← Back to video archive

Airdroplet AI summary

Runway Gen-4 is actually insane (text-to-video)

April 1, 2025Matthew BermanAI score 10049,526 views

Watch original on YouTube ↗

AI-generated summary

Runway just dropped their Gen 4 text-to-video AI model, and it's seriously impressive, marking a big leap in creating videos just from text prompts or images. The quality and realism shown in the examples are pretty wild, showing off major improvements in how realistic things look, how smoothly they move, and how well the AI understands the real world compared to older versions.

Here's a breakdown of what's covered and the key takeaways:

Introduction & Initial Wow Factor: Runway Gen 4 is here, pushing text-to-video generation forward significantly. The quality in the demo examples is described as "insane." An example shows a crosswalk sign figure realistically jumping off the sign and walking around, with accurate lighting and cars. Another example features unsettling bird-guy hybrids, labeled "nightmare fuel," including chicken, flamingo, and turkey versions. The flamingo-guy turning its head all the way around is highlighted as particularly creepy. The turkey version with huge necks is called "absolutely gross."

Hyper-Realistic & Cinematic Examples: Several demos look like they could be from actual movies. A scene resembling the movie Dune is shown. A woman walking through a forest demonstrates impressive object permanence – she looks correct even when briefly obscured by branches and leaves. Examples include a beautiful Chevy, a futuristic sci-fi scene (like Star Wars), a foggy environment (like Alaska), realistic water physics, and a Game of Thrones-style scene. Another car example doesn't look quite as good, but the physics are still noted as solid. A rocket launch is shown, described as "beautiful." Other impressive visuals include asteroids, an explosion, a woman realistically on fire ("fire looks crazy"), jellyfish with good physics, a Final Fantasy-esque scene, another spaceship, a stylized robot, a woman made of rope, and a monkey. A bridge collapsing on fire has some "oddness" but is generally incredible.

User-Generated Content & Flaws: An example by user Eleanor, created in 20 minutes using images from Reeve, shows people. Most people and hands look perfect in Eleanor's example. However, a mistake is spotted: one man's arm appears to be pointing the wrong way, and another arm seems to come out of a different person. Old cars driving by look good, and a woman's face, blinking, glasses, and hands are called "flawless." Another minor issue is noted with car wheels spinning strangely.

Official Gen 4 Announcement Details: Runway introduced Gen 4 as their new state-of-the-art AI model series for media generation and world consistency. It's positioned as a significant step up in fidelity (visual quality), dynamic motion, and controllability. Image-to-video features are rolling out immediately to paid and enterprise plans. Gen 4 is stated to be a marked improvement over the previous Gen 3 Alpha. Key strengths highlighted: generating highly dynamic videos, realistic motion, subject/object/style consistency, superior prompt adherence, and best-in-class world understanding. This "world understanding" (physics, object interaction) is noted as a crucial aspect, similar to what made OpenAI's Sora stand out.

More Diverse Examples: The presenter emphasizes that the current output is "the worst it will ever be," implying rapid future improvements. Cartoon examples looking like Pixar movies are shown, described as "gorgeous." A short film example, "The Lonely Little Flame," features clay-style animation (claymation/stop motion look). The movement style matches the perceived material (clay/stop motion). Features a pigeon, a skunk, a flame character, and a rock character. The story involves the flame setting the skunk on fire, then befriending a rock. Overall, the short looks "beautiful" despite the odd story. Another short, "New York is a Zoo," shows various animals (rhino, giraffe, elephant, bear, zebras, lion, monkeys) in NYC settings. Some awkwardness in movement is noted, but overall it looks "really good." The monkeys on a traffic light look "flawless," reminiscent of Jumanji. "The Herd" short film features a man chased by cows at night, created using Gen 4 and image references combined with Act 1 software. The cows look "flawless." There's an "uncanny valley" feeling with the human character. A scene with gasoline, a match, and fire shows "incredible" and "very realistic" flames and smoke. The silhouettes of cows against the fire are highlighted as impressive. The potential for generating full movies with just prompts ("vibe movie making") is discussed.

Detailed Analysis of User Examples: A subway scene shows a man standing still; the way a pole realistically passes between him and the camera, preserving details on his shirt (like folds), is called "very cool" and "very impressive." A man running looks "very natural." Paint splashing looks "really good" with beautiful detail. A car on a snowy mountain road looks "really good." A flame in a glass ball is shown. A woman with an umbrella walking: while beautiful, a flaw is noted where puddles react to her movement even when she doesn't touch them. A guy in a room with light balls. A volcano scene looks "gorgeous." Fireworks/light explosion shows some minor mistakes. A guy skateboarding with an umbrella looks "really cool." An awkward scene of a mother and daughter arguing: characters don't look at each other, daughter points randomly, mother's posture changes oddly, daughter fumbles unrealistically with a suitcase. This example is deemed "not good actually." A fabric/felt-style animation example: Features a bird, caterpillar, snake, porcupine, fox, deer, raccoon, and ducks. This style looks "phenomenal" and is a personal favorite of the presenter. The snake is called "unbelievable." The cartoonish, fabric-y look is consistent. Minor critique: water dripping from the raccoon's mouth looks good, but the water it hits doesn't react quite right. The fabric ducks floating (instead of absorbing water) are noted but still look "so cute." A rabbit running with a light trail looks "pretty good." A subway scene combining a hyper-realistic human, a 2D cartoon, and a fuzzy teddy bear: The lighting is praised for reflecting correctly and consistently across all three distinct styles. Called "very nice." A Pixar-style shot by Mad Pencil shows a character with electricity/lightning; the lighting is again called "beautiful." Futuristic cars backing up: Wheels spin unrealistically fast for the movement speed. Looks "okay" and "cool" but "far from realistic." A view from inside a luxury train looking out at a misty forest: Trees and scenery pass by consistently and realistically. Called "very nice." Hyper-realistic wolves: Details like wet matted fur, teeth, tongue, nose, and hair look "really good." Focus shifts (defocus/unfocus) and water interaction are well-handled and "really nice." Hyper-realistic screaming woman with turquoise glow: Teeth remain consistent. Some slight, awkward morphing of the tongue is observed. Overall, considered "very, very good." Hyper-realistic model shoot style: Dewy skin, shadows moving across the face. Called "very impressive."

Concluding Thoughts: Gen 4 handles an immense amount of detail and nuance, making this a "huge launch" and "very, very impressive." Runway Gen 4 is definitely giving competitors like OpenAI's Sora "a run for its money."

Video transcript

Open transcript
Runway Gen 4 has arrived. The text to video revolution is here. Look how insane the quality is of these examples. All right, so we have a regular looking scene. We have a crosswalk and then all of a sudden the guy from the crosswalk sign jumps out and starts walking around. All of the lighting looks realistic. The cars look realistic. It is incredible. Here is one that is absolutely terrifying. It's a guy, bird, bird that looks like a guy, guy that looks like a bird. Doesn't matter. Here's your nightmare fuel. We have some chicken versions. We have this odd looking flamingo guy. Absolutely terrifying. And of course he pulls his head all the way around. We have a turkey, a bunch of turkeys. Yep. With huge necks. Absolutely gross. And here are some hyper realistic demos that really could be any movie. Look at this. This looks like a scene out of the movie Dune. Okay. We have a woman walking around the forest. And what you'll notice is as she's walking, there are branches and leaves that are basically going across the screen, covering her up. And when she reappears on the other side of them, it's still exactly the physics that you would expect. Look at this. Really cool. Here's a beautiful Chevy. Here's something that looks like it was out of a Star Wars or a futuristic sci-fi movie. Here is something with a lot of fog. Looks like maybe Alaska water physics. Here's something that looks like Game of Thrones. Another car. This car doesn't look actually great, but the physics are still really good. And a rocket taking off. Yeah, that looks beautiful. Look at that. All right. We have some asteroids, an explosion, a woman on fire. The fire looks crazy. The jellyfish, the physics on the jellyfish. This looks like a scene out of Final Fantasy. Here's another spaceship. Really beautiful spaceship. Here's a robot, like kind of highly stylized robot. Look at that. Woman made out of rope, monkey. I mean, all of it. Look at that. Unreal. There's a bridge on fire coming down. Oh yeah. A little bit of oddness there, but just all of these look really incredible. All right. Here's something that Eleanor built. Holy cow. I did this in 20 minutes with runway gen four images made by Reeve image. Okay. So the initial images were made by Reeve and then uploaded to runway. And here we go. Let's look at what these look like. All right. Really, really good. All of the people look perfect. Look at the hands. The hands all look good. So she's talking. Not sure who she's talking to, but fine. They're all looking up pointing. Oh, I see a little mistake here. Look at that. So we have this guy who's kind of taking a photo, but his arm is out the wrong direction. And then we have kind of his other arm coming out of this guy here. So a little flawed, but still really good. Beautiful old cars going by. The woman's face looks flawless. Yeah. Look at that. I can't find a single mistake with her blinking her glasses, her hands. I mean, really just perfect. Here's a little bit of issue. So look how the wheels in this car kind of come out and spin out. Can't tell what's forward, what's backwards, but still really, really good. So today we're introducing gen four, our new series of state of the art AI models for media generation and world consistency. Gen four is a significant step forward for fidelity, dynamic motion and controllability and generative media. All right. Image to video rolling out today to all paid plans and enterprise customers. Gen four sets a new standard for video generation and is a marked improvement over gen three alpha. It excels in its ability to generate highly dynamic videos with realistic motion, as well as subject object and style consistency with superior prompt adherence and best in class world understanding. That's crucial. That was kind of when Sora first came out what it was really identified for, which was it has an understanding of the world, the physics of the world. And now we have something that could really be better. Okay. Little cartoons, polar bear jumping under. I mean, all of the people look incredibly real. And I think right now you only get like 10 or 20 seconds worth of video. But again, this is all the worst it'll ever be. Keep that in mind. Every time I show you something, this is the worst it will ever be. Okay. So cartoons looks like a Pixar movie, just gorgeous. All right, let's look at this other one. This is a short movie, the lonely little flame. So beautiful pigeon kind of clay style. Everything looks really good. And you can even see the movement looks like the movement. It should be based on the materials that are being used. So we have a little skunk walking around. And you can kind of see the low frame rate in it, which maybe you would think this is like stop motion or something. So here we go. The lonely flame flame looks good. Coming up to the skunk. And yeah, don't touch the flame skunk. Yep. There you go. Now you're on fire. Flame sad. Okay. I don't know what character this is. But yeah, it just looks like claymation or stop motion animation. There's the flame. Oh, the rock. It's a rock. Okay, sure. And maybe the rock is the only thing the flame is going to be able to have fun with. Yeah. So weird little short, but nonetheless, it looks beautiful. All right. So this is New York is a zoo. All right. We have a rhino in the zoo. Yeah, yeah. Different animals walking around New York City. Yeah, I mean, a little bit of awkwardness in the movement. But overall, it looks really good. Okay, we have a giraffe walking around and elephant just sitting there. A bear looking into kind of a fisheye lens. Zebras drinking out of the river. Here's maybe Chinatown. Zebra looking at the camera. Lion running around. Some monkeys on a traffic light. I mean, those look flawless. Something out of Jumanji maybe. Yeah, all of these look really good. So this next one is The Herd, a short film following a young man being chased through a field of cows at night. It was created using Gen 4 and just a few image references to build out each of the shots of the characters and the misty field of cows. It was then combined with Act 1 to bring the story together. All right. So we got a guy crawling through the field. Uh-oh. The cows look really flawless. That would be worth it. Definitely a little uncanny valley feeling here, but still really good. Okay. A little gasoline going in the water. Let's see if he's going to light it. Let's see the lighting if he does so. Yep. There's the match. The flames look incredible. Very realistic. The smoke is stunning. Look how the cows, you can see the silhouette against the back drop of the flames. I mean, this is, this is very impressive. Yeah. You put some music behind it, put a little dialogue behind it. And it's, uh, you could imagine in a year, two years down the line, when we're generating hour long movies from this and you get to just say what you want. This is vibe movie making. I mean, it's incredible. All right, let's look at some more user generated examples now. So I'm going to turn the music off. So my video doesn't get claimed and, uh, let's just watch. So we have first the guy in a subway, just standing there. Let's see what happens. So he's looking around and by the way, let's rewind for a second. See this pole right here. Look how it goes right over in between the camera and him. And the guy comes out the other end. Look at all the little folds on the shirt. So look at this upside down V shape fold right there. Now I'm going to rewind it and we're going to watch it. So there it is. And then it's covered up and it comes out the other end exactly the same. That is very cool. Very impressive. Let's keep watching. Okay. We have a guy running, running looks very natural. Here's some paint splashing, which, wow, that looked really good. Let's look at the detail in this one. So we have a bunch of paint coming down and look at all of that. Beautiful. Here's a car on a snowy road up in the mountains driving around. This looks really good. Okay. We have the flame and a little glass ball. We have a woman walking down the street with an umbrella. Now, one thing I do notice is look at all the water on the ground. There are some mistakes. If you look right here, look right here, I'm going to play it. And you can kind of see it doesn't quite look right as she's walking by, even though she's not touching the water, the water is kind of reacting to her movement. So not perfect, but still beautiful. Let's keep going. So guy in a room with a bunch of light balls. Here's a volcano kind of looking like a volcano. Gorgeous, gorgeous scene right here. A little explosion of fireworks and light, a little bit of mistakes. And here's a guy riding a skateboard with an umbrella. Really cool. All right. Next we have what looks to be kind of a mother and daughter arguing with each other, although it's definitely awkward. They're not looking at each other. She's using her phone. She's pointing at something very angry. And yeah, the mother is not actually looking at the daughter. The daughter is pointing out into the middle of nowhere. She's very upset. Oh, mama standing. Now she's sitting. She looks very unhappy. Daughter's packing her bag, ready to go. Yeah. None of that looked good, actually, to be honest. Look how she's fumbling with her suitcase. It just does not look real at all. And the mother's just standing there looking. Okay. I mean, look, I'm criticizing, but at the same time, this stuff didn't exist at all a year ago. All right. Here's another one. This is kind of like fabric looking scene, definitely cartoonish. We have a bird that looks like it was made out of like pillow material, some fabric and let's give it a watch. Yeah, this looks really good. Okay. We have the wings flapping away. Here's a caterpillar. Same thing. Very much like felt fabric material. This looks phenomenal. This may be my favorite one so far. This easily could be a full length movie if they just put a bunch of these different scenes together and a storyline and some sound effects. There's a gorgeous snake. I mean, that is unbelievable. A porcupine coming out of a little hole. Again, everything is very like cartoonish and fabric-y. Okay. That is odd. Little fox running around doesn't look very real, but fine. There's a deer and a raccoon. The water dripping looks phenomenal, except let's go back for a second. So the dripping out of the mouth looks really good, except when it's hitting the water, it's not quite reacting how I would think it would react, but that's okay. Oh my God. That's so cute. Look at that little ducks floating around in the water. And of course they're made of fabric, so they would be absorbing the water, but they're not. That is really good. I'm very impressed by this one. Good job, Timmy. All right. Here is a rabbit running with some kind of light trail coming off of its back, almost like a, maybe a rocket ship trail. And yeah, that looks pretty good. Here's a really cool one. We have a subway human, a cartoon on this side and let's see what happens. Look at the lighting, by the way. So we have some kind of exterior lighting happening and reflecting perfectly on the cartoon. Also on the real subway, just the lighting came out perfectly. And then we have this, what looks to be a teddy bear holding an ice cream cone coming in. Okay. Sitting right in the middle and just look at it. So it doesn't matter what style it's in the hyper-realistic human getting the lighting perfectly, the kind of fuzzy teddy bear getting the lighting perfectly. And then the more 2d cartoon getting the lighting perfectly as well. Very nice. This is by Mad Pencil. All right. And this may as well be a shot from a Pixar movie. Look at this. So we have a nice cute character and kind of doing something with lightning or electricity. The lighting again looks beautiful. Just a very short 10 second scene, but what you see could be extended. And this could be made into a movie once again. Now we have a set of cars, hyper-realistic, but very futuristic looking cars. And let's see what happens. So they light up, they back up. Now the wheels are spinning much faster than they should be based on the speed of the car movement. And it looks okay. Not, not perfect, but it's cool. Just far from realistic. All right, next, we have a scene from a looks to be inside of a train looking out into kind of a misty forest as the train progresses. And it looks like maybe a luxury train. And this looks really good. All of the forest, all of the trees are consistent. They're going by the window exactly how I would expect them to. Very nice. Next, we have a hyper-realistic scene of a few kind of wolves coming at you. Look at the wet matted fur. That looks really good. All of the teeth, the tongue, the nose, all of the hair, the fur looks good. Let's keep a close eye on the lighting, the defocus, the unfocus in the background, and plus how they all interact with the water and the water physics. Let's watch. The front wolf kind of goes out of focus now as the back ones become more in focus. Really nice. All right, now we're going for hyper-realism. We have a woman looking directly at the camera, some green or turquoise glow going on and she is just screaming. Let's look. Okay. The teeth are all staying the same. The tongue is kind of a little awkward. I think I saw. Yeah, a little bit of morphing with the tongue, but very, very good nonetheless. And another one we're going for hyper-realism. This looks like it could be maybe a model shoot. Let's take a look. Very like dewy skin, a lot of silhouette action shadows going over her face. Yeah. Very impressive. There's so much detail and nuance going on in these videos and Gen 4 has to keep track of all of it. So I think this is a huge launch. I think this is very, very impressive. Now I don't have a runway account, but if you want to see me put together a long video of trial and error and trying to get these videos to look really good, let me know in the comments, but that's it. Let me know what you think. This is giving Sora a run for its money for sure. If you enjoyed this video, please consider giving a like and subscribe, and I'll see you in the next one.