Let's hold our breath. Those are specifically crafted hand-picked good videos, where there wasn't any requirement but "write a generic prompt and pick something that looks good", with no particular requirements. Which is very different from the actual process where you have a very specific idea and want the machine to make it happen.
DALL-E presentation also looked cool and everyone was stoked about it. Now that we know of its limitations and oddities? YMMV, but I'd say not so much - Stable Diffusion is still the go-to solution. I strongly suspect the same thing with Sora.
The examples are most certainly cherry-picked. But the problem is there are 50 of them. And even if you gave me 24 hour full access to SVD1.1/Pika/Runway (anything out there that I can use), I won't be able to get 5 examples that match these in quality (~temporal consistency/motions/prompt following) and more importantly in the length. Maybe I am overly optimistic, but this seems too good.
Credit to OpenAI for including some videos with failures (extra limbs, etc.). I also wonder how closely any of these videos might match one from the training set. Maybe they chose prompts that lined up pretty closely with a few videos that were already in there.
Lack of quality in the details yes but the fact that characters and scenes depict consistent and real movement and evolution as opposed to the cinemagraph and frame morphing stuff we have had so far is still remarkable!
That particular example seems to have more a "cheap 3d" style to it but the actual synthesis seems on par with the examples. If the prompt had specified a different style it'd have that style instead. This kind of generation isn't like actual animating, "cheap 3d" style and "realistic cinematic" style take roughly the same amount of work to look right.
Sarah is a video sorter, this was her life. She graduated top of her class in film, and all she could find was the monotonous job of selecting videos that looked just real enough.
Until one day, she couldn't believe it. It was her. A video of of her in that very moment sorting. She went to pause the video, but stopped when he doppelganger did the same.
> Stable Diffusion is still the go-to solution. I strongly suspect the same thing with Sora.
Sure, for people who want detailed control with AI-generated video, workflows built around SD + AnimateDiff, Stable Video Diffusion, MotionDiff, etc., are still going to beat Sora for the immediate future, and OpenAI's approach structurally isn't as friendly to developing a broad ecosystem adding power on top of the base models.
OTOH, the basic simple prompt-to-video capacity of Sora now is good enough for some uses, and where detailed control is not essential that space is going to keep expanding -- one question is how much their plans for safety checking (which they state will apply both to the prompt and every frame of output) will cripple this versus alternatives, and how much the regulatory environment will or won't make it possible to compete with that.
> I suspect given equal effort into prompting both, Sora probably provides superior results
Strictly to prompting, probably, just as that is the case with Dall-E 3 vs, say, SDXL.
The thing is, there’s a lot more that you can do than just tweaking prompting with open models, compared to hosted models that offer limited interaction options.
In the past the examples tweeted by OpenAI have been fairly representative of the actual capabilities of the model. i.e. maybe they do two or three generations and pick the best, but they aren't spending a huge amount of effort cherry-picking.
While Sora might be able to generate short 60-90 second videos, how well it would scale with a larger prompt or a longer video remains yet to be seen.
And the general logic of having the model do 90% of the work for you and then you edit what is required might be harder with videos.
Most fictional long-form video (whether live-action movies or cartoons, etc) is composed of many shots, most of them much shorter than 7 seconds, let alone 60.
I think the main factor that will be key to generate a whole movie is being able to pass some reference images of the characters/places/objects so they remain congruent between two generations.
You could already write a whole book in GPT-3 from running a series of one-short-chapter-at-a-time generations and passing the summary/outline of what's happened so far. (I know I did, in a time that feels like ages ago but was just early last year)
> I think the main factor that will be key to generate a whole movie is being able to pass some reference images of the characters/places/objects so they remain congruent between two generations.
I partly agree with this. The congruency however needs to extend to more than 2 generations. If a single scene is composed of multiple shots, then those multiple shots need to be part of the same world the scene is being shot in.
If you check the video with the title `A beautiful homemade video showing the people of Lagos, Nigeria in the year 2056. Shot with a mobile phone camera.` the surroundings do not seem to make sense as the view starts with a market, spirals around a point and then ends with a bridge which does not fit into the market.
If the the different shots generated the model did fit together seamlessly, trying to make the fit together is where the difficulty comes in. However I do not have any experience in video editing, so it's just speculation.
The CGI industry is about to be turned upside down. They charge hundreds of thousands per minute, and it takes them forever to produce the finished product.
I'm almost speechless. I've been keeping an eye on the text-to-video models, and if these example videos are truly indicative of the model, this is an order of magnitude better than anything currently available.
In particular, looking at the video titled "Borneo wildlife on the Kinabatangan River" (number 7 in the third group), the accurate parallax of the tree stood out to me. I'm so curious to learn how this is working.
holy cow, is that the future of gaming? instead of 3D renders it's real-time video generation, complete with audio and music and dialog and intelligent AI conversations and it's a unique experience no one else has ever played. gameplay mechanics could even change on the fly
DLSS is essentially this, isn't it? It uses a low quality render from the game and then increases the fidelity with something very similar to a diffusion model.
Yeah, but I mean who knows why. I know some people can't, my GF is one of them.
I've often wondered if im ok with it because im used to the object on head stuff (like 25 odd years of motorcycle riding/ergo helmet wearing) and close up, high fov coverage fast past gaming? (I play on a 32" maybe 70 cms from the eyes give or take.)
> I am prone to sea sickness. Maybe it is related.
I'd think it might be given my understanding of why illness in many is triggered. It's odd because I never got sick from it, but i've seen others get INCREDIBLY ill in two different ways.
1. My GF tried to use simple locomotion in a game and almost vomited as an immediate reaction
2. A friend who was fine at first, but then randomly started getting very slowly ill over a matter of like an hour, just getting more and more nausea after the fact.
It's unfortunate, because due to lack of bad feelings/nausea/discomfort etc, I love VR. I equally from those around me can see no real path forward for it as it stands today though because of those impacts and limitations.
That being said, maybe they get smaller, lighter, we learn to induce motion sickness less, I dunno. I'm not optimistic.
Even otherwise, and no matter how good the screen and speakers are, a screen and speakers can only be so immersive. People oversell the potential for VR when they describe it as being as good as or better than reality. Nothing less than the Matrix is going to work in that regard.
Yep, once your brain gets over the immediate novelty of VR, it’s very difficult to get back that “Ready Player One” feeling due to the absence of sensory feedback.
If/once they get it working though, society will shift fast.
There’s an XR app called Brink Traveler that’s full of handcrafted photogrammetry recreations of scenic landmarks. On especially gloomy PNW winter days, I’ll lug a heat lamp to my kitchen and let it warm up the tiled stone a bit, put a floor fan on random oscillation, toss on some good headphones, load up a sunny desert location in VR, and just lounge on the warm stone floor for an hour.
My conscious brain “knows” this isn’t real and just visuals alone can’t fool it anymore, but after about 15 minutes of visuals + sensory input matching, it stops caring entirely. I’ve caught myself reflexively squinting at the virtual sun even though my headset doesn’t have HDR.
For games like 2D/3D fighting games where you don't to generate a lot of terrain, the possibilities of randomly generating stages with unique terrain and obstacles is interesting.
The diffusion is almost certainly taking place over some sort of compressed latent, from the visual quirks of the output I suspect that the process of turning that latent into images goes latent -> nerf / splat -> image, not latent -> convolutional decoder -> image
Agreed. It's amazing how much of a head start OpenAI appears to have over everyone else. Even Microsoft who has access to everything OpenAI is doing. Only Microsoft could be given the keys to the kingdom and still not figure out how to open any doors with them.
Microsoft doesn’t have access to OpenAI’s research, this was part of the deal. They only have access to the weights and inference code of production models and even then who has access to that inside MS is extremely gated and only a few employees have access to this based on absolute need to actually run the service.
AI researcher at MSFT barely have more insights about OpenAI than you do reading HN.
No. They have early access. Example: MSFT was using Dall-e Exp (early 3 version) in PUBLIC, since February of 2023.
In the same month, they were also using GPT4 in public - before OpenAI.
And they had access to GPT4 in 2022 (which was when they decided to create Bing Chat, now called Copilot).
All the current GPT4 models at MSFT are also finetuned versions (literally Creative and Precise mode runs different finetuned versions of GPT4). It runs finetuned versions since launch even...
Microsoft said that they could continue OpenAI's research with no slowdown if OpenAI cut them off by hiring all OpenAI's people, so from that statement it sounds like they have access.
Except they keep trying to shove AI into everything they own. CoPilot Studio is an example of how laughably bad at it they are. I honestly don't understand why they don't contract out to OpenAI to help them do some of these integrations.
Every company is trying to shove AI into everything they own. It's what investors currently demand.
OpenAI is likely limited by how fast they are able to scale their hiring. They had 778 FTEs when all the board drama occurred, up 100% YoY. Microsoft has 221,000. It seems difficult to delegate enough headcount to all the exploratory projects of MSFT and it's hard to scale headcount quicker while preserving some semblance of culture.
The only official statement from Micorosft is "While details of our agreement remain confidential, it is important to note that Microsoft does not own any portion of OpenAI and is simply entitled to share of profit distributions," said company spokesman Frank Shaw.
I suspect it's less about being puritanical about violence and nudity in and of themself, and more a blanket ban to make up for the inability to prevent the generation of actually controversial material (nude images of pop stars, violence against politicians, hate speech)
Put like that, it's a bit like the Chumra in Judaism [1]. The fence, or moat, around the law that extends even further than the law itself, to prevent you from accidentally commiting a sin.
I am guessing a movie studio will get different access with controls dropped. Of course, that does mean they need to be VERY careful when editing, and making sure not to release a vagina that appears for 1 or 2 frames when a woman is picking up a cat in some random scene.
We can't do narrative sequences with persistent characters and settings, even with static images.
These video clips just generic stock clips. You cut cut them together to make a sequence of random flashy whatever, but you still can't do storytelling in any conventional sense. We don't appear to be close to being able to use these tools for the hypothetical disruptive use case we worry about.
Nonetheless, The stock video and photo people are in trouble. So long as the details don't matter this stuff is presumably useful.
I wonder how much of it is really "concern for the children" type stuff vs not wanting to deal with fights on what should be allowed and how and to who right now. When film was new towns and states started to make censorship review boards. When mature content became viewable on the web battles (still ongoing) about how much you need to do to prevent minors from accessing it came up. Now useful AI generated content is the new thing and you can avoid this kind of distraction by going this route instead.
I'm not supporting it in any way, I think you should be able to generate and distribute any legal content with the tools, but just giving a possible motive for OpenAI being so conservative whenever it comes to ethics and what they are making.
I've been watching 80s movies recently, and amount of nudity and sex scenes often feels unnecessary. I'm definitely not a prude. I watch porn, I talk about sex with friends, I go to kinky parties sometimes. But it really feels that a lot of movies sacrificed stories to increase sex appeal — and now that people have free and unlimited access to porn, movies can finally be movies.
Where is the training material for this coming from? The only resource I can think of that's broad enough for a general purpose video model is YouTube, but I can't imagine Google would allow a third party to scrape all of YT without putting up a fight.
You can still have a broad dataset and use RLHF to steer it more towards the aesthetic like midjourney and SDXL did through discord feedback. I think there was still some aesthetic selection in the dataset as well but it still included a lot of crap.
The big stand out to me beyond almost any other text video solution is that the video duration is tremendously longer (minute+). Everything else that I've seen can't get beyond 15 to 20 seconds at the absolute maximum.
In terms of following the prompt and generating visually interesting results, I think they're comparable. But the resolution for Sora seems so far ahead.
Worth noting that Google also has Phenaki [0] and VideoPoet [1] and Imagen Video [2]
I know it's Runway (and has all manner of those dream-like AI artifacts) but I like what this person is doing with just a bunch 4 second clips and an awesome soundtrack:
