Google's New Video AI puts SORA to Shame...

Guys, I am losing my mind over this brand new Google VO2AI video generator. It is so good. The quality is above anything I've ever seen before. There are reputable community members that have access to VO2. We've seen the tests and it just flat out it's the best model I think I've ever seen for video generation pretty much all around. Let me show you what I'm talking about. Here is Google's webpage for VO2. It is their state of the art video generation model, as you guessed. And right off the bat, take a look at some of these examples. We've got a woman in a yellow lab coat peering into a microscope. I can't really notice any details that signal to me that that is AI. I mean, I'm sure if you looked close at the microscope you could tell, but that's a pretty complex object that the video generator kind of just nails and I get it. Yes, we are on Google's own project page. So these are cherry picked examples, but I do have community made ones that are not cherry picked. Plenty of those to off today. Next up, you can see a dog diving into water wearing the goggles. And it's so temporally consistent, meaning consistent over time, over the course of the video. There's no morphing with the dog's face. The ears have great physics. Those blue goggles stay exactly where you expect them on the dog's face. There's just nothing really all that weird about the video generation. Oddly, it just looks kind of pristine. Now we get into some more interesting stuff. Like we've got this animated esque girl working her magic in the stars or something like that. You know, it's very ethereal, not necessarily realistic, but pretty darn good. The only thing I think with this bottom generation that clues me into, hey, this might be AI generated is the way that the morphing lines from the stars are kind of blending in or attaching to this girl's hands. It looks kind of cool, but that's something that you oftentimes see in AI generated video. I sort of like motion blend like stringy attachments with moving entities in the video. If I'm explaining that well enough at all. Down here, you know, we've got some nice flamingos. Again you. This one's pretty small. You can't really tell anything's wrong with it. And then finally we've got some animation over here. Good hair physics and overall super clean. The background is rock solid. No weird movement at all. And these aren't even the Most impressive examples, by the way. They say realistic motion. They say super high quality output up to 4K resolution, which is pretty dang insane. Not something you really see from a lot of AI video generators. And I mean, Sora from OpenAI just released publicly, and that does up to 1080p. This is four times that resolution. And they also promise extensive camera controls. We've got another pretty standout example, a nice close up shot of this woman jamming out to some music. We've got some nice lights in front of her. Overall, it's a pretty decent composition. The physics are crazy impressive. Again, if you showed me this like a year ago, I would have said you were crazy to claim that it's AI generated. I mean, it's so perfect. They do give us some nice YouTube links for all of these individual examples up to 4K. This video isn't 4K itself, but I have a 2K monitor, so I can do my best to extrapolate for you guys the quality levels here in 4K. Yeah, this is just an 8 second clip. Again, that little animated girl talking. If I pause it here and we zoom in, there is definitely quite a lot of detail going on in the clothes, but definitely not as much detail as like a real 4K render of 3D animation. It looks a lot like an upscaled 1080p, which is pretty much what it is from what I can gather. Still though, adding decent detail in the right areas. The skin is nice and smooth, while the clothes have more roughness to them, which you expect. I think where you mostly notice the loss of detail from upscaling is in the hair. It is absolutely not natively in quotes, generating at 4K. We've got another video demo here where we have some pancakes and some syrup being draped over them, and then a close of a mug of coffee, I assume being filled up. Now you'll notice that the syrup for the most part looks pretty darn good. It dribbles off the side of the pancakes quite realistically. You'll notice the most motion issues though, right in the center here with the syrup kind of piling up. It's getting close, but it's still a little bit blurry. The bacon looks quite delectable. And then you can see there's some weird crackling in the background, at least in the beginning of the video. Some weird little motion artifacts, but for the most part very clean. And obviously the coffee being poured. We've got lots of physics, little bubbles going on. Again, still definitely looks like upscaled bubbles. Not necessarily super, super Consistent like this over here. To me, sort of looks like this over sharpened, over smoothed out type of deal. But other parts look a lot better. Like the extra details on the side of the glass came out fantastic. It really depends on the specific application of what's going on in the video for those high fidelity details to come through or not. We've also got a nice video of a beekeeper. This is a particularly complex test because we have all those bees swarming around and that's difficult for the AI to recreate. And while it's nowhere near perfect, it does a pretty decent job. We get the idea there's bees swarming everywhere, he's got some honey and he's a beekeeper. Point out how realistic the lighting looks, how realistic the guy looks with his clothing, holding the things. I mean, his hands have all the fingers that we would expect. The sunflowers look pretty darn good in the background. They're not mushing together too much. They look like individual sunflower stalks. It's pretty fabulous. They claim here that VO2 outperforms other leading video generation models based on human evaluations of its performance. And I mean, this example right here, cars and automotive stuff, is particularly difficult for these AI video generators to get right. And man, that is very, very consistent and very, very good looking over time. You typically don't see motion this good with vehicles, especially in such a complex area like a city with AI video generators today. That is very, very much impressive stuff. I mean, it's kicking up smoke everywhere. It's drifting around what appears to be New York City. And the headlights as it turns, by the way, you don't see them anymore and they come back around and they still look the same. So impressive. Again, we've got that same video with the microscope and we actually have prompts here, very complex ones. By the way, she's running through like this fractal, I don't even know. Miniature world blending hand drawn 2D animation with dynamic light effects. What a crazy big prompt. So they did a benchmark. They showed participants over a thousand prompts. And most of the time people preferred veo. There still was a decent amount of tying, which you can see with the white bars. But it appears 50 to 60% of the time VO is going to be the better option. And keep in mind these comparisons were all done at 720p. I mean, you have to think these high quality examples, they're very, very good. I think trying to push the model as high as 4k is introducing quite a lot of artifacting, which Honestly would look better if this was just in 1080p straight up. I mean, you'll see it in the water and it's distracting. But then you look at the dog and you look at the motion of the tail swinging back and forth and you realize how stable this thing is. You look at the pink flamingo float and you see it slowly curve as the dog is moving it with its paws. And it makes sense in a physics way. It feels almost as if this scene here with this dog floating in the swimming pool really did happen. But spoiler alert, it's all AI generated. I mean, this stuff, guys, is really on another level of video generation. It's not perfect, but it's closer, I think, than we've ever seen from another generator before. It really is, I think the fact that objects don't morph, skew and change over time, at least for the most part, some of that clearly still is an issue. Take a look at this video right here with the skater doing some fancy tricks. You can see that the board goes all wacko fifth dimension on him every time he goes up in the air. And sometimes he also flips around as well. It's not very good at keeping him consistent, so there is still some motion issues. The camera movement is looking very smooth though. The park as a whole looks pretty consistent over time as it turns and flips around. But the skater himself, the subject here, is absolutely struggling. Now, before I dive into some of the things that the community has discovered about the models, some of those less cherry picked results, let's talk about actually getting access to VO2. Basically, it's a wait list right now, and it's the same wait list as VO1, which was a long time ago. I'll link this down below. You got to click sign up to try vofx. Log in with your Google account and when you're brought back here, if you're not in the app to actually use VO2, then you still don't have access. It also doesn't hurt to go up here and click the Join waitlist button, which will basically allow you to fill out this Google form to get access. But again, if you already filled this out for V01, you should still be on that wait list in terms of getting access. I still don't have access. Most people I know still don't have access and it seems to also be country limited. But I do know two people in the United States on my Discord server that have access access right now. So it's possible. It does also seem like if you do have access and are a part of this Early VO2 Access program, you have to have like regular meetings with Google to talk about what you've been generating and some of the issues you might have been seeing. So I think selections are sort of based on that and obviously how well integrated and ingrained in the AI space you're in as a whole. I mean, I don't have access, so it's, it's pretty hard to get in it seems. I also asked those folks in the Discord server how long it took to generate videos for a tenant 1080p clip. It seems to take only about three minutes or so to get an eight second length video, which is pretty darn good. It's quick and it's snappy and it's definitely very high quality. Alright. Taking a look at some things from the community, Joseph Carlson here brings us a comparison between Google VO2 and OpenAI Sora. Keep in mind this isn't the full Sora model, but it is the one that the public has access to. This is Sora Turbo. The prompt here is a pair of hands skillfully slicing a ripe tomato on a wooden cutting board. First we'll start off looking at the Sora clip here. We can see that we've got the hands there and cutting the tomato, but it doesn't actually get chopped in half and fall down onto the cutting board. He's just sort of massaging through the tomato with the knife and you can see like we don't even see the stem until he removes and picks up the knife, which doesn't look great. And then he actually starts to cut through his own fingers and the knife is kind of just sliding through everything. It doesn't look the best. I think Sora could probably do a better gen than this, maybe with a different prompt or even just giving it another go. Either way, not a great generation. Here is the one that was produced with Google VO2. I mean it's a night and day difference though. No matter how poorly cherry picked, you could say the Sora generation is and how cherry picked this generation is. I mean it's like a no contest. We've got the tomato slicing and falling into place. There's like even the juices and the slight like wobble with the tomato. It's like, is really crazy. This one, I can barely tell is AI generated. This one's like, oh, obviously that's AI generated. You also have to think here that Google does own YouTube and as Davon points out in the comments down here, Google Basically has direct access to all of that video data. And OpenAI is going to be struggling to get that data in comparison. We also have other people in the comments of that tweet testing the same prompt with Sora themselves. And you can see like this, for example, is a slightly better generation. I mean, it's not perfect. He's like cutting like half of the tomato up. But it might be a little bit more plausible, better than at least the last generation we saw. Still nowhere near as good as the VO2 gen. Here's an 8B quantization of 1 Yun video running on only an RTX 4090. So this is possible on local hardware. It's still expensive consumer hardware, but consumer hardware nonetheless. And that is even, in my opinion, a better tomato cut than what we saw from the OpenAI Sora video. So maybe Sora just isn't great at those cooking tutorial style things, but still, I think Google's video is better than this. I've got some other fun experiments here. This one comes from Shlomy Fruster Fructor. Sorry if I pronounced that wrong. All of these Twitter links are going to be down below, by the way. Starting off here with some blueberries getting dropped into the water. And again, this is just like a flawless generation. It's really shocking to me how good this is. Honestly, it's kind of just borderline frightening. We start off here with already some blueberries kind of being dropped in, but they keep going and they splash in the water exactly as you would expect. I can't think of anything that should be happening here in reality that this isn't portraying. We even have little air bubbles attached to the blueberries, which is like exactly what you would expect from this sort of thing. And the way they slowly like float up to the surface, I mean, dang, dude, that's good enough to be in a. In a commercial. I would say this is pretty easy to recreate with a professional camera though, but you can generate it in three minutes with Google vo. We've also got some strawberries now, and these ones are more in slow motion and also more close up, like a really detailed macro shot. And again, there's nothing in this video that stands out to me as saying, hey, this shouldn't be here, or hey, this is missing, that you would normally see if you recorded this in real life. I don't want to say it's flawless because it's not flawless. There's a few little, you know, specs that might be a questionable, I guess, but it's It's. It's good enough where I wish I could just pluck one of those strawberries out and take a bite. You know, we've got an apple just being flunked down into the water, and this one does have an issue. I will say. When the apple is first dropped down, it looks like the bottom of the apple has some weird malformed looking stuff. I mean, apples can get malformed from time to time, but when it kind of turns, it looks a little bit sharp. And that is just, I think, a little bit weird. But quickly it rotates and then you don't even see that anymore and it kind of just floats there in the water. As you would expect. The way that the water calms down at the top is so crazy. Not to mention the reflections of the apple on the surface there, the little water bubbles, the particles. It's that extra detail, that extra mile that Google VO produces that makes you want it so bad. In comparison to the other video generators, we've got more specific stuff like green cherry tomatoes. And again, you're kind of probably starting to get sick of this. I like that the cherry tomatoes float, though, where the apple. Don't know how realistic that is. Though I'm pretty sure the apple is supposed to float. Yeah. As Plasmo says down here, shut up and take my money right now, Google. If they launch this with a $200 a month unlimited plan, I think it's a lot easier to justify that in comparison to Sora, for example. Next up, I've got a few more demos. This one comes from MBZ here on Twitter. This is a sitcom TV show about potatoes, and it seems to do pretty good. Went for that 3D animated aesthetic. Lots of great details on the potato. You know, he's a chef, clearly, and as he moves around, it looks very reasonable. It cuts to this lady. Her clothing looks pretty good. I mean, it's an anthropomorphic potato. That's not easy to do. I would know. I do a lot of anthropomorphic fruits and vegetables. The glasses are decent as well. They're not morphing or malforming. But this one, we're missing glasses on this side, which is. Isn't great to see. So that is absolutely an error. Also this. This one's eye is doing something a little frightening in the back, but yeah, I mean, sitcom about potatoes. It's pretty good. It's pretty darn good. I like the hands especially. We've got another muscle car drifting around a corner. Again, we already talked about how this is Impressive. But the license plate here does say VO and it keeps that consistent the whole time. One thing that's interesting to me is the lights turning on and off on the vehicle. Like it's about to just click on there. That's something that obviously you're going to want to be keeping consistent throughout your. Your shot of the car. I don't think any video generator, to be honest, is ready to generate decent car scenes. I mean, I think little short clips to work into a story just to like say, hey, these people are on the road or something. And then it cuts to when they're in the car is fine. Long driving scenes where multiple things are happening. I don't think the video generators are just there yet, but I like the foggy background. I like the camera movement. Veo is trained very on a good source of data, which is, I mean, it's YouTube, let's be honest. We've got a high energy music video with these dancing animated potatoes here. This is crazy. Kind of looks like a scene from Sausage Party or something like that. Until you get to the very end and they have like human legs and they're doing human dance moves. Like that is just wild to see. But it's. It's really intriguing to me how this model is able to take concepts like that. We've got 3D animation combined with more realistic physics, with all of their head pieces moving around and then realistic dancing legs and arms. It's not something that you see a lot in 3D animation. Right. This is like its own style almost. It's taking these multiple concepts and stringing them together. It's AI generated no matter how much it looks on the surface like 3D animation. And again, we've got more potato stuff. This guy is absolutely obsessed with the potatoes. Kind of like me with the lemons. But you know, we've got him dancing around some streets here. He's got this black leather jacket on that stays consistent impressively. And he's, you know, running through the streets while it's raining. I like the reflect on his body, but his mouth does get, I think, a little screwed up, at least towards the end. Yeah, he gets like a second mouth, which is a little strange. Pretty typical for video generators. But I'm impressed that it only happened right at the end there. Nearly perfect. Alright, we'll go through more here. We've got the potato putting makeup on in front of the mirror. We've got the monocle on, the top hat too, which is just so cool. I really like this one. Just like this Tiny potato in front of a huge mirror. That's exactly what you'd expect. This one's just sprinting for the Olympics here in the water now. Ceremony of the largest potato award, which I love very much. I love this silly, goofy stuff that's always the best to see. This guy is diving into a pool full of potatoes, and you can see it is kind of a pool with water in it. But look at all those potatoes just bouncing up, all hitting each other with seemingly correct physics. This guy really does like the potatoes. You've even got anime potatoes down here here. People walking up a mountain to a statue of a potato. I mean, you can create your own potato universe. I think that's sort of what we're trying to go about. The documentary is crazy. Like this close up shot of this guy. If you didn't have this hat with the words and the logo, you would have no idea that this person doesn't even exist. I'll link all this down below. Potato powered by Veo. You can see it absolutely can do logos and text pretty well. It's a versatile model for sure. This is especially really cool. Here's a cat looking into the mirror at his reflection, and it's a lion. And it actually syncs up perfectly with the cat. So it's like his reflection is the lion. That's how he sees himself. To me, this is just a very impressive amount of prompt understanding. It's a deep coherence from the text to what you expect in the end product. Video AI is just so crazy. Finally, to wrap this up, we're going to take a look at my own Direct Discord server, which I really suggest you join if you want to stay up to date in AI and also see stuff like you're about to see when it happens. One of my moderators, Noah, has been taking prompts for some users in my Discord server. And here's an example of that. Intricate organic shapes, flowing lines that evoke depth and texture. Abstract composition, Sophisticated and artistic. It apparently also is giving you four videos at a time, which is really, really cool. I mean, this is highly detailed and it matches the prompt pretty well. But these lines, I think down here are kind of a giveaway of AI generatedness. Straight lines and keeping them that way across the video can be pretty tough for video generators. Overall though, man, it's pretty darn cool. Look at the way that the light sort of reflects and turns off of everything. I mean, you can do some really crazy stuff. Like I said, that is a diverse latent space inside of this model, we've got another user, Roro, who has access. We've got this guy screaming in a box pool at McDonald's. So this is where you're gonna start to see some more of those fun generations, some of those guilty pleasure type of generations where you're just doing silly, wacky and crazy stuff. Here's another one. He's got the burger and it's like, you know, this GoPro camera view. Is he gonna take a bite of the burger? Probably not. He's just kind of screaming inside of the McDonald's. But yeah, you can see eight seconds is a pretty decent time for these generations. And overall I'm happy with it. I mean, this is awesome. This one is a little bit more professional looking, but yeah, he's just screaming in this little baby McDonald's and no one seems to care. They've also been testing it on things like Minecraft and you can see it is very, very good with the Minecraft stuff. There's plenty of Minecraft content on YouTube and it is trained quite well on it. I mean, even most of the textures are looking almost identically correct. Obviously when the motion starts to get a little bit crazy, things can start to get a little bit more screwed up. But overall, guys, not bad, right? It's pretty cool. We've got a YouTuber right here opening up their Hogwarts letter, which is pretty craz. You know, just sort of talking about it like a vlogger or something. You can see the nice setup in the background. One of my Discord server members, Roro, made this quick little mini movie, I guess you could call it. It's literally just 50 seconds long, but it's kind of just like this going to space vlog, if you will. So take a look at this.

Why don't you turn around and let, let them get a view from there and see what the field of view looks like backing into the cable. Okay.

Oh, we're at the moon.

Well, look, look at that. We can, we can go over, over there. It might be. Look at that.

Chris.

Chris, we need you back. Oh my gosh.

Kind of a crazy ending, but you can already see some of the creative possibilities blossoming when you have higher coherency generation like this. I don't think it took him too long to get good clips overlooking the earth. Good clips using your phone to take pictures seemingly just out in space wearing nothing, which is pretty funny, you know, finally getting on the moon, actually, you know, having someone run out into the moon itself, I think that looks pretty decent. And especially the end here, you know, with this crazy black hole, like, you could tell some really interesting stories quite easily. These are 3D watercolor animations and text. Check this out. Look at how correct the clocks are in this video. I mean, it's not just an image, it's a video. And all of the clocks seem to have correct times 12, 1, 2, 3, 4. I mean, it's counting up correctly. That is insane. Something you don't even really see with image generation. And now we're seeing it with Google's VO2 and video generation in this crazy, surreal watercolor style. Right? Here's some more of that just so you guys can check out these demos. Super cool stuff. Ooh. And it appears that VO2 does have the ability to do some more censored things like blood or gore. We could see the spaceman has clearly had a bad time vlogging on the moon. So that's, that's not too great. Was able to do two of these generations. I don't know if maybe my users were just good enough to bypass the restrictions that were in place for Google VO2. But the detail here is great. I mean, that looks like a movie scene. That looks like something out of Interstellar right there. Man in car footage of rhinos on the highway. They're kind of morphing into each other, but nevertheless pretty cool to see. Obviously something you'd never see in real life. And it did a. An okay job. Right here you can see some zombie esque movie clips. You know, the typical things that you would see. Walking Dead or anything cinematic with the zombies running around ready to go munch on some brains. Right? It does a good job. Stable camera movements, stable people running around through these barren wastelands. Lots of detail. I just have to say I don't have a lot of complaints when it comes to this model. It really, really is impressing me. Obviously it's not perfect, but is it the best I've seen? I've seen a lot of clips, a lot of examples, and I have to say it has the least amount of morphing. It's the best temporally over time. Great amounts of detail and resolution, realism. I mean, it really is all there and understands complex things. It's doing copyrighted works right here. Here's a guy reviewing the Mario on the toilet. I mean, look at that. That's kind of wildly good. So specific. It understands things so well. You can even see the Funko pops on the wall. This is the type of guy that would, you know, have something like this. Apparently so wild. And it goes way deeper. I mean there's so many videos for you guys to check out on my Discord server and of course linked down below on Twitter and such. It can create video games that don't exist. It can do so much. It's that YouTub data I think at the end of the day that is making it so good. But for the generation time of only three minutes to get four clips, man, that is nuts. I love these POV cam things with the GoPro esque footage. You can even see it says Police spelled correctly on it. I mean we're really getting to some high fidelity levels here guys. I want to get my hands on this thing so darn bad. But with how the wait list has gone so far, it doesn't seem like I'm getting access anytime soon. Hopefully early January or something like that. They'll release this in some capacity to us to compete with Sora, but I don't know to give you guys some final information on this, it does seem to have multiple modes like image input as well as just straight up text to video. It really is a mind blowingly good model. I think it's the best video generator I've ever seen. I don't have access to it, but I've seen a lot of non cherry picked clips for the most part and it's doing things that I have never seen before. It's been a while since we've seen an uptick quite like this. There's a lot of like minor little improvements up the slope of AI video quality. But this was like a pretty big uptick, man. Definitely I think better than Sora, which is crazy because it's coming from Google. Just this, you know, quiet little drop after they released Sora, which was super, super hyped up. I don't know, I think Google's feeling pretty smug right now about their recent releases because I just did a video on Google's Gemini 2.0 which was also kind of a knock out of the park. If you will. Let me know what you think down in the comments below. Let's have a discussion about this. I'm excited either way. I really want to get access to this thing. See you in the next time guys and have a good one.