One year ago, Sora 1 redefined what was possible with moving images. Today, we're announcing the Sora app, powered by the allnew Sora 2. [Music] It's the most powerful imagination engine ever built. and it's packed with new features. I'll pass it to Bill for more details. [Music] Now, every video comes with sound. Sora 2 is also the state-of-the-art for motion, physics, IQ, and body mechanics, marking a giant leap forward in realism. And we're introducing Cameo, giving you the power to step into any world or scene and letting your friends cast you and theirs. [Music] On the path to AGI, the gains aren't just about productivity. It's about creating new possibilities. It's also about creativity and joy. 1 2 3 4 [Applause] That's why we're launching Sora 2 inside the Sora app, allowing everyone to push the limits of their imagination and create in ways we never thought possible. [Applause] Welcome back to reality. I'm Bill. I'm the head of Sora. I'm Rohan. I lead the Sora product team. I'm Thomas. I lead Sora engineering. Back in February 2024, we introduced Sora 1. We really view that internally as being the GPT-1 moment for video generation. It was the first moment video really felt like it was starting to work and simple behaviors like object permanence started to emerge from scaling up pre-training. Since then, the Sora research team has been hard at work delivering the next step function change and model capability. And we're super pumped to show you guys Sora 2 today. Sora 2 is our flagship video and audio generation system. And you just saw a taste of what it can do. The first thing you'll notice when you get your hands on this model is how much smarter it is at physical interactions than any prior video generation system. In the past, really complex dynamics like an Olympics gymnastics routine or maybe doing a backflip on a wakeboard were really tough. Sora 2 is much more robust at handling these types of complex collisions and modeling dynamics in a way that feels extremely natural. The team's also done a lot of work to improve the steerability of Sora 2 relative to prior models. Oftent times you have to use video generation systems in kind of a shotby-shot manner. It's really tough to get out a longer narrative that contains multiple shots all in the same generation. Sora 2 is a lot better at this and can tell longer, more coherent stories all in one go. Of course, the big feature here is audio generation. This is the first Sora model that simultaneously generates both video along with audio. And it's a very general purpose system. You can generate dialogue in a variety of languages spanning multiple speakers. You can generate sound effects and even soundscapes. We're really excited for you guys to get your hands on this model, but there's one feature above all else that we're really pumped about. It's a new feature called Cameo and it's unique to Sora 2. The way it works is by observing a short clip of say me, Rohan, or Thomas. You can then take that individual and insert them into any Sora generated environment. You just saw a few examples with me and Sam in the previous video, but this is a very general purpose capability that emerges from our world simulation models. The way it works is that ultimately by just observing any clip of not even a human, but even a pet or an object, the model understands it really deeply and then can inject it into any prompt as if it were just another text token. We're really excited about getting you guys access to this model. But what we really want to show is what we're doing on the product side to really capture all of the magic of this model. You know, in the early days as we were developing these features, Sora researchers really felt like this was a new kind of communication. What originally started out as like text messages and then moved to emojis or voice notes really felt like it was progressing into a new video-based medium with this Cameo feature. And it became really clear to us over time that we needed to develop a new product surface to really capture all these amazing capabilities of the model and to really get this into as many people's hands as possible. Rohan and Thomas have been doing a lot of awesome work here and I'll hand it over to them to tell you a little bit more. Great. So, I know everyone's eager to see the app. Before we go in there, just a little bit of play setting. So, what you're going to see is a very familiar interface if you've used social media before. Uh there's a notion of identity. Uh you have a profile. You can follow other people that you're connected with. Um, but all the content inside of it is going to be AI generated. It's not posted by bots. It's posted by humans, but it's all AI generated. And it has this very, very interesting feel that's uh quite different than basically anything else I've used. It really does feel like a new medium. Um, when you see the feed, you're going to see all the fun we've been having on the Sora team. Uh, some of memes have been emerging uh as we play with the product. Uh, there's always the constant need for GPUs to keep up with the increasing demand. Uh there's one about ketchup. I drinking ketchup for some reason which I think it's based on a true story I've yet to understand, but it's there. Uh and of course there's some fun ones about perfume and other things that stretch the model in different ways. But Rohan, why don't you just give it a Yeah, let's jump into the app. All right, I'm going to click the Sora app here and we're going to be dropped into a feed. Got it. Sora is pre-revenue. If you show revenue, people will ask how much and it will never be enough. The company that was the 100x or the thousand Xer is suddenly the 2x dog. But if you have no revenue, you can say you're pre-revenue. The app is indeed pre-revenue. So, there's a couple cool things to note here. So, this is an example of the cameo feature I was talking about before, but actually this is two cameos in one. So, this is me talking with Sam in the same scene. And you'll notice lots of little details here that make these videos feel really realistic. These little shot changes back and forth, the natural gestures and facial expressions on both me and Sam's face. The natural lip sync that really captures the dialogue accurately. All of this is brand new in Sora 2. All right, let's keep going. Okay, watch this. I'm going to turn the lights off. Whoa, wait a second. Why am I a cartoon? That wasn't supposed to happen. The lights are still on. This is kind of cool, though. I really love this one. I think the the dynamic range of Sora 2 is incredible. A lot of prior models out there seem to sort of collapse into a single aesthetic, and Sora has such a wide diverse range, uh, which is amazing. I can't wait for the the creativity of the internet to to have access to this. Let's keep going. Back on the news again, the man from last month can't stop eating McDonald's ketchup. Straight from the Look, it's not about the ketchup. It's about the experience. Health experts are concerned. Back on the news again, the man from last here. He feels alive like a painting come to life. Every breed carries a story. Stay close and we might hear him. Let's see. So, this cameo feature is really general. Like I was saying, you can use it on humans, but you can also use it on pets. This is actually my real dog, Rocket. Uh, render it in an anime style. As Rohan was saying, this is a really general model in terms of the stylistic range that it can produce. Uh, and it can cover anything from realism to anime and everything in between. Awesome. I'm feeling inspired. I think we should run a generation. So, at the bottom of the screen, you'll see this plus button. I'm going to click there, and we'll be dropped into our simple composer. Here, you can describe any any idea you have in any style, uh, any scene, transcripts, all that kind of stuff, and you can get a video. Um, you'll see this tray above cameos and you'll see me here, Rohan all the way on the left and some friends I have on the network here who've given me permission to cameo them. Um, let's run a generation with a cameo. Maybe someone most people know. Let's do Sam. Uh, Sam, what do you think? He's got to be celebrating how well this live stream celebrating how amazing this live stream is going. He's screaming about it. pumping his fists and screaming. All right, we'll fire that off. Much like SO one, these generations can take a couple minutes. So, while that's happening in the background, I'm going to walk you through our Cameo feature in a little bit more detail. You're probably wondering, how do you set this up? What do the permissions look like? How do we keep this safe? Um, so let's jump in here. It's my profile. I'm going to click edit Cameo. In this screen, you'll see a number of Cameo settings. Before we jump into the settings, I want to talk about how you actually upload your came. So, I'm going to click retake here. Um, so in this flow, you'll be asked to record a dynamic audio prompt. Um, so, you know, we'll present you with some random audio challenge. Then there'll be a livveness check. You'll be asked to move your head in certain directions and then that'll get sent to our systems where we do tons of validation basically to ensure that no one's impersonating you and um, this is indeed you on the network. Once you've done that and your cameo is approved, uh you can set who can use this cameo. You can decide only I can use my cameo, people I approve, mutuals, everyone. You are in full control of your likeness on this network. There is no way for someone to generate you without you having given explicit permission and gone through this cameo flow which is a very important principle for us. Um a couple other notes, you can guide the model on how you want to be portrayed. The model is amazing but it's not perfect. Sometimes it might hallucinate things, might give me skinny jeans or a weird accent or something like that. So, I can go into Cameo preferences and sort of tune this as I run generations, which you know, I suggest everyone who sets this up does. Um, and we'll be adding things like advanced flows to give you more control over this. Um, but in the meantime, we have a couple of ways of doing this already. You can also use this for a lot of fun, which is uh we've been having fun on the team where we give ourselves kind of funny hats or weird things that you do. Rowan always has a gold chain which you'll see later but you use those instructions to sort of guide the model in just different fun ways. Um another thing that is really important to us is the idea of ownership over your own identity. So everything that uh is created with your Cameo when you've authorized somebody to give with permission you have full rights over in the sense that you can delete it. Uh you're treated like an owner of that video. Yeah, exactly. Um cool. Let's go back to the feed. See a couple more gems. Sora 2, the new fragrance from Sora. Fresh, clean, and unapologetic. I don't use perfume, but if it were Sora themed, I would consider it. Possibility. One of my favorite features of this app and this model. And I think something that is uniquely like made possible with this technology is the ability to immediately participate in a trend, in a storyline, in some lore that you know, some creator is is working on a universe via this remix feature. So, I see this. I'm feeling inspired. I want to see my own variation of this. I can just click this remix button here. New fragrance from I can click the remix button here and say, make this an ad. Oh, make this an ad for any ideas. Top hat. A top hat for a top hat with giant feathers. Nice. Okay. Um, and boom, I fire off that generation. Uh, and Sora will go working on my contribution here. Um, in the meantime, let's actually see other remixes of this perfume. Sora 2, the new toothpaste from Sora. Fresh, clean, and unapologetic for whoever you choose to be. Sora 2, the smile of possibility. Smile of possibility. I don't speak Korean IRL, but in Sora, anything is possible. All right, let's keep going in their feet. Guys, check my kick flip. This is our coworker Minia doing a kick flip. Uh, this is incredible physics. I haven't seen anything like this with any other video generation models. I've been trying to do this myself for about 20 years. Still working on my kickflip. But yeah, just in an incredible display of physics by the model here. The dream. Championship point for Rohan. Yes. Um, I want to thank the haters for for fueling me. You can see my gold chain in there. And if Thomas were to make a gen, he would unsuspectingly get a gold chain of me in there, which is kind of the fun of this feature a little bit. The dance ridiculous. Ladies and gentlemen, make some noise. And finally, it's a really good range. Download the Sora app. We'll tell you how soon. Uh, cool. Let's check on our generations, but they might still be going, but Oh, we got Sam's. All right, this live stream is going so well. Let's go. I'm fired up. It's amazing. We're crushing it. Come on. Thank you all. This is a great Hey, thanks. Thanks. All right, I think the other one's still generating. So, while that's going, I'm going to hand it back to Thomas to talk a little bit more about our philosophy on this app. So, thanks, Ron. Uh, I want to admit that when we were initially going through this project, we weren't really sure that this would be something that we wanted to go through as a company and and commit to. Um, we were all a little bit skeptical of the idea of having an AI generated feed and what that would feel like and whether you would lose touch with like actual human connections. Uh, I think once we've started using this Cameo feature, it really does feel different. It feels like a new medium, like a new way of connecting with your friends in a way that I was even surprised by. It's a very different mode of operation when I'm scrolling a feed and I'm thinking through like, oh, what could I do to riff on that just a little bit or wait, can I put myself in that video? Um, it just does feel very very different. And so I'm very very happy with the way this is turning out on on the team with the notion of like connections. Um, one thing we're know we've noticed over time is a lot of social media uh in general has moved away from the idea of friends and kind of family connections. Um, we believe that Sora can lean into this because it's just so easy to create. It's so easy to create in a way that wasn't possible before. And with that in the feed, we're going to be heavily prioritizing connected content. You also have a following feed always available where you can see just connected content alone. And we also have some new features that gives you some agency in the way you control your feed. So there's a beta feature on the top of feed that we'll be working on uh where you can select uh the type of content you want to see. So if you're in like say a relaxing mood, you can say I I want to be relaxed. Animals. Animals. We always have fun with that. Seeing only cute animals. Um and so you can guide this the model to show you content that is really aligned with what you want to do at that time. Um we're also going to be heavily optimizing the feed to encourage you to be creative, to inspire you in a way that's not just about scrolling the feed. Awesome. I think our Jen is done. Let's look at that. The new top hat with giant feathers, bold, elegant, and unapologetic for whoever you choose to be. Plume, the hat of possibility. I would buy cool. Um, yeah, and I'll talk a little bit more about how we're sort of approaching safety and moderation on this on this network. Obviously, like Thomas said, we've been, you know, pleasantly surprised internally. We were skeptics of a just a purely AI generated feed, but have felt this human connection. Um, we felt like this was the best form factor, but we wanted to make sure we amplified the good of form factors like this and, you know, mitigated the bad that often comes with sort of short form video. Um, so there's a couple things we have in place here. One is for U8s, we have a separate set of policies for U8s. There's no infinite scroll by default. There'll be a stopping period with a cool down um, pretty quickly in your experience. Even for adults, we'll nudge you sort of later on in your scrolling process if you're, you know, if we if we think you're in sort of a doom scroll loop, we'll nudge you to creation since we think that is like a fun thing and usually feels good on this app. Um, another thing that's really important is that we want this content to be clearly labeled AI generated when it's ever off our platform. So, we have several provenence techniques. First and foremost, things are visibly um watermarked when you export them off our app. So if these things are floating around other networks, you'll see the Sora animation there. Um we also have some techniques internally to always trace back generations we see to Sora if we see things floating around the internet. And we have C2PA as well. And then we're working on top on top of all the amazing moderation that's come with Sora 1 and image gen. We have reasoning models under the hood that make sure it's very difficult to create harmful content on this network. obviously extremely important in the Cameo feature that no one can create, you know, X-rated or violent content and that is uh the case with all the sort of guard bells we put in place, which is amazing. Um, obviously we're, you know, we're starting a little conservative with our moderation here. Uh, you might see overblocking. We're sorry in advance. Memes about that. Yeah, people are memeing us internally for overblopping blocking. So, we're finding this balance, you know, of user freedom and people who who might be, you know, trying to do bad things on network. And so we'll be working on that over time. Um lastly, before I hand it back to Bill, I want to talk about um a couple of other services that we'll be deploying. Sora one, sora.com, our existing web app will get this new model. There's a little bit of a facelift you'll see, but we'll also still have some awesome features like storyboard launching soon that might come in a week or so that lets you really control like shot by shot um how the model creates a scene. Like Bill mentioned, there's so much controllability and power in this model. We really want to invest in amazing creator tools so you can create amazing content for our network. Um, and we'll be launching an API in the coming weeks as well. There's a long tale of use cases where people can do amazing things where we might not want to build like fine grain editing controls, but others might. People might want to integrate this into their own video editors. And now that's possible with Sora 2. Um, yeah. Want to talk about how how we're rolling this out? Yeah. So today's the day you'll be able to download the Sora iOS app in the App Store starting later this afternoon. Uh we're starting only on iOS. The team is hard at work trying to get an Android version up and running, but bear with us. Uh we're launching initially in the US and Canada and we're doing an invite- based rollout. So like we've said, we think it's really important that you come into this app with your friends. This is really best experienced in a social way almost as a new form of messaging. And so when you get off the wait list and you'll get notified of that when you download the app and get a push notification, then you'll get four invite codes automatically that you can use to give to your friends to make sure they come in with you. We're super excited for you guys to get your hands on these models. You know, we started the Sora research program back in early 2023 to really build AI systems that deeply understand the physical world. We think that's going to be a paramount capability in order to get to truly generalist AGI. Along the way, we're training a lot of models that we think the world can have a great deal of fun with and can bring a lot of joy. So, we're really excited to see what you'll ultimately create on this app. We'll see you guys on Sora. [Music] Ready [Music] when you are. 3 2 1 go. Dude, you all right? I'm good. Steady. Damn it.