New Promising AI Video Generator! VEO 2 Alternative?

Hello everybody. Welcome back to the Matvid Pro AI YouTube channel. I know my hair looks a little bit messy. I just got out of the shower so don't worry, I'm not like sweaty or anything gross. I know how much you guys like to make comments about my hair. At any rate guys, today I'd like to talk about Luma Labs new AI video generation model. It's called Ray 2. There's a lot of buzz floating around online about how this model is pretty darn good. How good, you might ask? Well, I did a quick poll on my X account, which you should follow by the way, along with the Discord serv to get a sentiment analysis for how people generally feel about the three latest AI video generator releases. Google's VO2 Ray 2 by Luma, which we're talking about today and then finally V2 which is going to be a separate episode now. The results of this poll are pretty staggering. Alright, a lot of people prefer VO2 by Google, 85% but still we have that 9% or so that prefers RAY2 Luma's model. And here's the thing, we can't really access Google's VO2 right now anyways. The the model is not open for everybody, it is not public. I don't even have access to it. And while it is super good, there's no telling how long Google is going to keep VO2 under private wraps. So my question for today's video is can ray2act at least in the intermedium before vo2 hopefully becomes released as a substitute? Like I said, it seems pretty darn good. Well, let's take a look on the website here. They're claiming realistic visuals with natural coherent motion and it has a strong understanding of text instructions and can take images and video input as well. So we're starting off with some good capabilities and some pretty substantial claims from Luma Labs. I'm mostly interested in the strong understanding of text instructions. A lot of AI video models can do pretty realistic videos, especially if you're prompting for what's already in the training data. And most of them have been fine tuned well enough for natural coherent motion. It's nailing those difficult prompts that I'm really interested in. Anyways, they've got plenty of examples to have a peek at. Here we've got some natural motion, this guy in the fur coat running away from explosions that already looks like a movie scene to be honest with you, at least at first glance. I like the shaky cam that still keeps the subject relatively in focus. We've got the winding car through autumn painted mountains. This one looks a little bit more surreal to my eyes. There still would be a lot of green and yellow in autumn leaves. But vehicles are traditionally very difficult for AI video generators and that's a pretty good result in terms of the vehicle at the very least. We've got fencing down here as well and that is mighty impressive. There's a strong focus here on human movements that are otherwise pretty difficult for AI video generators to reproduce. And then we've also got the physics of the fencing blades which wiggle and wobble around. And the physics on the blades isn't perfect, but the human movement is pretty close here. We've got abstract ideas. We've got a gorilla surfing on a wave. This is a really, really good example. All of these examples, by the way, are going to be somewhat cherry picked no doubt. Luma is trying to advertise their new model. So these are all the best case scenarios. But yeah, that gorilla looks awesome. I'd love to see a dog or a cat maybe try to surf a wave. I think that would be a little bit difficult because we're dealing with a quadruped. Gorillas are fairly similar to humans in anatomy, which might make it AI model. Next up we've got a humpback whale. What I see as the most impressive portion of this is the fact that the whale isn't really morphing too much. Its fins stay in line where they're supposed to. And I love the particles in the water as well. It's quite beautiful. We've got a tank plowing through some water here. That looks great. Another hyper futuristic sports car. We've got some snowy mountains, a turtle. And I like the insane camera fly through in this particular prompt. That leads to some pretty interesting results. But typically keeping something like this turtle stable and non morphing as it does the insane camera fly through is not easy for video generators. These are examples they provide for impressive instruction following. We've got a snow leopard. I don't think the leopard in particular is insanely impressive because again, videos like this are probably in the training data themselves. But I do love the physics on the fat and the skin of the animal as it crawls around. And I like that we don't have any morphing on the feet or the tail. But cherry picked results of course. Relaxed woman on a motorbike looks pretty good. Tiny kitten, miniature kitten walking and exploring on the surface of a fingertip. That's interesting. That's definitely not in the training data and it Did a pretty great job at producing this. This is the kind of stuff that I'd like to see. These are the more difficult tests for the AI models. And this right here is already pretty good proof that this might make a great substitute for VO2, which, again, none of us have access to, unfortunately. Next up, we've got another thing that's probably not in the training data. We have polar bears relaxing on an iceberg, drinking drinks, wearing nice Hawaii shirts, and they also have sunglasses on. It produced this pretty well. It's a fairly still shot, but I'm a big fan. We've got a sloth swimming underwater. This video footage looks pretty great, but again, probably not something that's in the training data. Ship captain smoking a pipe, turning and looking at a looming storm. I'm sure there's plenty of humans, especially smoking pipes, in the training data. But bringing all of this together, the complete uniform, no morphing as he turns around, blowing the smoke out of his mouth. The fizz in the background with the sails, it's pretty awesome. Next up, cutting meat, something that VO2 is famously good at. It seems to nail pretty decently, at least in the example they show here. We've got a bunch of animals trapped in ice, sculpting some clay, pottery ballerina dancing on a block of ice in the Arctic Ocean. This one is a little bit trippy in my mind. There's a little bit of weirdness going on with the legs and as she spins. Not super, super believable, but it's still did a pretty decent job. Now we're moving on to physics and simulation. We've got a massive orb of water floating in a backlit forest. Definitely not going to be in the training data, because while you can't really do this in real life, it violates physics, but it does a great job of imagining what that would look like. So I like that example. Girl in the red velvet dress floating underwater. Pretty cool. I like the slow motion, the slow physics of the dress. It looks pretty dang good. I mean, a lot of these examples, guys, a lot of these examples are impressing me. I'm not going to lie. This looks like the kind of footage I've been seeing from VO2 floating around Twitter. Again, I don't have access to VO2, unfortunately, but we can access this today. The maple syrup on the pancakes is pretty good, but not perfect. I don't think I've seen a single AI video generator nail a particular test like this with maple syrup or some goopy substance pouring onto something flat There's a little bit of jumping around. It's not quite perfect. Drinking from a bottle of water. Truck driving through the Mexico jungle. Splashing through it again. Wow. The slow motion raindrops are crazy. I bet this is so good specifically because there's a lot of this in the training data. You can film something similar to this on your iPhone. Mukbang video of a man eating a sand castle. Okay, that's the stuff I like. That's not in the training data for sure. He seems to be enjoying that sandcastle. It looks like he's having a good time. Simple campfire, but good fire physics overall. And then finally we get into some photorealism examples. A spider web at sunrise. That's a pretty insane example with the wind blowing through the spider webs. Sculpting clay, painting nails. We've got a subway that's pretty stable for the most part, although the subway looks a little bit long. Seems to be kind of getting all the lines of the doors confused towards the apex of the image. I mean, man, a lot of these are pretty great. A lot of these are pretty great. They're probably cherry picked. As I said though, they've also got some cinematic examples here. I mean, never ending examples, right? We could, we could spend this whole video just looking at examples of this thing. Okay, so like I said, this is available now in Dream Machine. It will also be available in the Luma API soon. And for those of you wondering, the Luma API is going to enable other websites like Korea AI or Polo AI to implement Ray 2, although it's probably going to be more expensive than it is directly on Luma Labs. So here we are in the generation interface. Obviously we're going to go ahead and make sure that we're using the correct model. We'll use Ray 2 at 720p. It looks like we can do up to 10 seconds. I'm going to stick with five seconds for video generation. Right now we've got some good aspect ratios by the way. 16:9, which is standard. That's what this video is right now. A 219 which is way more cinematic. It's widescreen 4 3, which is a little bit older of an aspect ratio and a one by one. Good for sharing across all social media, I suppose. And we've also got 9 by 16 which is your phone's aspect ratio typically. And 3x4 which is a reverse of 4x3. Let's just do a basic text to video prompt. Slow motion cinematic video footage featuring a Shih Tzu which is a breed of dog running and scrambling to reach a tennis ball that is bouncing away from it. Let's send that one off. Let's try another difficult prompt. Ultra close up of an old man. A green water balloon collides with his face and drenches him in water. And by the way, we do have the ability on this website to run multiple generations at once. It looks like it puts us in a queue though, so we're waiting our turn to even start to generate this video. Let's just keep plugging them through. Let's grab a nice starter image of yours truly. Video footage of a man giving a thumbs up. Suddenly from behind, a huge truck breaks through the wall, headed right for the man. Oh no, this is not good guys. Currently Ray 2 only works in describe mode, at least for me on my account. If I try to select Ray 2, it's just grayed out. So we can only do this with Ray 1.6. Well, this is a pretty funny prompt so I'm gonna run it anyways. Still waiting for this prompt to begin to generate. That's three minutes ago. Not liking the speed so far. Let's try some of that. 219 ultra cinematic footage. Cinematic movie footage featuring Egypt in its heyday. City streets are bustling. Vibrant greenery fills the area. A perfect pyramid can be seen in the dist. Send that one off. Oh, and by the way, our Rey 1.6 videos are done here. You can see the truck is kind of just morphing in there in the background and appearing out of thin air and is headed right towards me in both of these clips. These aren't great clips, so hopefully Ray 2 is a lot better when it does get the image to video feature. Interestingly enough, our Egypt prompt here was the first one to get filled out. Maybe it has something to do with the aspect ratio, but we're still waiting on the second example for this one. But the first one, it's pretty good. Definitely only 720p. I'd like to upscale this to get a better cinematic experience from it, but it's a nice smooth pan up showing you, you know, Egypt in its quote unquote heyday. I like the greenery it added. I like the city streets that are bustling. The pyramid in the background definitely is there, so it followed the prompt. But it's, you know, obviously the pyramid that we see today, not the nice clean, beautiful looking pyramid that you would see in EGYP heyday. Still also waiting for the first and second generation. We're going on 13 minutes or so now. Well, if these prompts are going to take some time, I'm gonna put a bunch through right now. I'll do that off camera and when they all generate, we'll go through them together. Alrighty guys, I've been spamming this thing for some time now. We can see our top generations are now out of the queue and I believe are generating. We do have this one video posted online from an iPhone walking through a forest on the east coast and running into a gray alien who looks surprised. I actually like these examples. We've got the gray alien who seems to just be showing the person with the iPhone around. This one's a little bit more blurry. But this clip right here is awesome. I am really super impressed by it. It looks pretty realistic. Obviously the alien kind of looks like they're from a movie or something and there's a little bit of weirdness going on with his feet but. But overall pretty darn good. And obviously we can't make the alien look super, super real. It's good. It's a great generation. Next up We've got a 3D animated clip featuring an anthropomorphic lemon character jumping rope on an iceberg. Did not have high hopes for this one. And yeah, sure enough, it tries its best. Definitely got the 3D anthropomorphic lemon character on the iceberg, but he's not really jumping rope. He's kind of just spazzing around. The jump rope is glitching and looking weird. Not a super great result. Now we've got a macro close up video footage of a human hand squeezing a walnut open and cracking it to reveal a tiny person who was living inside it the whole time. And that person is supposed to look disgruntled. Alright, so definitely, you know, macro video footage. The walnuts already broken in this video and we're not really seeing anything inside it seems to have just prepared itself as well. And then we see, you know, some close up of the walnut. No person, no disgruntled person living in the walnut. Unfortunate. Was hoping to see something better than that. How about this example though? He's kind of cracking the walnut open. He's oh, there, there's like a little toy inside or something. And oh, there's another one. So it's kind of understanding my prompt a little bit, but not super, super close. It's trying its best. Okay. It's trying its best. I'd love to try a prompt like this on VO2 Google give please. Next up, YouTube video showing off a mystical and physics defying 4D Tesseract and it doesn't really give the presentation of a YouTube video necessarily, but it did produce pretty cool looking 4D Tesseract esque visuals in both of the examples. This one looks more like you're kind of transforming through a wormhole or something like that. But 4D Tesseract. I'll give it props. This is pretty cool. TikTok short video. I bought the entire USA. This was kind of just a prompt to see what it would do with something like this. The results are fascinating and creepy, I will say. Just morphing. People kind of like maps of the entire usa. It seems to be trying to do edits and transitions throughout the videos as well. And it's. It's confused. Reytu is confused. Not doing a great job. I mean this guy looks like a cyclops. Why is his. Why is the middle of his face scrunched up like that? Sir, what is going on? Do you need help? I think he just was able to buy the entire USA because people felt so bad for him. The man has no eyes and the center of his face is totally scrunched up. Okay, so now I do want to mention another feature that we have on Luma. Certain parts of portions of the text are circled or outlined and you can change them. So we'll take the alien example which was a good one. Let's switch it to the ocean floor on the Amazon rainforest. It's pretty cool. I like the UI here. I like what they're doing where I can morph, edit and adjust the prompts automatically. This idea of being able to change keywords and prompts was popularized by Google Fun fact. And I believe it's also accessible in VO2's interface. But don't quote me. We can also do a robotic lemon character Gothic fruit avatar. That's pretty cool. And you know what, let's actually just keep the lemon character. Let's say he is corrupt. Karate sparring. Oh, something went wrong. It did not like that. TikTok short video. I bought the entire USA. I bought the moon. Sure it's a pretty. It's a cool feature on the website. Still these top videos have yet to generate. Alright guys, I just had some lunch and I've come back to see that we are still waiting on generations. Unfortunately these first two just don't seem to be generating at all. It seems like they got out of the queue and then maybe stuck in an error. So the system that they have for the website right now is still having issues. I don't know if it's bugs. I don't know if they just have too many people in the queue. Regardless, not all my generations are going through. I'm just letting you guys know that that is something that might happen to you. Even if you go off and purchase one of these plans, the lowest entry level price to try and actually use ray 2 is 10 bucks a month. And once you start getting into paid plans, there really isn't much of an excuse for having generations not go through Luma. If you're watching, this is something I would absolutely fix asap. I know not everyone is immune. Pretty much no one's immune from having server issues, especially with new model launches. Even Sora from OpenAI had issues at launch, but I continuously am seeing this problem so it's more of a warning to all companies. Make sure your infrastructure is set up correctly to handle the load of a new model launch, at least to the best of your expectations. To have generations not go through at a paid plan level is just unacceptable. Since we're having some issues, I'm going to take a look at the community, see what the community is generating with this AI, and we'll get a basic sentiment analysis for how people are feeling right now. We're looking at the Twitter replies to the Luma announcement post for this model. Nathan Bowie here can confirm it's working. He's got a pretty funny generation here. Three identical twins holding the largest cheeseburgers I've ever seen. I think the people look good for the most part. I mean, like I said, they're all identical, which isn't very realistic. And the cheeseburger also are holding their own for the most part, although I would expect them to fall apart, especially as this guy all the way on the right basically tilts it sideways. The background's pretty nice and the butterflies are okay. One user restructuring democracy brings up can it do consistent characters? And that is a huge question. One of the largest things in my opinion, holding back the AI video generation space as a whole is the whole issue of consistent characters. We haven't quite cracked the IT with image generation yet, so how could we expect to crack it with video generation? There are some companies, vdu, cling, poloai that are trying to do consistent characters, but none of them have really nailed it. We don't have a perfect use case that does faces and clothing consistently throughout multiple clips. The first video generation company or research lab to crack good consistent characters is going to have a huge advantage in my opinion. Hopefully it's something that we see get delivered to us from the AI overlords. This year, DaVinci Digital here used Luma Labs Ray 2 to create a last of us style clip or series of clips. It's about two minutes long by the way. Stuff like this is going to be linked down below if you want to take a good look at it for yourself. We're dealing with Twitter compression by the way here, so bear that in mind. Pretty interesting sounds here. Got the zebras. It's pretty cool. The more cinematic stuff with people, we got nice smooth camera movements. I mean, that seems to be what it's mainly trained on. Right. A lot of the tests that I were performing were a little bit more dynamic. I think the crux of this issue here, where we have, you know, good cinematics, smooth camera movement, nice colors, especially good people, is all a result specifically of the fact that a lot of companies are fine tuning and optimizing their models to be smaller and better for generalized use cases such as more cinematic things. So it might not be great in these other areas like the areas that I was testing myself. The problem is you probably could build a larger video generation model that can do it all right. It can handle those more complex, weird prompts like I was testing today and the stuff like this. The issue lies within the cost to generate videos and how long it's going to take to generate videos from a larger, more comprehensive model. So that's sort of where we're at. From what I've seen, no model is immune from that. As compute improves though, and as fine tuning improves and architecture, we'll start to see this stuff diminish. Though some folks mentioned that the motion is getting a little bit smoother. Leo asks if they allow image for reference. I wasn't able to access that particular feature. Anna points out that she just paid to test it and the waiting line keeps increasing. Yep, that's the same issue that I was having and I ended up with generations that never even went through. Huge problem here. Future AI World is also testing Ray 2 from Luma Labs. He's got some fires going on in Los Angeles, which is something that is really happening in Los Angeles, unfortunately. So not super tasteful generations, but I suppose examples from the model nevertheless. Ofaleko here on Twitter is Also Test driving Ray 2. Showing off spacecrafts and sci fi things that that you would see in the movies. Leans more on Ru's strengths rather than its weaknesses. I like the movement. I do like the cinematic aspect, Obviously. A little interstellar inspiration. Wow. We've got another little montage of Ray 2 from Nigel. Like I said all these are linked down below if you want to take a closer look at them. All of these examples really that I'm seeing floating around from other people are utilizing the more cinematic aspects. I mean, we have some more dynamic tests, at least in this one, but for the most part it really seems like Ray 2 is focused on cinematic storytelling and trying to elevate the model's ability to understand prompts specifically for that use case. I mean, it's doing some cool stuff. It's not a bad model by any means, by any stretch of the imagination. It's absolutely state of the art, it's competitive, it's probably better than OpenAI's Sora. Alright guys, so let's draw some conclusions, some pros and cons. First of all, let's talk about some of the things that I like from this release. I like the focus on cinematic things. A lot of people like to generate more cinematic looking movie esque TV show esque visuals, human characters, doing things that you've probably seen in a lot of other TV shows. The model excels in this area, might be better than a lot of other generators. You're getting 720p video footage, you get up to 10 seconds, it's 30 frames per second. And hey, the model is actually available today, which is nice. The paid plans start pretty low at only 10 bucks a month, which isn't too bad. And I like the innovations that they're making with the UI where you can scroll down, change keywords in your prompts and then regenerate easily. Let's talk about some of the cons I wasn't able to get all of my generations through and that is not acceptable, especially for having paid plans. More dynamic and strange prompts that push the model do not end up super good. This model seems to be highly optimized, highly fine tuned for specific styles and specific use cases, which might be the most popular use cases at the moment. And as I stressed earlier in the video, this is more of an issue related to compute costs, the market as a whole, and the fact that AI video generation just isn't all there yet to be able to do these more complex things. Is it a Google VO2 killer, so to speak? Absolutely not. It appears to be worse than that model, but that model isn't widely accessible, which kind of puts us as consumers in a little bit of a tough place. Listen, AI is a rapidly evolving space. Things are going to look completely different in like six months. So buckle in and don't worry too much about where things are at this particular moment is what I would say is my recommendation and all. And honestly, the more I use AI video generation tools, the more I explore these websites, the more I listen to you guys who are also actively testing this stuff out with me, the more I realize that a lot of these companies who are making AI video generation platforms are advertising them as maybe a little bit more complete and well rounded than they actually are. This technology is so, so brand new. It's been less than a year since we got state of the art modern day AI video generation even presented to us and way less than a year since we've got access. At the end of the day, Ray 2 is something that is going to be a lot of fun to play around with. It's a very nice look into certain use cases, but we're waiting on this technology to get a lot better. I want to be able to make my lemons jumping rope on icebergs. I want to make a very realistic style video where a tiktoker buys the moon and I want my generations to go through. And I think most of all we want consistent characters because it's really difficult to tell a good story without consistent looking characters. Again, not trying to be negative here. This space is moving extremely rapid. So I'm excited to see what cooks up in 2025. Thank you so much folks for watching. I'll see you in the video. More to come. Like I said, I'm going to be experimenting with VDU as well. There's other AI updates I think. I'm doing a whole roundup of general AI news tomorrow for you guys. Stick around. I'll see you in the next video.