구글의 Veo AI 동영상 생성기: 심층 검토 및 Sora와 비교하기

### Google의 Veo 살펴보기 AI 동영상 생성의 미래
* 🎥 Google의 Veo 플랫폼을 통해 AI 동영상 생성이 새로운 차원에 도달하면서 **실제 콘텐츠와 AI가 생성한 콘텐츠를 구분하는 것이 점점 더 어려워지고 있습니다.**
* 🔍 Veo의 얼리 액세스를 통해 **텍스트-비디오 및 텍스트-이미지-비디오 옵션**을 포함한 기능에 대한 인사이트를 얻을 수 있으며, 사용자는 **액세스 대기자 명단에 등록**할 수 있습니다.
* 🐶 수영장에 뛰어드는 닥스훈트의 동영상에서 볼 수 있듯이 **세부적인 안내를 통해 인상적인 결과**를 얻을 수 있지만, 일부 출력은 때때로 **의도한 동작과 반대되는 결과**를 낳기도 합니다.
* 🎬 사용자는 **시네마틱 스타일로 동영상을 제작**할 수 있으며, 플랫폼에서 여러 결과물을 생성할 수 있지만 현재 **이전 프롬프트를 저장할 수 있는 라이브러리 기능이 없습니다.**
* 📊 Veo와 Sora를 비교한 결과, 두 플랫폼 모두 여러 요소가 포함된 복잡한 프롬프트에 어려움을 겪지만 일반적으로 **Veo가 더 사실적인 결과**를 생성하는 것으로 나타났습니다.
* 💡 AI 동영상 생성의 잠재력은 무궁무진하며, **Veo는 업계에 새로운 표준**을 제시하고 있으며, 향후 개발을 통해 그 기능이 더욱 향상될 것으로 예상됩니다.
* 🔮 AI 기술이 계속 발전함에 따라, 곧 출시될 **베오와 소라 버전**은 이 분야의 빠른 발전을 보여주며 사용자들에게 더욱 큰 감동을 선사할 것으로 기대됩니다.
* 📅 관심 있는 분들을 위해 라이브 스트리밍을 통해 **Veo를 실험해 볼 수 있는 기회**를 제공하여 사용자가 프롬프트를 제출하고 실시간으로 결과를 확인할 수 있습니다.

Google의 새로운 Veo 동영상 생성기를 사용하면 AI가 생성한 동영상을 현실과 구별하기가 점점 더 어려워집니다.

The videos you're seeing on the screen right now, some of them are real and some of them are AI-generated. I saw Marques do this in one of his recent videos about Sora, and I feel like even since Sora's come out, we've leveled up even more with AI video generation.

Now, people that are really paying close attention can still probably tell the difference between what's real and what's not. But now that Google has started to roll out access to their new Veo video generator platform, we're truly getting to a point

where it's more and more difficult to distinguish between reality and AI. Now, I was one of the people that was fortunate enough to get early access to this new Veo platform inside of Google Labs video effects. But if you're interested in playing with this, you can go to this URL

labs.google/tools/video-effects, and there's a waiting list that you can join. Ideally, if you join that waiting list, you'll have access fairly soon as well. Once you're logged into video effects, you actually have two options: you have a text-to-video and a text-to-image to

video option. Now, please keep in mind that I did get early access, so by the time this rolls out, there might be more settings, the layout may have changed a little bit, and there may be more or possibly less guardrails on the videos that you generate with it. To get

a video, I can give it a prompt. I like to go more detailed, so I gave it a cinematic, high-action tracking shot that follows an incredibly cute dachshund wearing swimming goggles as it leaps into a crystal-clear pool. The camera plunges underwater with the dog, capturing the

joyful moment of submersion and the ensuing flurry of paddling with adorable little paws. Sunlight filters through the water, illuminating Doc and Sleek's wet fur. I find it helpful to start with a basic prompt and then use something like ChatGPT or Claude to improve the prompt and add more details.

I can now click create videos, and after roughly 3 minutes, it gives you four videos. So we got this one here, which looks more like a dog jumping out and back into the water, but it still looks pretty realistic. Honestly, we got this one of a dog underwater; it's kind of shooting more coming out of the water, so it sort of reversed my instructions. But again, you can see the realism is there.

Here's another one with the dog kind of popping out of the water as opposed to coming into the water. And once again, another one of a Dachshund wearing goggles swimming underwater. It sort of messed up the order of things that I asked for in my prompt. It was supposed to LEAP into a crystal clear pool while the camera plunges underwater with the dog.

But I seem to have gotten a lot of videos that are kind of the reverse. Then we have the text-to-image-to-video model. So this is like image-to-video that we've seen, but right now it looks like you can't upload your own images. You have to essentially generate an image with Google's Image first and then turn that into a video.

I'm going to go ahead and do the 'I'm feeling lucky' and we get a cinematic, moody shot captured of a fair-skinned man in a dark leather jacket leaning against a silver car.

Foreground, let's just go ahead and see what happens when I try to get lucky. We get this starting shot here, which looks pretty decent, and then I can move on to step two, which is the motion step. We can describe the motion we want to actually see in the video. I'm going to

go ahead and use the prompt that it sort of pre-filled in for me because it's a pretty good one: a cinematic short film opens with a low-light shot of a man in a leather jacket leaning against a vintage car, etc. Let's see how it animates.

This way, let's go ahead and create the video. Once again, it's going to take a couple of minutes, and then we should get four outputs from it. After about five minutes, we got these four outputs here that all look pretty good. I mean, they

look like a cinematic shot that you would see from the opening of a TV show or a movie or something like that. I'm pretty excited about it. Now, if we like what came out here but we want to see more, we have the option to get more

variations. Once again, it'll jump back to our original image and start to generate more variations. But I'm going to go ahead and hold off on that because I took to social media. I went to LinkedIn, I went to Threads, and I went to

X, and I asked you to give me some prompts that you'd like to see me generate in Veo. I received over a hundred comments on X, plus a handful more on LinkedIn and Threads, so I wasn't able to generate every single one, but I did

I managed to test about 30 of the prompts that came across, and again, all 30 of them came with four generations. I don't even know if I'm going to get around to showing every single thing I generated, but I will let you know these.

These were all from the first prompt. I did not do any cherry-picking; I took the prompt that you suggested, plugged it into Video Effects here, and then I downloaded the four videos that it provided. So there is no cherry-picking here; this is what they gave me, and you're going to see.

The good, the bad, the ugly—all of it. One thing that this Video Effects platform currently lacks is that it doesn't save the previous prompts that I made. I can't go back through and look through a library of videos that I

created. So instead, I downloaded every single generation that I made based on your prompts. I actually even recorded myself entering the prompts and showing the four generations on a screen like this, but actually showing you all of this video playback of me entering the prompts and showing the results seems

like it would be a bit of a waste of your time. So instead, I'm just going to show you the outputs that I came up with, and you're going to have to trust me that I didn't specifically cherry-pick any of them. I actually entered the

prompt and downloaded the four videos. Just for fun, as a comparison, every single prompt that I entered into Veo, I also tossed into Sora, so we can do a quick side-by-side comparison of how well Sora did it versus how well Veo did it.

So let's start with a simple prompt here. My buddy Michael Stellner said, 'Show a dog surfing in La Jolla.' Here's the first video it generated, where it kind of looks like the dog is running on the surfboard. So, not the best generation, and

then the surfboard kind of disappears. Here's the second generation. This one kind of looks like there's a GoPro or something stuck to the front of the board looking at the dog. This one came out pretty impressive, in my opinion. This one was probably my favorite generation

of the four. It looks like a sort of Collie on a surfboard, and it looks real. I mean, it looks like a drone shot sort of hovering over the dog, and this actually kind of looks like La Jolla in the background. So that one, to me, came

out pretty good. Finally, we have this one, which looks like another sort of GoPro shot, but the dog kind of changes age. If you notice, at the very beginning, the dog looks like a little puppy, and then as the video goes on, the dog stands up and

looks like an older dog. So there's a little bit of a coherency issue there, but still pretty good. And to compare it to what Sora generated, here are my two dog surfing images that Sora made. One, uh, the one on the left, not great

actually neither of them are really that great. I think almost every single one that Veo did was better. Next up, Barnabas Lawson here suggested three. I actually tested a group of people parachuting out of a plane. Here's the first generation that came out of it. You don't

I can really see the plane and all of these people seem awfully close together, but uh, that's what it came up with. Here's the second one; it's like a tandem jump. There's an extra arm going on here. I think one, two, three, four, five—I'm counting five arms. But uh, yeah, so that's the second generation. Not the best. Here's our third generation; the selfie stick here just kind of appeared out of nowhere, but I do like that angle where it put the selfie stick out and you can see it. Uh, but it's just weird how it sort of appeared out of nowhere.

And then here's the final generation where two people are strapped to one person, like a three-way tandem parachute jump. So if we take a look at Sora here, these are the two that Sora generated. The one on the left kind of looks like they're just hovering near the plane. The one on the right has two people that somehow eventually merge into one. Neither of them actually really generated great videos with this prompt. Honestly, Matthew Sher asked to do Will Smith eating spaghetti underwater. Unfortunately, Veo will not let you generate real names.

So if I try to do Will Smith eating spaghetti and then create the video, we'll probably get an error. Yeah, we couldn't create what you asked for; try a different prompt. So whenever you try to put somebody's real name in there, it doesn't work. Sora, on the other hand, did let me enter Will Smith eating spaghetti underwater, but none of these videos look anything like Will Smith. So it's a guy eating spaghetti underwater, but I can't really call that Will Smith. Chris Johnson here said a hamster the size of a skyscraper.

I think that's quite an interesting idea!

Dressed as a samurai, battles a swarm of tiny flying drones carrying paintball guns. I liked that prompt. Here was the first video it generated, which kind of looks more like a still image that has some parallax effects going on. Here's the second video; once again, it sort of looks like a still image with some paintball animations going on in the background. Here's the third one; this one's probably the best looking one, where you can see the drones kind of coming over the shoulders. That looks like a very cool cinematic shot.

Probably the best of the group. And then finally, it gave us this one, where once again the hamster looks totally frozen, and the drones are just kind of going around in circles, you know, spitting paintballs. So one out of the four wasn't too bad. Comparing the same prompt over on Sora, here's what we got. Once again, the one on the left looks like a static image that just sort of zooms out. The one on the right is a little bit more of what we were looking for. Sora probably actually kind of wins this one.

Now, one thing to note is that with Sora, I was generating in 480p, so do your best to sort of ignore the pixelation there because I was trying to get fast generations, not the highest resolution generations. Moving over to Threads, here Nathan said a dolphin is breaching from a sea of honey. I was curious what that would look like. Here's video one of a dolphin breaching from a sea of honey. You don't really see the breaching portion; it almost just looks like a dolphin swimming in some yellow.

A sea of honey. I was curious what that would look like. Here's video one of a dolphin breaching from a sea of honey. You don't really see the breaching portion; it almost just looks like a dolphin swimming in some yellow.

Water, here's the second generation once again. It kind of looks like a dolphin swimming in some yellow water, and right before it looks like it's about to breach, the video starts over. Here's the third one, where we actually see the dolphin breaching what looks like a sea.

Of honey, you can actually see it's got that sort of honey viscosity. But as the dolphin lands, it feels like it's in the air for a little too long, like it's kind of hovering there a little too much. But that does look like that sort of honey viscosity, and it does look like a dolphin jumping out of it.

And then finally, here's another dolphin breaching the sea of honey. Once again, this one to me just kind of looks like yellow water, and there are some weird bubbles there. One out of the four looks pretty good. Comparing this to Sora, here are the two videos that I got out of Sora. Uh, yeah, they don't look great. The one on the right almost looks like the dolphin's turning into honey. I'm jumping over to X.

Now, this is where I got the most requests, and I apologize if I wasn't able to actually get to your request. Here's one that I actually posted just in one of my replies to X, where the prompt was 'mysterious young woman conjures Lamora, her hands enveloped in swirling translucent flames, fed by her raw emotions.' The crimson fire twists and turns, illuminating her sharp features and casting shadows of her inner turmoil. And this right here is what was generated, and I think it follows the prompt pretty well. Honestly, this one was interesting to me: a slow, melancholic pan.

Across a dimly lit DVD rental store, the camera focuses on a middle-aged customer wearing casual attire. His face shows disappointment as he sighs deeply, then gently places a DVD case on a shelf with reverence. As he steps back, the camera

smoothly zooms in on the DVD, revealing the title 'Squid Game' in bold neon red lettering. The case sits nestled between 'Squid Game 2' and 'Squid Game 3', their spines glowing faintly in the store's fluorescent lighting. The shot lingers

emphasizing the irony of the original's placement among its sequels. So, the generations you just saw were what we did, and here's what Sora managed to generate with that exact same prompt. I definitely think we did it better. Steven Greenwald asked for a man playing

with his dog at the beach. In the background, dolphins are surfing the waves. He goes on to say most leading tools have issues with this, and as we can see from some of these video generations here, it also seems to have issues with this. Here's another one

where the dolphins just seem to be frozen in midair. I mean, you've got the guy playing with his dog. At least here's another one: the guy throws the ball up in the air, and then it cuts to a shot of dolphins with a ball floating above them.

And here's the final one: a guy petting his dog and then some just like frozen dolphin images in the back. So yes, Sora struggles with this one as well. Somebody said to give it the prompt AGI, so I was curious what that would look like, so

Right now on the screen, I'm showing you the four videos that it generated when I gave it the prompt 'AGI'. All of them are pretty wildly different. I always like giving abstract prompts just to see what comes out of it. When I gave Sora the AGI prompt, these were the two videos that Sora generated for me, both pretty similar actually. Someone suggested a wolf howling at the moon, which I always like to test. Here's the first video I got of a wolf howling at a moon, which looks pretty good. I don't know about the

moon and why it's moving so fast. I guess it could sort of be a camera panning around the wolf. Here's the second video it gave me with a weird red moon, and the video literally just freezes halfway through, so this one got screwed up for some reason. Here's the third wolf howling at a moon, which looks pretty good, but the background is red and you can't really see the moon. And here's the fourth video, which looks really realistic to me, but once again, the moon isn't really in the shot. But

compared to Sora, giving Sora that same really basic prompt, it just gave me a frozen picture with a parallax effect on it. Gabo here gave me this long prompt. I'm not going to read the whole thing, but basically it's about two plant seedlings growing taller together and very close by. I would say Veo kind of struggled with this one. We can see some plants blooming here, but yeah, they're a little bit funky. Here's another one showing the plants blooming, but not much

Else, so it sort of ignored a lot of the other details that were put into the prompt. Here's the third video where the plant—you can't even see it, cut off at the top—but you can kind of see some movement here in the soil, and then

finally, it flashes a flower at the end. This one I actually thought was cool looking because it kind of looks like an After Effects animation showing two plants growing, but it also really didn't follow the prompt very well because there's nothing really growing. Here's that

same prompt when given to Sora—also not great, so neither really did that one too well. Colton Batz here said a vaporwave-styled cinematic wide shot of a robot detective under a street light and it's raining. I actually think these generations came out pretty decent. I've

got all four of them up on the screen right now so you can compare them. Here's Sora's version of that same exact prompt. Once again, I definitely think Veo wins in that competition. Amed Raaza here says Bollywood dance sequence in Alaska featuring grizzlies and polar

bears. I've got all four of those up on the screen right now. I actually think these came out really, really good. I think some of the faces don't look amazing, but I was pretty impressed with how this came out in general. When

you compare it to Sora, here's what Sora gave us. Again, the 480p that I generated isn't really doing us any favors, but it's not horrible as well. There's even less detail in the faces, but I think both of them did this prompt pretty decently.

I'm honest, this was a really fun one from Jake. Dan here, a punk rock squirrel in a studded leather jacket shouted into a microphone while standing on a stump. This one was fun, and I think a few of these actually look pretty solid. Here's our comparison of Sora giving

that same one. The one on the left almost looks like a little animatronic squirrel or something. The one on the right, the squirrel's just like eating the microphone. Still, pretty funny videos though. Andre suggested creating a rough seas and stormy weather lighthouse in

the middle of the sea, big tall tsunami waves, cloudy and galaxy sky, realistic minimalist real. So that's what you're seeing on the screen right now. I'm just putting all four of them up, and here is that exact same prompt when given to Sora. I actually think the one on the

right for Sora doesn't look too bad, but I also think that Veo definitely nailed it a little bit better than Sora did. M.L. Laac here says a spider Kaiju sipping tea with a dragon Kaiju over the burning remains of a city they just destroyed.

Now, the spider part wasn't too great, but it seems to have gotten the Kaiju dragon, and it also seems to have gotten the destruction part. Comparing that to Sora, here's what we got. It got the spider, but it didn't really get the dragon portion.

It also did get the city destruction pretty well. Ovium Gamer said cat riding a unicorn through space, and what you're seeing right now is the cat riding a unicorn through space from Veo. They're okay, not too bad, but let's go ahead and compare that to what Sora gave us.

I actually really like the one that Sora generated on the right. I think that's kind of hilarious. It also didn't give the unicorn wings, whereas Veo, for whatever reason, gave the unicorn wings. Restructuring democracy suggested a cat wearing a cat suit, eating a cat in its mouth, standing in a room decorated with cat pictures and full of cats eating dogs. I don't know if it really nailed the last part about the cat eating dogs, but this is what you get from it, and here's our Sora version for

comparison. Once again, no cats eating dogs, but it's a pretty silly little animation nonetheless. After Light asked for a person solving a Rubik's Cube, it really seems to struggle with the Rubik's Cube motion and the colors actually changing properly, so I wasn't super impressed by

what it managed to generate with this prompt. But comparing it to Sora, I think it actually did a little bit better than Sora because in Sora, one guy is just touching it and the colors are changing, whereas Veo actually looked like they were

doing the motion properly. Gooey here said two astronauts playing ping pong on the moon while Earth explodes in the background. Well, we've got the astronauts, we've got a ping pong ball, but it doesn't really look like they're interacting with it properly, and you can definitely

see Earth exploding in the background. And here it is compared to what Sora generated. Sora has less of the explosion in the background, and they actually look a little more like astronauts. They're actually wearing the proper space suits and everything. Neither of them really nailed this one, let's be honest. I also

I saw this prompt from my buddy Baval Sadu: a side-scrolling shot of a rhinoceros walking through a dry field of low grass. You can see his comparison with Veo and Sora here, and then I went ahead and did my own version of that.

I got this one and thought it was like the most realistic AI-generated rhino I've ever seen. It is really, really good. Here’s some other footage that I tested. These weren't from actual recommendations; they were just random prompts that I tested for myself: CCTV

footage of an alien walking in front of a convenience store. We can see the four videos that it generated there. None of them really scream CCTV footage to me, especially since the camera is moving, but still pretty cool. Here’s a first-person view out the window of a jet

fighter. I thought all of these looked really, really good. The one on the top right looked especially good. I was so impressed with that found footage in a dark forest with a mysterious monster, and for whatever reason, it only generated three out of the four and didn't generate the fourth one. I really

like the found footage up on the top right. I thought that one came out really well. Here’s one of the image-to-video videos that I made with the circus tent. Once again, it only generated two out of four. I don't know why, but the two that it generated

were both pretty impressive-looking establishing shots that could easily be used for a movie about the circus. So there you have it: a sort of deep dive into Veo and what it's capable of. Again, I tried to test as many prompts

As I could, we'll probably play with it even more on one of my Monday live streams. If you're not familiar, I am doing live streams every Monday at 11:00 a.m. Pacific Time. If you want to test out Veo, we can do that on the next

live stream, and I will take your prompts live. We'll generate them on that stream and see what we come up with. But just quickly giving my thoughts, I think that we have a new standard for AI video. I think Veo is much better than what

we're getting out of Sora Turbo right now. We haven't seen the full version of Sora; the Turbo model is a sort of quicker processing model. It doesn't generate the greatest results, so I think OpenAI still has a better Sora model.

But of what's available for people to use right now, Veo is definitely winning. And yes, a lot of the generations you're seeing online are cherry-picked, but Veo is kind of designed for cherry-picking, right? You give it a prompt and it's

going to give you four videos. Everything I showed you was the four videos that it generated. I never generated any of these prompts twice; it gives you four videos, and then you cherry-pick from the four that it generated. It does still tend to struggle

with people doing gymnastics and a lot of really complex prompts with things supposed to be happening in the foreground and the background, like we saw with the guy playing with his dog and the dolphins. It almost focused on just one or the other, and as we

I always say, and I know this is cliché, but this is as bad as it's going to get. This is a really, really impressive model. I'm grateful that Google gave me early access so I can demo it, play around with it, and show it off to you, but AI

video seems to be jumping by leaps and bounds, and it's only a matter of time before this is publicly accessible to everybody. And man, that next version of Sora and that next version of Veo are probably going to really blow some people's minds, because this one is already

starting to blow people's minds. So that's what I got for you today. That's my breakdown of Veo. Again, we'll probably do more tests on a future live stream, so make sure you join me for one of those. But I really appreciate you tuning in

and checking out what this new technology is capable of with me. It's been really, really fun to play with, and I'm excited to dive deep into even more tools like this in future videos. So if you're not already subscribed, make sure you subscribe to the channel. Thank

you once again for tuning in. I really, really appreciate you. I'll see you in the next video. Bye-bye!

This New AI Video Tool is The Best I've Seen! (Hands on with Veo)

AI, No-Code, Tech, Futurism - I'm a tech nerd and talk about tech nerd stuff

Matt Wolfe

Google의 새로운 AI 동영상 생성 플랫폼인 Veo를 소개하고 Sora와 비교합니다.

실제 동영상과 AI가 생성한 동영상을 구별하는 것이 점점 더 어려워지는 것에 대해 논의합니다.

Google의 Veo 출시에 대해 설명하고 대기자 명단에 대한 URL을 제공합니다.

Veo의 두 가지 옵션인 텍스트-투-비디오 및 텍스트-이미지-투-비디오에 대해 설명합니다.

얼리 액세스 및 플랫폼의 잠재적 변화에 대한 개인적인 경험을 공유합니다.

사용자가 제출한 메시지를 포함한 다양한 프롬프트를 사용하여 Veo의 동영상 생성 기능을 시연합니다.

자세한 예제 프롬프트와 생성된 동영상을 제공합니다.

생성된 동영상을 분석하여 사실성과 간혹 프롬프트의 잘못된 해석에 주목합니다.

텍스트-이미지-동영상 변환 기능과 그 한계를 살펴봅니다.

소셜 미디어에서 프롬프트를 수집하고 Veo에서 테스트하는 방법에 대해 설명합니다.

사용자 프롬프트에 따라 생성된 다양한 동영상을 Sora의 결과물과 비교하여 보여줍니다.

Veo에 대한 생각, 향후 계획 및 행동 유도 문안을 마무리합니다.

Veo의 기능을 요약하고 향후 라이브 스트리밍 데모 계획을 설명합니다.

베오와 소라 터보를 비교하고 향후 버전의 잠재력에 대해 논의합니다.

현재의 한계를 인정하면서도 AI 동영상 생성의 인상적인 발전을 강조합니다.

시청자가 향후 라이브 스트림을 구독하고 참여하도록 유도합니다.