구글 딥마인드의 텍스트-비디오 모델 v2: 소라보다 낫다고? 놀라운 사실감과 물리학을 보여주는

### 구글 딥마인드 v2: 텍스트-비디오 기술의 판도를 바꾸는 기술
* 🎥 구글 딥마인드는 이전 모델인 소라보다 뛰어난 사실감과 물리학을 보여주는 **텍스트-비디오 모델인 v2**를 공개해 호평을 받고 있습니다.
* ✨ 첫 번째 예시에서는 **여성 DJ가 등장**하며, 생생한 네온 컬러와 사실적인 카메라 움직임으로 매혹적인 존재감을 강조한 영상이 눈길을 끕니다.
* 📸 v2는 사용자가 **샷 유형과 움직임**을 지정할 수 있어 이전 모델에 비해 **출력 품질이 크게 향상**되는 등 세부적인 카메라 설명에 탁월합니다.
* 📊 벤치마크에 따르면 **v2는 상당수의 사용자가 소라 터보와 같은 다른 모델보다 선호도가 높음**에 따라, 그 결과 상당수의 사용자가 선호하는 것으로 나타났습니다.
* 👑 르네상스 시대 궁전의 여왕을 보여주는 한 가지 놀라운 예는 **거울에 비친 정교한 디테일과 반사가 매혹적인 분위기를 조성**하여 풍부한 내러티브를 만들어내는 모델의 능력을 보여줍니다.
* 🐝 이러한 발전에도 불구하고 체조 선수와 스케이트보더의 예에서 볼 수 있듯이 **사실감이 떨어지는 고난이도 동작**은 여전히 v2에서 해결해야 할 과제입니다.
* 🍳 이 모델은 **시럽이 쏟아지고 수증기가 피어오르는 물리학이 아름답게 렌더링되는 아침 식사 장면**에서 빛을 발하지만 일부 디테일은 놓쳤습니다.
* 💻 이 비디오의 파트너인 커리어리스트는 기술 분야의 부트 캠프를 제공하며, **소프트웨어 QA 역할에 대한 수요가 증가**하고 있으며, 이는 **수익성 높은 연봉**으로 이어질 수 있어 업계에 진출하려는 사람들에게 훌륭한 기회가 될 수 있다고 강조합니다.

구글 딥마인드의 v2 텍스트-비디오 모델은 뛰어난 사실감과 물리학을 자랑하며 경쟁사를 능가합니다.

Google DeepMind just announced v2, and it is a text-to-video model that many people are saying is better than Sora. Looking at the results, it actually might be; it's very impressive. Let me show you. All right, so check out this first video. What I've noticed about v2

is that the quality of the realistic video is astounding. And not only that, the physics are really, really good—much better than Sora, even. So here's the prompt: An extreme close-up shot focuses on the face of a female DJ. Her beautiful, voluminous, black, curly hair frames her

features as she becomes completely absorbed in the music. Her eyes are closed, lost in the rhythm, and a slight smile plays on her lips. The camera captures the subtle movements of her head as she nods and sways to the beat, her body instinctively responding

to the music pulsating through her headphones and out into the crowd. The shallow depth of field blurs the background; she's surrounded by vibrant neon colors. The close-up emphasizes her captivating presence and the power of music to transport and transcend. That is

exactly what we're seeing in this video—the quality, the eyes; there is no uncanny valley in this video. And what v2 does really, really well, apparently, and that's what we're seeing, is camera descriptions. You can tell it the type of camera,

the type of camera movement, the lens type—everything. Here's another one. This is more of a cartoon theme. This medium shot, with a shallow depth of field—and pausing for a second, that's an important part—you can describe the type of shot much more accurately than you could with

Sora portrays a cute cartoon girl with wavy brown hair, sitting upright in a 1980s kitchen. Her hair is medium length and wavy. She has a small, slightly upturned nose and small, rounded ears. She is very animated and excited as she talks to the camera, and here is probably

one of the most impressive ones. So this is a car driving through the street. Reflections everywhere! Car burning out; we're seeing smoke come out of the wheels. The smoke looks completely real. Here's the prompt: Low-angle tracking shot, 18mm lens; the car drifts, leaving

trails of light and tire smoke, creating a visually striking and abstract composition. The camera tracks low, capturing the sleek, olive-green muscle car as it approaches a corner. As the car executes a dramatic drift, the shot becomes more stylized, and it's that type

of description where, again, you're describing the type of scene that you want to see, the type of shot, the type of camera. This is all stuff I haven't seen before, and we have some benchmarks, and I'm going to show you some more examples, but let me talk about the benchmarks

really quickly. So here, here we go. This is basically human preference, and we have four different text-to-video models: We have Meta's Make-A-Video, Gen-2, MiniMax, and Sora Turbo. And where you see this green is where Vo is

preferred over the other models. So you can see, and surprisingly, compared to Sora Turbo, it has the largest preference percentage. All right, you want to see another wild example? Check this out! This is a reflection in the mirror that is absolutely wild! This woman is looking

Into the mirror, and you can actually see the reflection perfectly. As the camera pans around the back of her head, the reflection in the mirror looks flawless. Let me read the prompt: As you're watching, the camera moves in a slow dolly shot, revealing the opulence

of a Renaissance palace chamber, adorned with gold inlaid furniture, velvet drapes, and chandeliers casting soft, flickering light. A queen sits motionless at a gilded desk; her crimson silk gown cascading on the floor like spilled blood. Now, we don't really see that—we

see a little bit—look at that, a little bit under the actual person, just a glimpse of it—but it's there. On the desk lies an unsigned letter; its edges curled with age. There it is; you can kind of see it's a little wavy.

The camera frames her from behind, catching the reflection of her stoic face in a massive, ornate mirror in the background. Couriers murmur; there are silhouettes dancing like ghosts in the candlelight. The room feels heavy; every gilded detail amplifying an air of betrayal and

paranoia. The color palette alternates between deep, regal reds and cold golds, with chiaroscuro—which hopefully I'm pronouncing that right; I feel like I'm not—lighting intensifying the drama. Shot on 70mm film for rich texture, evoking

the grandeur of historical masterpieces. Really beautiful! The level of detail, the reflection, the physics—it's all so good. Thank you to Careerist for partnering with me on this video. I've worked with Careerist before, and I'm excited to tell you about them again. Careerist provides

accessible and interactive online boot camps to help individuals break into tech, specifically software QA engineering. These courses are tailored for people with no tech background, offering mentorship, hands-on practice, and career advice. With over a thousand graduates employed across 42 states, Careerist is a trusted platform for

entering the tech industry. The tech industry offers a lot of benefits, including really competitive salaries, high demand for professionals, and plenty of opportunities for growth. Obviously, I've been in tech my whole life and I love it. These tech roles can offer salaries of up to $113k per year with no

coding experience required. Quality assurance is going to be more in demand than ever because of artificial intelligence. All of this code being written by AI needs to be checked thoroughly. Plus, there's a wide range of different projects you can work on, and it is very remote-work friendly, so

invest in yourself. Check out Careerist; they're offering $600 off exclusively to my viewers, so use the promo code Matthew to get that $600 off. I'll drop all of the links down below. Please check them out if you're considering getting into tech. Definitely just get in touch with

them and check it out and learn more. And thank you again to Careerist for partnering with me on this video. Now, back to the video. All right, here's a limitation, and I've actually seen a bunch of examples of this type of video.

Let me show you the Sora version of it first. This is a gymnast where she's performing some gymnastics, but her legs and her torso do not look connected at all. It's kind of silly, and we're seeing the same thing here. So now this is V2.

This is a woman ice skating, really performing a beautiful dance. But you could tell her torso kind of gets separated from her legs; her legs spin around, they disappear for a little second. It looks like she would be paralyzed after doing these moves, so definitely not real, but still impressive.

And here's another example. It seems like highly physical movements are still a big issue for text-to-video models. This is obviously a skateboarder doing moves, but as soon as he takes off the ground, all of the physics go out the window.

It still looks really cool, in my opinion, but it still doesn't look realistic. He's doing moves; his legs are flung all over the place; it gets a bit blurry; it just doesn't look good whatsoever in terms of realism. All right, here's another incredible example of realistic physics.

Both liquid physics and thicker liquid physics, but check this out: The sun rises slowly behind a perfectly plated breakfast scene. Thick, golden maple syrup pours in slow motion over a stack of fluffy pancakes, each one releasing a soft, warm steam cloud. A close-up of

crispy bacon sizzles, sending tiny embers of golden grease into the air. We don't necessarily see that, but it still looks good. Coffee pours in smooth, swirling motion into a crystal clear cup, filling it with deep brown layers of crema. The scene

ends with a camera swoop into a fresh-cut orange, revealing its bright, juicy segments in stunning macro detail. So we don't see the fresh-cut orange; I don't see that anywhere, so it didn't pick up some of the details of the prompts, but

Still beautiful output, and check this out: this one is a beekeeper, and he looks absolutely real. The movements of the bees look completely real. Let me read this prompt as you watch the video: The camera floats gently through rows of pastel-painted wooden beehives, buzzing

honeybees gliding in and out of frame. The motion settles on the fine farmer standing at the center; his pristine white beekeeping suit gleaming in the golden afternoon light. He lifts a jar of honey, tilting it slightly to catch the light. Behind him, tall sunflowers sway

rhythmically in the breeze. We're seeing all of that so far; their petals glowing in the warm sunlight. The camera tilts upward to reveal a retro farmhouse with mint green shutters. I don't see that at all.

Shot with a 35mm lens on Kodak Portra 400 film. That type of description of the lens, the film type, the camera type—that's only stuff that I've seen in text-to-image models, and specifically stuff like Midjourney. Really impressive that that can translate to video models now. The golden

light creates rich textures on the farmer's glove, marmalade jar, and weathered wood of the beehives—all of which we are seeing there. So here's what Google DeepMind is showing off as their kind of core differentiating features versus other video models: enhanced

realism and fidelity, significantly improved over other AI video models in terms of detail, realism, and artifact reduction. Advanced motion capabilities: it represents motion to a high degree of accuracy thanks to its understanding of physics and its ability to follow detailed instructions, and greater camera

Control options interprets instructions precisely to create a wide range of shot styles, angles, movements, and combinations of all of these. All right, and let me show you a few more examples that I found. Thanks, Bill Wall, for showing me this. So this is VO2 versus Sora, so

this is going to really show you the difference in quality. So here is VO on the top, a sidescrolling shot of a rhinoceros walking through a dry field of low grass plants. Now on the top, what you're seeing is VO, which I'd say this

is really good. The body is kind of rigid and it looks obviously fake, but compared to Sora, which actually the movements of Sora look realer, but all of a sudden the physics completely break down about halfway through the video and it just seems like the rhinoceros is walking on the

moon. So definitely weird, but I would give the win to VO in terms of realistic movements. For the first few seconds, Sora might have the win, but overall VO is better. So the next example is Matt Wolf, and this one looks phenomenal. This is VO2

and this is straight out of a nature documentary, as Matt Wolf says. I mean, I cannot really find any mistakes with it: the ears, the horn, the movements, the physics around it, it all looks really good. All right, so here's another one: horse running, beautiful horse running

down a path, goes right through the fence, but overall really good. Here's another one: horse coming towards the camera, not showing the full body. This one looks more real. Here's another one: horse running down a tiled path, and yeah, it looks pretty good. I think the horse has

Slightly unrealistic movements, but okay. And here's another one. I think all of them actually look quite good. I don't see anything wrong with the legs; the movement... trying to find anything wrong with it, but it just looks really good. So here's a couple from the HighLow video.

model, just to show as comparison. Here's one; I think it looks okay, definitely not nearly as good. The galloping doesn't look as realistic. Here's another one, same one... this one looks horrible; they're kind of floating; one kind of morphs into

the Mustang car... just really bad. Here's one from Sora Turbo. Definitely the legs are kind of shifting and moving like they wouldn't. I don't think rhinoceroses have that type of dynamic movement, so not great. And the ears seem to morph a

little bit if you just watch as it's moving... just very awkward. So here's a pelican riding a bike; obviously not very real at all, but still I think it looks good. This car driving down the street, obviously in a mountain road, and now going off a waterfall. I think this looks

fantastic; obviously not real, but it looks real. This shot is incredible. This is the future of video games right here. This is kind of a market bizarre in the sky. Here's one that could be a Pixar movie. All of these I think are just stunning. Here's another one where

there's a car drifting, kind of what we've seen before. Obviously, look at this! Look how impressive this is! The license plate says 'VO' and doesn't move at all; it's just flawless. Here's another with some potatoes as chefs, and then look at

This is a piece of fabric falling on a piece of marble. Just stunning physics! I don't see anything wrong with it. All right, the next one is a woman playing a guitar with a cat sitting next to her. A man walks right by and, as

it covers up the woman and the cat, it reappears after he walks by and just is perfect. I wouldn't be able to tell that this was AI-generated. Here's an octopus on the ocean floor. Again, the physics just look phenomenal; a little bit of

morphing. Here are some cards being shuffled together. This looks okay, not great to be honest. The fingers actually look really good. Here's a cat running, jumping on something. Okay, tell me this isn't flawless! I don't see a single thing wrong with it. Here's a couple

of giraffes running. Yep, again, very flawless animal movement. It seems to have nailed it perfectly. Now, this is the one that people are beyond excited about: the physics in the cutting of this tomato are flawless. Look at this! Look at that! All of the reflections,

the physics of the tomato moving—so impressive! So that's it. Let me know what you think. It's pretty darn impressive, right? I am beyond blown away! Let me know what you think in the comments. If you enjoyed this video, please consider giving a like and subscribing, and I'll

Google's Veo 2 - STUNNING AI Video (gg Sora)

Artificial Intelligence (AI), Open Source, Generative Art, AI Art, Futurism, ChatGPT, Large Language Models (LLM), Machine Learning, Technology, Coding, Tutorials, AI News, and more

** Exclusive Pinecone Partner **

Matthew Berman

구글 딥마인드 v2: 인상적인 텍스트-비디오 모델

품질 면에서 소라를 능가할 가능성이 있는 텍스트-비디오 모델인 Google DeepMind의 v2를 소개합니다.

소라에 비해 놀랍도록 사실적인 영상 품질과 뛰어난 물리 효과를 제공합니다.

카메라 설명(유형, 움직임, 렌즈)에 대한 탁월한 처리.

다양한 촬영 유형(예: 미디엄 샷, 얕은 피사계 심도)을 정확하게 묘사합니다.

반사 및 연기와 같은 효과로 복잡한 장면을 사실적으로 렌더링합니다.

인간 선호도 테스트 결과 V2가 소라 터보와 같은 다른 모델보다 성능이 뛰어난 것으로 나타났습니다.

거울에서 완벽한 리플렉션 렌더링을 보여주는 예시입니다.

복잡한 장면과 조명을 캡처하는 능력을 보여주는 자세한 프롬프트.

기술 경력을 위한 커리어리스트의 온라인 부트캠프 광고.

멘토링 및 커리어 조언을 포함한 커리어리스트의 과정에 대한 자세한 내용입니다.

경쟁력 있는 급여와 원격 근무 옵션 등 기술 커리어의 이점을 강조합니다.

AI로 인해 증가하는 QA 엔지니어 수요에 집중하세요.

시청자를 위한 프로모션 코드 제공.

v2의 한계 고도의 신체 움직임 관련 문제

고도의 신체적 움직임(예: 체조, 아이스 스케이팅)을 사실적으로 묘사하는 데 어려움이 있음을 보여주는 예시입니다.

일부 생성된 동영상에서 신체 부위 간 연결이 끊어집니다.

역동적인 움직임이 포함된 장면의 물리 문제.

액션 장면(예: 스케이트보드 타기)에서 비현실적인 피직스의 추가 예시입니다.

V2의 사실적인 피직스 및 디테일 예시

사실적인 액체 물리학(메이플 시럽, 커피)의 쇼케이스.

사실적인 요소로 아침 식사 장면을 자세히 설명합니다.

벌의 움직임이 사실적인 양봉 장면의 예시입니다.

카메라 및 필름 유형을 포함한 자세한 프롬프트가 표시됩니다.

향상된 사실감과 충실도, 향상된 디테일, 아티팩트 감소.

정확한 물리학 및 세부 지침을 따르는 고급 모션 기능을 제공합니다.

다양한 촬영 스타일과 움직임을 위한 카메라 제어 옵션이 향상되었습니다.

추가 비교: V2와 다른 모델 비교

v2와 소라를 직접 비교하여 품질과 사실감의 차이를 강조합니다.

코뿔소 장면을 분석한 결과, 약간의 강성에도 불구하고 V2의 뛰어난 사실적인 움직임을 확인할 수 있습니다.

자연 다큐멘터리급 동영상을 제작하는 v2의 기능을 보여주는 Matt Wolf의 예시입니다.

움직이는 말의 여러 예시를 통해 일반적으로 훌륭하지만 때로는 약간 비현실적인 움직임을 강조합니다.

하이로우 및 소라 터보 모델과 비교하여 v2의 우수성을 보여줍니다.

비디오 게임과 애니메이션에서 v2의 잠재력을 보여주는 환상적인 장면의 예시입니다.

디테일한 피직스가 적용된 인상적인 장면(예: 천이 떨어지거나 토마토가 잘리는 장면)을 보여줍니다.

매끄러운 전환과 사실적인 동물 움직임(예: 고양이, 문어, 기린)의 예시입니다.