini started generated impacted of people again? My trial has ended and I haven’t been following the issue.

Sora is also movement limited to a certain range if you look at the clips closely. Probably something like filtering by some function of optical flow in both cases.

> The shots here [..] almost all look like slow motion or slow panning shots.

I think this is arguably better than the alternative. With slow-mo generated videos, you can always speed them up in editing. It's much harder to take a fast-paced video and slow it down without terrible loss in quality.

A commercially available tool that can turn still images into depth-conscious panning shots is still tremendously impactful across all sorts of industries, especially tourism and hospitality. I’m really excited to see what this can do.

Not just that, but anything with a subject in it felt uncanny valleyish... like that cowboy clip, the gate of the horse stood out as odd and then I gave it some attention . It seems like a camel's gate. And whole thing seems to be hovering, gliding rather than walking. Sora indeed seems to have an advantage

I thought a camel's gait is much closer to two legs moving almost at the same time. Granted, I don't see camels often. Out of curiosity can you explain that more?

Also the horse just looks weird, just like the buildings and peppers.

It's impressive as hell though. Even if it would only be used to extrapolate existing video.

They’re not showing people because that can get hairy quickly.

I imagine thats just a function of how much training data you throw at it.

Could also be the doing of google. if Veo screws up , the weight falls on Alphabet stock. While open AI is not public and doesn't have to worry about anything . Like even if open AI faked some of their AI videos(not saying they did), it wouldn't affect them the way it would affect Veo--> Google-->Alphabet

being cautious often puts a dent in innovation

You mean like how they faked some Gemini stuff?

https://www.bbc.com/news/technology-67650807

Reply on news.ycombinator.com