AI Video Generation is EXPLODING! | New Top Tier Model & News!

AI video is evolving so much more rapidly than I originally thought. I remember seeing COG Video, which was the first AI video generator, and it looked pretty garbage. It was super low resolution, the quality was terrible and it took forever to generate. And now we have a whole slew of competitive AI companies that do just AI video, trying to get the best video generator out there. Anyways, we're diving into the latest advance advancements in the AI video generation space today. So first of all, for those of you who might be new to the AI video generation space, let's just give you a rundown of all of the big players. First up, we have Runway gen 3. This is a pretty decent generator, lots of features, nice ui. It's also got a turbo mode for fast generation. Most people put this one towards the middle, I think, of their tier list of all the video generators. Then we've got Helio AI, also known as Minimax. And this one is known for not having necessarily as many features, but being very high quality, good adherence to whatever prompt you put into it. I mean, you can just see right off the bat, the quality is pretty darn fantastic from this thing. It's very stable, it's very clear. Next up we have ltx, which is more of a platform than a video generator itself. Although video generation is built in, this one focuses much, much more on the UI design and having a bunch of little nuanced features to control your videos and really create something for whatever vision you have in mind. It's the only one that actually allows you to combine multiple videos together to actually get some larger end result. Then we have Cling AI, which is also known for its extremely high quality, but also having a ton of features and being the only one to really have a good motion brush feature. It's worth noting as well that all of these do have paid plans and some of them have, you know, a little bit of a free trial, but it's going to cost you money to generate AI video for the most part. Then there is also vdo, which is another really good video generator. This one lends itself more to 2D animation, but can do 3D realistic stuff quite well. Not as many features as the other one, but this is kind of a fan favorite. If you go to my Discord server, which is linked down below, by the way, you'll find plenty of people who love to use vdo. Now we also have Pika Labs and this one just got an update. So we're actually going to be doing a little bit of hands on testing today with Pikalabs. And finally Luma Dream Machine, which is known for maybe not the best quality in the world, but they're continuously updating it and improving its quality. And it just had an update that makes it possibly the fastest generating AI video in the world. And finally, for my honorable mention here, I do want to talk about Black Forest Labs. This is not out yet, but they are planning on doing open source text to video, which means that this will be completely free to run on any machine anywhere, any computer anywhere, if you own the hardware or if you can get on a website that could run this program for you. Basically it makes it very possible that completely free AI video generation could exist as well as kind of uplifting the whole AI video sphere just by using the latest technology and sharing it for free, essentially showing their steps, showing the process, sharing the code. It's incredible. This is the same company that made Flux 1, if you've heard of that, AI image generator. And then there's also Sora by the legendary OpenAI, although no one has access to this one. And honestly we're not even sure if it's quality at the this point is up to snuff with the rest of all of these generators. The less said about Sora the better at this point. So those are all of the different generators that you can access. I'm gonna link all of them down below. And if you're just skipping to this portion of the video because you already know about all those generators, welcome. Okay, now let's talk about the latest happenings in the AI video sphere. First up, a company that I'm sure you've heard of before, known as Adobe, actually is working on their own video model. They actually have their own image generator as well, known as firef. And it's pretty good. Kind of just your standard run of the mill AI image generator at this point, but it has some interesting UI design that I like. Anyways, they're working on a video model and this is pretty much what we can see from it so far. That's a gif. So it's a little bit low quality, but it looks decent I guess. But this is something that I guarantee. Most of the video generators that we have on the market right now could pull off, you know, just a zoomed in video of putting some glasses on. This gives us a sneak peek at where we'll probably be able to use it first, which looks like it's going to be the Adobe Firefly page, which is their whole generative AI section. You could test out their text to image there right now, but it seems like generative video will be coming there first. And again with Adobe, they actually do have good UI I think for their AI generation models. I like to see this frames per second, 24. Most video generators target about 30 right now, but 24 is more filmic, it's more cinematic. And of course Adobe, working in that creative artistic space, is going to have a completely different idea of what they should do for the ui. And this is another one of those examples. Hopefully they allow you to do things like 60 or 30 FPS video generation. I think that would be pretty cool. Not all creative video is at 24fps, but either way it's interesting to see that they did that. You'll also notice that they have a camera section where you can actually set, you know, shot size camera angle and then they have motion zoom in, zoom out. And actually with the motion stuff, you will see that on some video generators, cling AI comes to mind. And cling AI also has the really cool motion brush feature, but it's not all focused around the camera. And again, this is a part of the different philosophy that these companies might take with the same exact underlying technology. Right, but we've got some interesting motion. Pan left, pan right, tilt up, tilt down. I think that the demo subject here of the Sunflower is a pretty good example as well to kind of give you a pretty decent idea of what's going to happen with your end result video if you select one of these options. But it's nice to see that we're going to have some level of control overall. They've also got a general motion intensity slider. This is something we've seen on video generators. Plenty of them have this, the slider for motion. I would prefer to see something that's a little bit more advanced, like the selection and move tool that CLINGAI has. But for now that's I would say, better than nothing, that's for sure. Gen 3, for example, has no motion intensity tools, which bugs me. You can also do image uploading, which is something that everyone is going to be very happy to see. It's honestly one of the things that has made AI video generation so popular this year is the ability to upload an image you generated in midjourney, ideogram, whatever, and then convert it to a video more or less. And of course you've got your basic prompt, text field, but at this point, you know, they're just running through and showing us all of the different videos that have been generated with this model. It doesn't look too bad. Some things I will note though, fireworks don't look great over here. They're kind of meshing and melding in with each other. In this example though, they show how you can actually change the kind of shot and adjust your video. It looks like this one might be better for more realistic stuff like the snow or the fire or smoke effects, for example. They also have, you know, this demo video where they generate some 2D element and then use it in an animation, which is a cool idea. And then this also gets me pretty excited. It seems to have pretty good ability at spelling and understanding what you're putting into the prompt. So you could do custom name templates, for example, like you would see on my intro to this channel. And then finally they also show it being used in a more professional scenario. So directly with Adobe's tools. I don't personally use Adobe's tools to edit videos, so this is kind of not that useful to me. But I mean if they could integrate this in a deeper way. I mean, picture this. What if you could just prompt the Adobe Video Editor to generate a bunch of clips for you with voiceover text and all edited together? I think that would be pretty cool. I think long term there's definitely some value there for Adobe to look into, but short term, you know, you can just generate on any website and then pull clips in. The examples here are, I'd say mid tier, mid range for what we see from modern video generators. The Chinese ones have been getting very good lately and it's tough to compete with them. So yeah, if you're interested in getting access to this thing, the waitlist will be linked right here and I'm obviously gonna link this page below. But like Sora for example, lock the way for now. So speaking of the Chinese models though, ByteDance, which is the owner of TikTok, I believe Chinese company, has released a new AI video model that we can't access right here in the States. I don't believe yet. And I don't know how many people in China access to this thing, but it looks pretty incredible. Fellow AI video creator, theoretically Media, AKA Tim also just did a video on this one and I think he did a pretty good job. So I'm gonna link that video down below. But either way I still want to talk about the quality of this video generator. Here's the first clip, for example. This is 10 seconds long and if you play the clip we can see a woman becoming sad and then he leans down and looks up at her. First of all, consistent Character all the way. And I can barely tell that this is AI generated. Look at the textures. Look at the patterns on his shirt, for example. Look at the details on his eyes, his eyebrows, the skin and the light, how it reflects as he looks up at her. Not to mention the movement as well. Notice the motion blur when he moves down. This looks like actual cinematic level motion blur that you would get from for having a shutter rate that is twice the speed of the video. Or I guess in film terms, it would be like an 180 degree shutter angle. Film nerds, feel free to correct me. Anyways, finally, as he looks up, we can see tears streaming down her face as she cries. And again, consistent character. It's the same character that we saw in the first scene. The crying teardrops look a little questionable. They look kind of like glue, like hot glue. But other than that, I think the facial expression is really good. Again, the. The jewelry, the clothing, the fur, all the textures, the skin, it looks freaking great. And I gotta say, the color grading on this scene right here, that's freaking impressive. Next up, we've got some crazy surfboarding. And this should showcase some good water physics. I wish we could get this one in a higher resolution. It looks like it's at like 360p, but you can see it's just a few frames. He gets closer. It looks pretty darn good though, overall. Like, look at this thing, the water splashing there. It's in basically slow motion. He gets right up to the camera and it gets blurry for a second because he got closer. So as cameras do they have to auto focus. And this one auto focuses pretty quickly and shows off some of the details on his face. Overall, his face looks pretty good. I gotta say, the teeth, which is consistently pretty difficult for these video generators, don't look bad. It looks like realistic human teeth that this dude might have while he's surfing. And then the camera movement moves a little bit. He puts a thumbs up. His nose warped a little bit there. But again, this is freaking impressive. It's all very cinematic looking. This reminds me of that, you know, original designer girl look from the SORA demos. But this looks super cinematic. And she takes her glasses off. Again, that's something difficult for these generators to get right. Understanding what glasses look like when they're on someone's head and then how it changes when they take them off. And being able to replicate that in a photo realistic, near perfect way, I mean, and then again, the camera changes focus from her to the people in the distance and. And it all works pretty great. Only thing that I see wrong here is obviously the lines are a little bit squiggly, a little bit not perfect, but these people look pretty good. This is another pretty insane example. Again, taking glasses off and walking away. And then of course, you know, this happens. A fine, beautiful work of art that we can see on this image. The curvatures and the nuances in this are just absolutely incredible. I mean, it makes looks like it belongs in a museum. And it's absolutely in the museum. Oh, you thought I was talking about the lady. Get your mind out of the gutter now. But in all seriousness, though, the. The human generation is mighty impressive from this model. Watch the way that she again takes the glasses off. We got good motion blur. She gets up and then walks away doing this anatomy and getting it all right with the hands and the legs and the walking. It's difficult. It's definitely trained on a lot of people. It's trained on a lot of videos of people doing things because that's what people are going to be generating most. Also, again, really nice color grading on this here. We've got an example of an entirely static shot. The other ones had moving cameras where this man just picks up, you know, a tea glass, sips on a little bit of tea in a pretty realistic way. Someone walks behind him and then he puts it down. There's a lot of impressive stuff even in this one. Again, it's a perfectly static background. Nothing is moving, nothing is warping. It all looks nice and clear. He picks up the tea glass. It reacts how you would expect. I mean, you'd be hard pressed, honestly to tell that it's a fully AI generated video. You'd be hard pressed to tell that this isn't real. Drinking from the tea looks pretty decent, but the person going up behind him and being slightly more blurred out and then coming up and actually touching the leather jacket in a realistic way. That is easily the most impressive part. It's combining all of these elements together. Emotions seems to be a huge target for AI video generation models, and it's easy to understand why. I think that the people that are working on AI video generators are pretty well aligned with what the consumer wants in this regard. We definitely saw some decent alignment in that regard with AI image generation, for example, but I think it's a lot easier for the companies developing the video generators to understand that storytelling is a crucial process in developing these models and it needs to be accounted for. So being able to produce emotions like this, for example, is massively important. And Again, it's not the only video generator that's able to do stuff like this. Cling comes to mind. Quite a few of them come to mind. Even Gen 3 by Runway does fantastic emotion replication. But it's nice to see that it's prioritized. This is pretty crazy as well. Apparently this thing has a feature where it does multi lens video. With just a picture and a prompt, you can generate a multi camera video that has the same style overall scene and characters. And you can see it does multiple variations of this wolf, multiple camera angles where it's howling at the moon. You also see some pretty cool horror stuff in this as well. I mean, look at this guy straight out of Dark Souls. And we've got what appears to be Greta Thunberg, absolutely terrified out of her mind as this thing walks over super close to it and yeah, she's screaming for help. Maybe not exactly Greta Thunberg, but just some lady is very scared of this grim reaper thing. I think the emotion here is. Alright, it looks kind of like a fake ish scream, but it looks like something you'd see on Netflix for sure. The most impressive part of this example demo though is definitely the physics on the cloth and how he walks up to her. I mean, that is just absolutely stunning. We can see the individual legs moving how you would expect, and the camera moves and he's just shaking with all these, you know, strings of physics that follow his movements perfectly. We're going to talk a little bit more about physics in relation to AI video generator models in this video, but this is sort of a preface that physics and the understanding of how objects interact in the real world is a super important part of making AI video generators that feel real. And clearly they have it down here. We've got, you know, like some Martian space exploration movie style demo. He's shooting some giant laser gun at some alien thing and it blows up. Looks pretty decent. Again, for the most part, that stuff that I've seen other video generators do. I think that this clip right here is actually showing off maybe some of the weaker sections of the model. Which is good to see though, because you need to see the good and the bad to understand what you're supposed to expect when you get to use it. It's still not terrible though. It's still pretty incredible how far we've come in such a short amount of time. I think it's weird that he misaligns these with his eyeballs and the depth just feels off as a whole with this scene, I think. I mean, at the end of the day, this video generator looks incredible. Probably the best that we have ever laid eyes on. Probably the highest quality, the most well, fully featured apparently. This is launched for enterprise invitation testing first and then launched on Volcano Ark, which I have. I don't know what that is. I assume it's like a Chinese app. It will be launched on Jimang again, probably another AI app to some degree. It's. It's the best video generator I think I've ever seen. In terms of just examples. I'm super excited. I hope we get some sort of a US version or at least this pushes the other companies to step up their game. I've got another quick little tidbit from the AI video generation space that actually relates to CLING AI if you aren't aware. Pretty recently they released a new 1.5 model that does higher resolution and better quality video. But now they've added more features to that brand new model. For example Lip Sync, which does exactly what you think it does. And they also have an API for developers for cling and I'm interested to see if the API is maybe a little bit faster than the Kling website itself. This is just a small little tidbit though. CLING and a lot of these other video generators have been updating and integrating new features and releases very quickly. This is a hot space that is evolving rapidly. I told you we were going to come back around to the physics. This is brought to us by AK here on Twitter. This is a fantastic AI account. Highly recommend. Following this is Phys Gen. It's a rigid body physics grounded image to video generation. It converts a single image and an input condition. Basically force or torque applied to an object and it will produce realistic, physically plausible and temporally consistent video just by the little demo clips that they have here. It's really, really cool. Again, this is something that difficult for video generators right now, but this could be well implemented into something like a video game. You could do full physics based video games with as many objects as you want and it could just be using AI in the background to actually generate what's happening. It is very consistent and I think the examples are pretty awesome. We've got the dominoes falling, we've got the billiards which is really, really cool to watch them spin around. I mean this reminds me of like a billiards game on your phone, but it's crazy. It's all AI generated just because it understands the torque and the trajectory of the physics and what's supposed to happen just by Training like, like basically this is what physics looks like. And it goes, oh, okay, that's what physics looks like. And then it produces accurate physics, which is so crazy it's not actually doing any physics math anyways. We've got like a little can falling over. We've got, you know, the shuffleboard right here with the Olympics. And it's cool because we actually see it go over the Olympics logo and is able to, to restore that fully. That's something that video generators also sometimes struggle with pushing the ball over. And then this one is just, you know, a little wooden block hitting a wall essentially. For now, this is just a research paper, but it could make it into the world for sure. Again, I think this could have huge applications for things like video games, for example, where you want to do physics in already demanding environments. Let's say you can see the various segmentation, you can see the input force and how it reacts to collisions, gravity, friction. It's. It's really interesting. And honestly, that's more or less the reason that I brought it up today. It's a paper, you can go read it. It's not open source as of yet. I hope they open source it though. I also think that this is something that could be applied to even the video generators we might use for other forms of entertainment, like just creating movies or shows. Because understanding physics can probably help elevate overall coherency of models, which was something that was actually noted in the first sora release by OpenAI. So next up, we have much faster generation with Luma Labs full quality Dream Machine 1.6 clips in under 20 seconds, which is really crazy. Apparently this also isn't a turbo or distilled model. It's just uncompromised quality that is much faster. This is also available on an API, which I would love to see some apps integrate because it is so much faster. And honestly, I have never seen video generation happen in 20 seconds. So I really want to test this out for myself. They also claim that this is ten times faster, but Benjamin down here makes a good point. 10 times faster under 20 seconds. He says that he's an API customer and in his testing the API was already generating in about 45 seconds. So probably not 10 times faster, maybe twice as fast. Still in my mind, twice as fast is a pretty big leap. I want to give this thing a quick little run to see how much faster it truly is. I remember this thing being very slow. We'll put a photo of me in there and we'll say, man, takes a Huge bite out of an iPhone. Alright. It's generating. It's generating. Alright. Okay. I mean I would say that this is probably already over 20 seconds of generation time. But as long as it's much faster than it was, which it used to be like minutes. Used to take like 6, 10 minutes to get a video. Oh, there it is. Okay. Well, it generated. How's the quality though? Ooh, yeah. In classic Luma 1.6 fashion, it's not terrible. It's not the worst I've ever seen, but definitely not, you know, near the big boys, which would be like Minimax or cling or even Gen 3. It's nice to see fast generations though. Oh, by the way, Luma also has motion now. Camera motion. So if you just type camera in, for example, it says camera move up, down, push in, pull out, camera push in, zoom into eye to reveal a gorgeous jungle. Alright. Man holds up 4D hyperdimensional test rack to the camera. I guess that's not too bad. I wish he was looking at the camera, but that's not too bad. At least the Generations are lightning quick. That is impressive. It's impressive that they were able to trim that much fat out of the model. Alright, Zooms into the eye and to reveal like a gorgeous jungle. Will it work? No, it just zooms into the eye. But anyways, faster generation. Like to see at Luma Labs, it still would be nice to switch between a slow but high quality model and a fast but still decent model. I like the approach that Runway took with gen 3 for that. And finally, guys, Pika Labs just released 1.5 and Pika was pretty far behind all of the other video generators. But it looks like that might be changing because this looks crispy, this looks very realistic. This looks high quality overall. But again, we got to test it out for ourselves. It's, it's. It's really interesting though. Like they're doing some really wild stuff in this trailer. For example, right off the bat we have first person view, which is difficult for generators to get right a lot of times because most things aren't recorded in first person view. And the quality also looks good. Like this guy looks like a real person. It looks like it's in New York City or some city zooms right into his face. And that's a particular type of zoom as well. Where you start with the lens actually more zoomed in and then you physically move closer to the subject and widen the lens like that is something that I haven't seen done really at all yet with an AI video generator. And they did it pretty good. That looks cool. We've then got these extremely crisp and clean looking shots. And this looks like 1080p video to me. You know, slow motion. He's bringing his foot down, stomping around. We've got this dinosaur rat thing chasing, I don't even know, a raccoon carrying something. But we've got motion blur on the sides. And it looks, looks pretty decent. I guess this plane is zooming around this guy on the skateboard with the chess pieces. That car looks pretty great. Some gorillas dancing, zooming in on the cheetah. We've got this lady looks pretty realistic. I actually really like this shot. And this also, again, it looks like it's 1080p and then it's. It's doing a whip pan effect which looks really good as well. I think in this example they're trying to show off some water physics. And it's just, it's interesting. They're doing, they're trying all kinds of crazy stuff, crazy fashion over here. This one seems to have a shutter rate, but that would make sense because it's so bright in that theoretical room. Although I gotta say, the details on the dress looking a little sloppy maybe. Not bad though. It's pretty high res. Now this looks realistic. Again, these are just like E girls or something like that that are made up by the AI. And I assume this is all image to video because that seems to be what it looks like. They look realistic, especially this one in the center, maybe this one, but this one does not look as realistic. And then these people also not looking super realistic. But that's pretty acceptable if you're just trying to have some fun. Now this is where my mind really started to get blown. Like they're just melting objects now in a realistic way, which is like traditionally very difficult again because of the physics requirements and the understanding of physics that that needs to be held within the model to produce a result like that. That was mind blowing. And then you see this and they're just inflating, you know, a dog. They're inflating all of these different objects, which is just weird. And then also crushing them in a real. And again, that's like an understanding of physics. This is an understanding of physics that I haven't really seen in a flagship video model yet. Or squishing things. And then they squish this cat, which just looks like really scary. And then they also are blowing stuff up, which again, that's traditionally very difficult for these generators. Cutting things Open. It's, it's all very difficult. Let's, let's actually play around with this thing. I want to see what it's got. Let's try to do an image of me biting into a cheeseburger. Oof. Looks like we're getting slow generations right now as well. Big pet peeve of mine with these video generators. Can't wait till this technology just improves rapidly. Let's do another one. We're gonna try to fill up my room with lemons. We'll try to turn my hair into fire. Now what's also really interesting, guys, is with this, they have the new Pika effects and that's how they do the inflating, the melting, the exploding, etc. But it's interesting. I don't know if they train specific models just to do this stuff or how it all works. If it's a pre prompt, I don't quite understand it. But we're gonna try to inflate my dog. Like that dog was inflated. I'm also uploading a nice image of a car here. We're gonna try to melt the car like in the demo. We'll see how that goes. Well, now we've got six generations waiting, so we'll come back to these. Well, guys, I've got some unfortunate news. I never usually like to do this in videos, but the generations are just not loading. It's been over an hour and still nothing. I refresh the page and still the only thing that I got was the melting abilities. Even in the inflation one isn't working right now. So here's what I'm gonna do. I'm gonna look at these two clips and I'm actually gonna save these for the next video that I put out this week. So be on the lookout if you want to see the results for Pika Labs. If we get some pretty mind blowing results from this thing, I might actually do a dedicated video just all about Pika Labs. But as of right now, I'm pretty disappointed in this generation. It looks like their servers are literally imploding. Anyway, anyways, the melting effect overall is pretty cool. It works as intended. And I'm impressed, especially with this car example, the way it melts and it reflects the sky color perfectly in the pool of, you know, molten red car. So I think it's, I think it's pretty cool. This is the one with the pumpkin and some of the melting occurs from the top, which is above the pumpkin, which I'm not a huge fan of. But other than that, it's pretty darn cool. You can also see on the side here, some of the leaves start to melt as well. Anyways, guys, that's gonna do it for today's video. Again, I'm pretty disappointed with the whole Pika Labs situation here. I really wanted to take a look at some of the new generations, and now this video is probably going to go out late because I was waiting so long for these generations to finish up. Usually this doesn't happen, but this is a circumstance in which it did. So I'm sorry I couldn't show you guys more examples. Like I said, though, I'm planning on making it up to you guys in my next video, so be on the lookout for that. Thank you so much for watching, everybody, and I'll see you in the next one. Goodbye.