Public Access to Open AI's Sora Video Generator just Leaked...

Well, it happened folks. OpenAI's Sora finally sort of released, but it's against OpenAI's wishes, you see, SORA released, but it was a leak and apparently from artists protesting against OpenAI also. Huge thanks to the community for keeping me well informed on this one, especially in my Discord server. You guys were on top of this leak as soon as it happened. Take a look at this. This morning a new hugging face space opened up allowing you to generate with Sora for completely free and with an open letter to open AI. Now before we take a look at the letter, take a look at some of the examples that are coming out of this hugging face space. These are all supposedly Sora generations and honestly guys, it's basically confirmed at this point. These are 100% all sort of generations. It's real. It's a real leak. Here's our first little clip. We've got a guy walking up to a portal. Yeah, the isn't going to be the best because of course these are all leaks coming from Twitter. But he's walking through the woods, the camera movement is pretty stable, the walking looks decent and he's you know, moving towards this glowing portal. Nothing too crazy, but we can see if I move my head, there's the OpenAI logo down there. And it also has some sort of a watermark as well, which I believe allows you to verify that it is indeed AI generated video. Here's another example from the public. We've got a graffiti stricken alley with some lights and just a bunch of dudes running down the alleyway. This one's not too impressive. We've got a lot of like duplicate people kind of running in place, but the movement I will say is pretty darn smooth. Overall I like the physics on the shirt as the guy runs. Finally we've got some Minecraft esque gameplay being generated here. And this looks a lot better than any other Minecraft gameplay generated by a typical video generator. So this was probably my first clue when I originally was looking into this, that yeah, this is a real sort of leak. It looks pretty good. We've got the hand movement and you know, you're looking around this, you know, generated world, very Minecrafty. Here is another one with just some city street, some building, some cars. Again all looking decently smooth. It cuts to then this where it's like a corporate room, some pencils. I'm gonna show you guys way more examples and absolutely some more impressive ones because this model actually looks to be pretty darn good. And it's also not the full SORA model. Let's take a look at this open letter though. Dear Corporate AI Overlor with some little guys flipping the bird, we received access to SORA with the promise to be early testers, red teamers and creative partners. However, we believe instead we are being lured into art washing to tell the world that SORA is a useful tool for artists. Obviously the people who leaked this were early testers or red teamers immediately. Also giving by the way OpenAI a nice list of potential suspects. Anyways, it continues on to say artists are not your unpaid rd. We are not your free bug testers PR puppets its training data validation tokens. Hundreds of artists provide unpaid labor through bug testing, feedback and experimental work for the program for a 150 billion valued company. While hundreds contribute for free, a select few will be chosen through a competition to have their SORA created films screened. And this offers minimal compensation which pales in comparison to the substantial PR and marketing value OpenAI receives in turn from the artists. They're saying it's a bad deal denormalize billion dollar brands exploiting artists for unpaid R and D NPR. Furthermore, every output needs to be approved by the OpenAI team before sharing. This early access program appears to be less about creative expression and critique and more about PR and advertisement. Anyways, they're releasing this tool to give everyone an opportunity to experiment with what 300 or so artists were offered a free and unlimited access to this tool. They are not against the use of AI technology as a tool for the arts. Makes sense. If they were, they probably wouldn't have been invited to the program. But what they don't agree with is how this artist program has been rolled out and how the tool is shaping up ahead of a possible public release. They're sharing this with the world with the hopes that OpenAI becomes more open, more artist friendly and supports arts beyond PR stunts. Very, very interesting. I'm sure we're all going to have varying opinions in regards to this and please feel free to share them in the comments and have a discussion down there. I think it's absolutely worth talking about this now. First off, what they did here is absolutely illegal. They can be taken to court for this and under US law penalized. I'm no lawyer, but it can't be good. So very risky move by these leakers. I don't know how they're planning on protecting themselves, but like I said, I mean if only 300 artists were given access here, it should not be too difficult for OpenAI to figure out who the mole is now they do say, you know, artists or not, they're unpaid R and D. And that's a fair criticism to OpenAI. I don't think there's anything wrong with that. The thing is, they very well could have said no as artists. Boycott OpenAI and say, we don't want access to Sora. Keep it for yourself. We're not going to do this for you. But clearly there's enough artists that want access to the tool and will accept whatever contract that is with OpenAI, even if it does offer very minimal compensation. I think a lot of people would argue and say, hey, if you don't want to be a part of OpenAI's PR stunt and marketing stunt, then just say no to OpenAI in the first place, because it does seem to be a very limited few select people that were offered access to SORA ahead of time. That's not to say, though, that there isn't some valid criticism here against OpenAI. It is very much true that OpenAI is not nearly as open as a lot of people would hope they would be, or even as open as they were in the past. And I think that, quote unquote, luring artists into this program to beta test SORA and then share the films made with it for unpaid R and D&PR is a scummy thing to do. I feel like that's a pretty dang reasonable criticism, but is this the right way to go about it? My personal opinion? Absolutely not. If I was in this situation and I was upset at OpenAI, I wouldn't be leaking things because it's against the law and they're probably gonna get caught. Now, with all of that being said, guys, it's pretty hard not to see the truth here. This is hilarious. I think it's so funny that this got leaked. I'm not the one doing it, but I can certainly sit back from afar and laugh. OpenAI, I think, to some degree, kind of deserves this karma. The company name is literally OpenAI. They used to be a nonprofit, but now they're gonna be for profit. No open source releases, not sharing as much of the research, and not being nearly as transparent as other AI companies are. And they've also been hammering down on censorship, all while a huge amount of people have been paying $20 a month for ChatGPT. Plus, Sora was teased earlier this year and it was like this mind blowing thing. And you know, months and months later, almost towards the end of the year, still no SORA access until it was leaked. So do I think it was right? No, I wouldn't do it. Personally, I wouldn't leak something like that. But do I think it's absolutely hilarious and do I think OpenAI kind of is getting a little bit of karma it deserves? Absolutely. Alright, with all that aside, let's talk about the model a little bit more. Let's talk about some of the capabilities and what we've learned from this leak so far. First thing you guys got to know here, this is not just normal Sora. This is a version of Sora that's limited to 10 second generations at 30 FPS and it can generate up to 1080p. But also, this is a more distilled version of the model. It appears to be a Sora Turbo model. So this isn't the big full bad Kahuna. This isn't the big beast Sora. This is just Sora's little brother, Sora Turbo. And this heavily alludes to the fact that there's still a larger Sora model out there under wraps that we have not seen yet. And we don't really know what it's capable of. Considering what Sora Turbo is capable of, the larger Sora Turbo model could be really good. Here is the link to the Hugging Face space in question, the open letter and the generate with Sora. And by the way, this is getting absolutely swamped. So much so that it looks like OpenAI has shut down the early access for all artists now to prevent this leak from getting worse. So yeah, I mean this, the space still exists, but you can't generate the Sora anymore. Still, I'm sure thousands of videos were generated. It also appears that the examples were taken right, right off of Hugging Face as well. And this leads me to believe that OpenAI probably sent a little message to Hugging Face and said, hey, take this down. And Hugging Face just complied for us. This really isn't a huge worry though, because there are so many Twitter posts with Sora demos being leaked. It's basically out there. Here's a nice video of a cat chasing a mouse. This looks pretty good and it kind of does have that slow motion esque aspect to it that you'll see with a lot of AI video generators. But I must say, the consistency of the cat's body in this video is really impressive. We've got the little mouse here. He kind of escapes decently. He's out of frame now, but we've got, you know, the hind legs of the cat, the tail up in the air. It's not misjudging the tail for the foot or the foot for the tail. It's not melding and mushing them together. It's all looking pretty darn good. I mean, there is a weird situation, I think, where you can see over here where, you know, the cat turns his foot and that just doesn't look right. It looks like it detached from his body, but it corrects itself. Very smooth movement, natural and fluid feeling in comparison to some of the other video generators I've seen. As Billowalt points out on here, apparently the OpenAI Discord went nuclear and Sora as a topic has been banned entirely. So, yeah, overall it's just not looking good for OpenAI. It's a rough morning for OpenAI. Here is another example of a baby floating around underwater. And I've got some more examples. These all coming from el.cine. this generation right here isn't all that impressive. I think you could see something around this quality level from Kling or Minimax, maybe even Runway ML. My favorite part though is that this baby is smiling right at the camera. Then some bubbles come up. So this baby definitely farted under the water and now we all have to deal with smelling it. This might be the most impressive example I've seen out of the bunch today. We've got a golden retriever dog. It looks like jumping around in the snow and yeah, it's doing a really good job at keeping him consistent overall, especially with the physics in the trees. And you can also see there's great physics on the dog's body, his tail, his fur, etc. It's very blurry. I know it's hard to see because these are all links from Twitter, but man, that movement is smooth and it really is nice. This one I really don't see too much incorrect with it's good quality. We've got a nice one here of a dude, you know, in some restaurant eating a burger. And this looks alright again, maybe not as good as like cling or Mini Max, but for eating. Those models were literally known for being able to generate good videos of people eating. But Sora does like an okay job. And keep in mind, guys, this is Sora Turbo. So it's not the full model assumedly, but he kind of goes in and takes like a big bite. I like the camera movement here. I think that's kind of fun. In this next little demo video, we just have some strawberries floating in like some pink chocolate or something. I guess it's supposed to kind of be like a chocolate fountain, but it looks like this pink chocolate stuff. Has a mind of its own and is able to just sprout out like that. But the strawberries, you know, slowly rotate. Obviously this is kind of like an homage to the whole strawberry thing we saw earlier this year. So that's pretty cool. We've also got this generation with this little guy in the water there. You can see the ripples coming off of his feet. He also seems to be kind of standing directly on the water. So all I gotta say is this is one holy bird. But yeah, I think this is a decent enough scene. It's not a very challenging scene. Again, something I wouldn't be surprised if I saw a cling or Minimax generate. This right here is probably one of the worst looking gens that I've seen so far come out of Sora. We can see it just sort of pans around his head and then it's another duplicate head on the other side with the arms. So it's not really understanding spatially what the camera is doing in relation to the robot. And again, guys, this is like a turbo model. So this is something that we could expect from a lower quality model anyways. Looks kind of right on par with the rest of what we've seen from the AI video Generation Space up until this point. And the list goes on here we've got a monkey playing with some sort of a cube. But again, he's not really doing too much, not too much movement. It's fluid, it's dynamic, but nothing striking is going on here. That makes me say, hey, you know, I don't think any of the video generators out there wouldn't be able to generate something like this. That's just not the case. I think that they could. We've got a cute little pirate cat though, on top of a little like Roomba or something, like a robot vacuum thing. And that's pretty cute. The cat kind of does some weird twitching though, and like glitching, like, what is that? The model tries to make it look realistic, especially when the cat lands back on its feet. But there's something weird going on there. We have some more videos here coming from Theoretically Media, AKA Tim. You guys already saw this Little City video. We've got another video of a guy eating though, and this time he's just trying to eat spaghetti, you know, similar to the whole Will Smith demo. And it doesn't look great. I've seen better noodle eating videos from Kling for sure. But again, keep in mind, this is just Sora Turbo after all. Look at him. He's trying to Put it in his mouth, but it doesn't really look like he takes a bite. It really seems like these models are better in certain circumstances, depending on what is inside of their training data, such as, you know, video footage of people walking around on the streets. This one's super blurry, but it looks to be a decently high quality video with the Eiffel Tower in the background. This one's pretty useful because it looks like someone went ahead and tried to replicate that original demo where we had that lady walking down in the middle of a city with all the lights. And it did a pretty great job at replicating this. Again though, this is Sora Turbo. But what's really impressive is the ability for the bag to stay consistent. The reflections on the bag here, and also the physics with her clothes as she walks closer towards the camera. This is really impressive. And again, I feel like this right here would be something that's pretty difficult to replicate with all the video generators we have access to now, cling, minimax, Runway, etc. Sora might potentially just be very, very trained on people. Maybe not people eating, but people walking around and doing other things. Like this almost looks like an entirely real video. It does look like it's a little bit in slow motion, but. But I gotta say, that quality is pretty stellar. Smooth, fluid movements, just the reflections on the bag and everything. It's so perfect. Seems like a very boomer bust model here. We've got someone who tried like an anime style with the model and it's kind of looking 3D and creepy and not all that great. But again, the consistency is probably the most impressive part about this model I would say. So far I have another little demo video from testing catalog and we've just got this little rotating galaxy that's pretty cool looking. Nothing too special though to be fair. You could probably just generate an image of a galaxy and then rotate it digitally to achieve the same effect. And finally, guys, I do have this example which seems to be more like a 3D game esque example. This looks really quite good. We've got this fox creature and the camera tilts around. It looks like a scene from a video game or something like that. Right? You can see the 1080p detail is quite good. It's very mushy. Again, all these videos are from Twitter, but I bet that raw 1080p video looks decent. Digging through, trying to find you guys a few more examples. We've got a polar bear with a Coca Cola can and this looks pretty good. I like the little snow particles flying around. It seems to be decently temporally consistent, but is it wildly better than anything we've seen up until this point? Not really. I'm interested to see if OpenAI will put a statement out about this kind of addressing it, confirming hopefully that it's real publicly. I mean, at this point maybe they should just release the Sora Turbo model for us all to test out. They probably don't have the servers to withstand that right now though, since apparently still only 300 or so artists have access to. I would have expected it to be a little bit larger by now, especially since we're coming up to the end of the year and we're kind of expecting a sore release this year, especially since Miramarati said that we would see it released this year. I really want to hear your thoughts though. Do you think that this is good for the community overall, or does it make the AI community look a little bit bad? It's definitely not legal what happened, but we learned a lot about Sora in the very short period of time we did have access to, and it appears that the Turbo variant of Sora is about as good as the rest of the state of the art models we have access to. Nothing too special. The real question is what's behind closed doors? What isn't OpenAI showing us with the larger Sora model? Do they have something that really is a neck breaking, mind blowing, insanely powerful video generation model still remains to be seen. I was going to do a larger news video today, but this seems to have taken up most of the news cycle in time, so I'm going to push that back just a little bit farther. Lots of really cool other updates that aren't around OpenAI Sora getting leaked to show off to you this week. Thanks for watching and I will see you in the next video. Goodbye.