Runway Gen 4 AI Video is Blowing My Mind! First Impressions

Hello all you wonderful people and welcome Back to the Matvid Pro AI YouTube channel. Today we're checking out a new AI video generator by RunwayML and they are a top competitor in this AI video generation race. So they're not to be snuffed at. I think this thing might blow some minds today. Yep, that's right. Runway ML has released Gen4, which is pretty crazy because I feel like Gen3 didn't come out that long ago. Alright, so kicking off their little demo video showcasing this thing's capabilities, we've got dunes during a sandstorm. I like that the sandstorm is kind of whipping across the screen and the physics there, the physics on this guy's cloak are pretty awesome. I also love that the light is shining through in a pretty reasonable way, like through the fabric. And we can already see his arm through the fabric as well. So that's good coherency of this character. And this shot, I mean, impressive. Moving on. We've got a girl here, very low depth of field, she's sort of wandering through the woods. We've got a Chevy Camaro, I believe blasting up here. And vehicles are traditionally pretty difficult for video generators. Now keep in mind, guys, as we watch this, there's no doubt in my mind that this is all cherry picked videos, but they definitely are coming out of Gen 4. We've got two massive structures and then a character standing in between them. Again, more physics in the desert. Some crazy eyes. Look at that vulture thing. Is that a vulture? But wow. Spreads the wings and then contracts them in. And that is impressive right there. Movement of those wings. I mean, you thought it was impressive seeing human movements done with AI. Check out animal movements, man. Like complex ones, like a bird. Wow. Probably the most impressive shot I've seen so far from Gen 4. And by the way, notice that the clarity here, the consistency in the background, it's all very much there. It doesn't have a lot of that AI scribbled, you know, mushy, hallucinatory mess factor going on. It's very clear, it's very defined, it's pretty professional looking at the end of the day. And I think Runway's always done a pretty good job in that area of the AI video generation sector. Here's a rocket taking off, man. That's pretty cool. The more abstract the concepts become, the more impressive these generations are to me. Like, yeah, there you go. Abstract concepts. Jellyfish movement. What we figured jellyfish movement out now. Man, this is crazy, all of this. You might notice that I don't have any Music in the background. This does have music, but I'm almost certain that they're putting copyrighted music for this ad. So I'm not even gonna try with the the YouTube monetization, but man. Yeah, these examples are crazy. Look at this little gray monkey, dude. Oh man. All the particles forming into that woman shape. Oh, the bridge is collapsing. Oh my gosh. Okay, that was a little bit of AI weirdness motion. You can still see it's not perfect, but that is. That's still pretty darn impressive. What a difficult shot to create too. I gotta give this thing a try for myself. The good thing is that it's available now. It's just. It's fully here. No wait list. No. Only certain users get access right now. Just anyone's got access the Gen 4. Okay, so Gen 3 is pretty old. It released on June 17th of last year, so definitely under a year until Gen 4 is released, which is pretty surprising. They have been improving Gen 3 consistently too, adding a turbo mode and stuff like that. Oh wow. They have all kinds of different little made with gen 4 videos now. We've got this little claymation style rock character. He's eating a flame of some form. We've got a cute little skunk. NYC is a zoo. Oh man. That's crazy. Wow, look at some of those shots, man. Okay, I'm kind of just blown away by this page. The herd. Yeah. Wow. Look at the detail in some of these shots, man. Oh man. AI video is. Is really coming along on the website when you go to pick gen 4 out of the model list. Labeled as the most consistent and controllable video model that they offer. Looks like we get durations up to 10 seconds, which is pretty good to see. Although I'd like to see that increase to at least like 20 seconds. So no changes from Gen 3 in that regard. Looks like it supports image uploading right off the bat. So I've just uploaded an image here made with GPT4 Omni. I am going to now crop it to an aspect ratio that's supported by Gen4, which appears to be all of these ones. So we've got your typical 16x9, which is what this video is you're watching right now. Or we have a more cinematic 21:9 or you can even do 4x3, which is pretty nice, which is like what it automatically recommended. You know, they've also got the more portrait styles as well. So a wealth of options to choose from. You know what, let's stick with four by three. I think that works for this character, we'll say man, stares into the camera for a beat, then sprints away in the distance. As he sprints away, the camera starts to pull away and fly upwards, showing the landscape. So we'll try that at five seconds. I also will try it at 10. By the way, if you're wondering what this generation is from, it's actually from my recent video on Saga AI, where we actually created an entire movie front to back, like a full plot and everything. So that was pretty crazy. So far, generation time seems to be pretty decently quick. So far, these generations are moving pretty quick. We're at 12% with the first one, 10% with the last. Last one. Definitely faster than just image generation on GPT4 omni, which has been pretty slow lately. This is a photo of my actual car. Alright, so the prompt for this, the camera moves towards the left hand side of the car. So I want the camera to sort of rotate around the back end of the car, stopping when it faces the rear of the car. Then the car is going to speed off, leaving smoke in a like cinematic comedic fire trail behind the tires. So we'll do 10 seconds for the first one. I'm also going to try a five second here. All right, looks like our first five second is done. So he's gonna stare at the camera and then sprint away. Okay. I wanted the camera to back away a little bit, but not too bad. He definitely stares in the camera for a beat and then he sprints away. I like his sprinting. I think the movement's good. Let's see. Even kicks up some decent dust clouds as he runs away. The motion's a little bit confounded as he gets smaller and smaller, but it's not too bad. Now the one thing that is pretty evident here, and one of the reasons I chose this image as a test, was he has no arm here. He's got no left arm. And I mean, this image was generated by GPT4OMNI. And it's even difficult to get a traditional diffusion generator to give you an armless guy. But GPT4O pulled it off. Putting it in Runway, though, doesn't seem to help Runway out, as the arm simply grows back when he sprints away. He's got a fall arm. So obviously that's still an issue here. If we wanted to have a character with a consistently gone arm, I think that would be pretty difficult. We could try reprompting this to see if we could get it, but I really don't think it would do it. He just grows his Arm back and then he sprints away. So pretty good though overall. Let's see if the 10 second generation did any better already. Stares at the camera, he turns and then, yep, his arm grows back again and he kind of skips away and runs away. It looks like a little bit of his running movements got a little washed up. But yeah, I mean, it's not bad. The camera sort of follows him and it did listen to my instruction. It pulls the camera up to show the landscape a little bit better as he runs away. Not bad, man. Now also, if we wanted to, we could upscale any of these to 4K resolution, which is pretty crazy. So I'm gonna run one of those now. Like to see what that looks like. If it gives it a little bit more detail, a little bit more punch. Okay, we've got the first car speeding away. 1. Let's see if it was able to manage this. There it goes. All right. Pretty awesome. It didn't rotate the camera the way I wanted to. I like the fire coming out of the tire. It did kind of like listen to my prompt. For sure. It's not doing it like in the typical cinematic way I expect. But also this is a 5 second generation, so the thing you have to understand about that is it doesn't exactly have a whole lot of time to generate everything I'm asking for. Anyways. Yeah, not bad. Pretty cool. You know, just to see what it's like. Let's try this with Gen 3 Alpha. And our 4K upscale is apparently done. How's this looking? Definitely a little bit AI esque looking, but it's upscaled. It's a better quality, I'd say at the end of the day. Definitely doesn't fix the motion though. Like if I pause any of these, it's pretty clearly upscaled. Some weird warping and stuff going on. All right, our 10 second speeding away car generation is done. Let's see. All right, the camera sort of rotates. Then the car starts to speed away. We've got way more fire this time. All right. Pretty cool. Yeah. All right now. Not too bad, not too bad. It didn't rotate the camera like in the first place exactly like I asked, but it's definitely still listening to the prompt. We've got crazy fire going on. It's pretty consistent. Alright. This is the same prompt but with Gen 3 instead. So we'll see. Yeah, Gen 3 is immediately way worse. Not even close. We don't even see the fire. The quality is worse. Overall, Gen 4 is, no doubt in my mind, like a Massive leap forward. That is way, way better. Not even close. Well, that's good to know. Let's upload a quick picture of me for a thumbnail in here. A truck smashes through the wall behind the man. He looks back, then looks at the camera, worried. Pretty simple over prompt. A truck though is hopefully gonna smash through the wall behind me. Do a five second and we'll do two ten seconds for that. Thousands of lemons rain down from the ceiling and begin to pile up on the floor and on the man, who looks very happy about it. So hopefully a bunch of lemons are gonna rain down. This should be a pretty good physics test. Not an easy prompt. We'll do 2,5 seconds for that. Here's another image generated with GPT4 omni. It's me looking very creepy in a pile of lemons. So we're gonna try to keep this creepy found VHS creepy footage. Man is obsessed with lemons and brings one up to the camera. Send a couple of those on through. Alright, let's go ahead and try some 3D animation. The camera follows a cute robot riding a rocket ship through space. The rocket lands on the moon and the robot gets out of the rocket. Cute 3D animation style. Obviously this is gonna be a little bit longer, so 10 seconds and we'll generate two of those. Looks like some of our gens are starting to come through here. Alright, this is the truck smashing through the wall. Oh, nothing. No truck. Wow. I like the zoom out with the camera there. It was very consistent. It kept my face pretty good too. Completely hallucinating the rest of my room there. But no truck smashing through the wall whatsoever. He looks behind him like a truck's about to smash through the wall, but it does not happen. Alright, this is the 10 second generation. Maybe it has a little bit more time for this. Oh, oh. And the truck is just there. Okay, the truck is in the room. Oh my God. It looks like it's trying to interact with me and I'm just like clearly trying to talk about the truck and I'm just like explaining what's going on or something. What is happening? Oh my gosh. The pillows falling out of the underside of the truck are crazy. Okay, so it doesn't exactly smash through the wall. With great physics, the truck sort of already there. Maybe we need to prompt it a little bit differently. We do have another generation though. So how did this one turn out? I'm just sort of talking at the camera explaining something. There we go. Okay, it's kind of smashing through the wall. Yep. All right, that was pretty good. I'm just completely unfazed by this truck. Just like, yeah, welcome back to another video. Yes. I like to hold my hands like that. Boom. Truck just sort of molds through the wall and I'm just like, yep, another day. Guys, you have to understand, this is what it's like for me to try to record every day. Trying to explain some AI tech. And then a truck materializes from behind me and I'm like, damn, gotta re record that. Alright, now for the lemons. Let's see. Okay. Better physics for sure than I was expecting. But yeah, they all just sort of fall in a pile right in front of me. The camera zooms out and they're just there. I mean, they're a little bit wishy washy. It's a lot of lemons. And you know, they're not as detailed as they would be in real life, but I seem to just be very happy about it. I'm like, yeah, check it out. This is Lemon world. We've got another generation of that though. Oh my God. This is like me actually getting buried with lemons. That's actually kind of cool. I like that it just sort of zooms out and there's thousands of lemons are piling on top of me and I'm just like loving it. This is exactly what it would look like in real life. Perfect. Oh, here's the creepy VHS footage. Oh God. So I'm supposed to hold the camera right up. Oh my God. Yeah, Lemon right up to the camera. That is terrifying. Guys, you want 11? Oh my gosh. I don't even know what to say. Yeah, this one's gonna need to go right to the freaking chat. By the way, if you make any cool generations with gen 4 or if you make any cool generations with AI, point blank, period. My Discord server would love to see them. Really awesome tight knit community on here. And they post some awesome stuff. They're Gonna love these. Gen 4. Okay, but quality wise, I gotta talk about the quality. This is actually pretty impressive. It maintained the VHS style for the most part. And yeah, I hold a lemon right up to the camera and super creepy. It nailed this. I think. Exactly what I asked for. Can it do it again? Oh yes, it can. That's even creepier. Oh my God. I have like insane guy face. Like I'm smiling but then like something strikes me and I'm like, oh, I need to show them the lemon. Look at the lemon. Very terrifying. We actually have a request from a community member. Whoosh. Trying Gen4 out with this style of animation? Why not? Obviously this image was created by Woosh with GPT4OMNI. So it's kind of like this crayon artistic style drawing. It's pretty unique. So before we check out the rocket ship video, let's send this guy on through. Red haired female character on the right talks whilst sipping her drink. Male character on the left looks earnestly at her while taking bites of his burger. So we'll do some 5 second generations and 110 second. Alright, well that loads up. Let's check out our 3D animation. So he's going to be riding his rocket. Wow, that looks freakin great. Oh my God. He's blasting off onto the moon. He sort of lands, I think the rocket change shape there which is a little strange. But he gets off and looks around. That is pretty stellar guys. Wow, that first part is like you couldn't even tell it's AI and then you know, some fire shows up in weird places and the rocket changes shape. So you know that's a little bit of an issue. But other than that like even when he first lands and he hops out, it's not too bad. He looks pretty happy. I like that. It keeps like his little robot character face consistent and adds like emotional characteristics and details. Like he's got clearly is having like a happy positive interaction here and happy to be on this rocket. And that's not like a human, you know, this is a robotic animated character. It just applied some like Pixar level character animation without even being prompted to. Man, that's crazy. It's good. Plenty of detail as well. Like the lunar surface. He sort of just hops out. We've got another one. This one's 10 seconds as well. So this one he seems to be moving away from Earth now he's got like an actual face, little digital face. He hops out onto the moon on like a little platform. It seems like he's talking to the camera or saying words of some form. So it's really interesting that it tries to make it narrative in some way. It really does push for that visually. But yeah, I mean maintaining that 3D style. This was a first really shot. This was like one shot man. That blew my freaking mind, man. Alright, now for Whoosh's prompt. Let's see. So the character on the side is gonna be yapping, sipping her drink does okay. He kind of pulls up and is focused on her eating his burger. It doesn't add too much to the animation to be honest. It keeps most of the frame still, which is really interesting. She does go to take a sip of the drink and I think that looks good. But she's actually not pulling the drink up. In the next couple of frames though, she actually does pull the drink up and take a sip properly or almost properly through her nose. Meanwhile, he bites the burger. And that looks pretty good. I think he came out a little bit better than her overall, but it's interesting that it didn't add too much movement in there. How about this one? Yeah, he's sort of yapping at her. She's taking a sip of her drink. He's trying to take a sip of a drink. Kind of similar. I think this was a little bit worse. And for the 10 second one, how did this turn out? Okay, he puts the burger down. She sips her drink, blinking her eyes. This one looks like it's coming out a little bit better. Oh, he's like surprised that she's taking a bigger drink. Yeah, that's not bad. Alright. I think this one's the best. Even though he doesn't really munch the burger too much. Alright, we set those examples back over to Woosh. Slow motion footage of a fox jumping around in the snow. A very pretty little image here for that input. This should be relatively easy. I want to see if it can consistently nail. This should be a pretty good demonstration of like physics and realistic movements with this creature. Alright, let's stop uploading images and start to prompt it. Raw physics simulation. Jelly is raining from the sky on 3D. Animated characters holding umbrellas in a city made from ice cream. Oh please upload an image. Oh my gosh. This thing folks, does not take regular prompts. You actually need an input image. You could always generate it with their built in frames. So that's probably worth showing off here. They have their own image generator here on Runway. It's frames. I don't think I. I don't think I ever really covered it or talked much about it on this channel. But it's something that exists. Not as good as, you know, any of the top generators. GPT4O obviously leading the pack. Ideogram 3.0 Reeve image generation. Those are going to be your top three contenders right now. But frames isn't bad. Okay. Got some pretty interesting generations. These are kind of not super coherent, so I don't want to use them as inputs. Let's do this on Ideogram A. By the way, the reason I'm using ideogram AI instead of native image generation with GPT4 omni is because native Image gen is very slow right now and this is like multitudes faster and should give me, yeah, decent enough quality images. Honestly, I like all of these. Let's grab this one. Let's also grab this one. Boom. Let's do a 10 second and a 5 second and for this one a 5 second and then another 10 second. Looks like we do have our fox generations coming through though. So the first one here is 10 seconds. Remember this is a slow motion footage. Definitely slow mo footage so far. The fox. Oh yeah, he's leaping and running through the snow a little bit. It's very close up, but yeah, he just sort of runs away. That was not bad. That was very coherent. Difficult, I would say, to place this as an AI generation to the untrained eye. Obviously in the fur details and stuff like that, you can see a little bit of shimmering, which is pretty typical of AI generated video. But yeah, he runs away. We got good snow physics. It's not a bad generation. Got another one. Can it do it consistently over time? Yes. The fox blinks and he sort of hops away. Oh, and he flies away like a reindeer. Okay, well that wasn't as good as the first one we saw, but still not terrible, I would say. I mean, feasibly he could just have a very powerful jump and he's just leaping away into the sky. Here we go for another generation. This one's a little bit quicker. Not as slow mo, but still a little bit slow mo, I would say. He just sort of picks up and runs away, which is pretty realistic. And our last one sort of gets a move on again, sort of hops away. I like that one. I like the motion with all four legs. You can kind of track it. Seems like he might have missed his leg somewhere in the middle of the video. It kind of disappears. But this first generation is definitely, I think, the most impressive and looks the most like real actual footage. He sort of hops away in slow motion. I like that. Here's a cool image I made a little bit ago in neogram v3. Let's plop this guy in deep macro zoom showing the individual molecules that make up the film reel. All right, we'll throw a couple of those in for 10 seconds and we'll do 15 second. Alright, looks like our physics simulations are completing now. Here's our first one. Wow. With all that pink blobby stuff and these characters kind of just staring around watching it, they look pretty happy about it. You know what, this is actually pretty consistent overall. You know, the initial frame had all of this raining down. So obviously all of those became like their own physics objects. The characters sort of watch it and then one more big blob comes down because we did prompt it as like physics simulation test. I like that. The ice cream clouds in the sky also begin to melt and like goop sort of starts to rain down from them, which is really interesting. So that's pretty cool overall. Interestingly though, gravity on this one seems to vary. Like it's got a certain level of gravity in the first here, but in the second half with the drips coming from the sky, it seems like gravity slows down a little bit. How did this one come out? Oh wow, this is a lot faster. You can see all of the umbrellas disappear and they sort of look around, all the drips raining from the sky. They're like, yeah, what's up? Alright, here's the next image. How does this look? All the goop sort of falling from the sky. That actually looks kind of cool. Oh, I like that. It's kind of satisfying as it rains down and we see it like sort of seep and it's got plenty of. And all the characters are just like, dang. I guess we didn't need the umbrellas because none of the goop landed on us. Amazing. The characters really do feel alive in a lot of these. Like they react to their environment and they're sort of following what's going on, which definitely adds to your generations. Even if it's like a simple physics demo like this. All right, now here is the longer variants in this one. We don't have any of these falling from the sky, at least not yet. Oh my God, they're getting gooped on. He's got another umbrella though. And the camera moves away to reveal what an egg at the side. Sometimes these generators can just be so random. Wow, okay. I mean, yeah, it's like a physics demo. It's pretty interesting. But then we've got an egg over here on the side. Who is this little fella? Alright, now for our macro zoom. And I believe, yep, we've got the AI kind of cornered around here because it doesn't know what to do or how to turn this video into like a macro zoom. Like I wanted it to zoom into the very fine details of this film strip and show me like the molecules, like an ultra close up zoom, but it sort of just makes like a pattern and then still has the clothes line there. So didn't really know what to do with it. Alright, this one instead. Can we get a nice zoom oh, we're getting closer and closer. Nice video so far. How close can we get? Will we have enough time to see molecules? Probably not, but it gets close. It's trying to do a macro zoom. I kind of like that transition though. Like, this was pretty cool. It just sort of slowly gets closer. You can actually see the focus change with the blur in the background. That's really impressive. As we get closer to the film strip and the focus of the camera, theoretically, I mean, there's no real camera, but it follows the focus on the film strip as it gets close. Thus the bokeh balls in the background get larger, which is what would happen in real life. So that's just really impressive. And how about this one? How close did we get? Oh, and the film strip has exploded into molecules. It seems very interesting. So doesn't know exactly what to do with it. The closest one was this first one up here. Did a pretty decent job at that. So, yeah, guys, that's first impressions of Runway gen 4. It's definitely an impressive model. I think it is very much attuned for realism, making movies and the real world. Right? I mean, that's what most people are doing with the examples. But it also seems to be pretty strong in 3D animation and really applying emotions to characters in 3D animation. And out of 3D animation, 2D animation seems to be a little bit more lackluster, especially in comparison to things like vdu. Vidoo kind of rules the roost, I think, in terms of 2D animation. But, you know, we're gonna start to see these different AI video generation models sprout off into their own, I think, areas. There's just way, way too much depth in the video space for them all to be able to occupy all of it. I don't know if that's possible right now. So it seems like Runway's really leaning into to their stuff a little bit more. And it is paying off for sure. It's absolutely, you know, a different league above Gen 3. It's a lot better and it's a lot more consistent. But people such as Eleanor here are already creating little mini movies. This is a one minute little mini movie about a race car driver made with Gen 4 in about 20 minutes. That's another thing. Gen 4 has great generation speed. So far. It's a really serious bump above Gen 3 while also being just as fast, if not faster. I'm sure we'll get a better idea of this model and its capabilities in the future as more and more community members test this thing out and play around with it. I'll be reposting stuff on Twitter and sharing more examples as I continue to generate on my Discord server. So please be on the lookout on those platforms if you want to see more. Thanks so much for watching guys. More to come this week and I'll see you in the next video. Goodbye.