Create Cinematic Ai Videos with Google VEO 3 (FULL COURSE)

0:00 Google Veil 3 has completely changed the AI filmmaking game with its ability to create cinematic AI videos along with realistic sound effects and character voices all inside one platform. This will be the last time you fail me. I know it's frustrating to watch tutorials which show clips of highquality videos, but when you go and try to create the AI videos yourself, you get lower quality clips like this one that look like PowerPoint slides. That's why I made this video specifically for people like you who want the real advice. Not just a surface level explanation, but a deep dive into practical tips that will actually help you with your AI films. May the force be with you.

0:46 Google Veil 3 is accessed through Flow, which is Google's new film making platform. So, let's go ahead and launch it. Once we're inside, we can see a list of our past projects. I'll go ahead and create a new project right now. From here, we can start creating our AI videos. There's a couple different options. Text to video, frames to video, ingredients to video. We're going to cover all of them, but let's start with text to video where we'll just enter a prompt and it'll create a video for us. What's new in Google Veil 3 is the ability to directly generate talking characters. So, I'll create a video of a Jedi from a Star Wars movie. I'm going to describe the appearance a little bit. She has green skin, these geometric tattoos across her face. Um, also give

1:33 her a yellow lightsaber, and we're going to tell her to talk about how to use the force inside this AI video. Let's go ahead and generate that. When you're generating the video, you need to make sure you've actually selected the newest Google Vo3 model. If we go to the settings tab right here, you can change the number of outputs per prompt. I've set that to one to generate one video at a time. But down here in the quality, make sure you've selected the highest quality with experimental audio. That's the newest Veil 3 model. The fast and quality options default to the older Veil True model, so you won't get the best looking videos if you use these. Here's what the video turned out to look like. The Force is a subtle

2:18 energy. You have to feel it, not just see it. What's impressive about this is that it animates a talking speaking character just from a text prop. And the quality of the lip sync and the emotions in the video are very high. We can also make the AI say specific words from a script if we want. I'm using the same prompt of a female Jedi from Star Wars with green skin. I've described her appearance a little bit closer. So this time I also told her the specific clothing to wear, which is a lightweight brown and gold robe. And she's also meditating inside a quiet swamp illuminated by bol luminescent plants on an alien world. This time I'm specifically prompting for her to say, "The force doesn't scream, it hums. Listen and you'll know which path is

3:05 open." The force doesn't scream, it hums. Listen and you'll know which path is open. Look at the way the hand motions are joined together with the speech and the facial expressions and the wrinkles on the face all go together very very well. Now I want to go a little more in depth on this. How much control do we have over the voices? Can we actually change the way the character's voice sound? So I rewrote this prompt. This time I told her to have a highpitched and nasal voice which should sound like she kind of has a stuffy nose a little bit. So then if we generate the video, here's what that character sounds like. The force doesn't scream, it

3:52 hums. Listen and you'll know which path is open. It's a great looking video, but it didn't actually follow my prompt of giving it a high-pitched and nasal tone to her voice. The way Google Video3 generates the voices is dependent on the way that the characters look. I generated a video of a Sith. He's an older weathered man. And if we listen to his voice, it's going to have a deeper, grittier tone to it. Pain is not punishment. It is revelation. Feel it and you will understand me. Depending on the video of the character that you're generating, Veil 3 will match the voice to the way that character looks. Target acquired. Regret will be brief.

4:42 But unfortunately, there isn't a way to actually change the way the character sounds inside the prompt. This might not seem like a huge problem, but sometimes it does generate voices that don't match the way the characters look at all. Luke, I am your father. But sticking with the theme of characters, can we actually generate videos of the same character inside different scenes? The answer is yeah, you can get pretty darn close as long as the description of the character's appearance matches closely. I've written down the way her skin looks, the patterns on her face, also the clothing that she's wearing. When we go and generate a new video of this character using text to video, it'll actually give someone that looks really, really close.

5:28 Control is not the absence of fear. It's the choice not to let fear decide. We can render them inside different scenes as well. We've spent so long guarding peace that we forgot how heavy peace can be. The more specific we are in the prompt in describing the way the character looks, the more consistent they'll look in the videos that we generate. For this prompt of the Sith, I'm extremely specific in the way that he looks. He's tall. He's lean. He has pale white ashy skin and dark veins on his gaunt face. I also specify the hairstyle. He has gray hair tied in a bun, and he wears a raggedy black cloak with sharp edge plating on the shoulders. Here's a bunch of videos I generated of this specific Sith Lord. The main thing you'll notice that is

6:15 kind of different is the shoulder pads he's wearing. For some of them, they look like plates. For other ones, there's some sharp needles pointing out, but we could have just removed that part of the prop that talks about the plating on his shoulders. At this point, I'm sure you've noticed that Veil 3 also generates sound effects that go along with the characters, the lightsabers, the steps that he's taken as he's walking. Although, from what I can tell, you don't actually have that much control over what specific sound effects are generated. The amount of voice control you have is limited to a certain extent. For example, in this prompt of the Sith, I specifically said that he reaches out with his hand and yells no. I want him to be kind of screeching out and screaming with a very exaggerated facial expression. No. Oh,

7:09 okay. So, we do hear him screaming a bit at the end, but the part where he says no happens at the beginning, and we don't even see his mouth moving along with that word. So, for super super expressive voice control, it doesn't seem like it's quite there just yet, but the video animations do look incredible. If you want to adding the sounds later in by yourself, these are 8second videos, by the way. There's a few different options to download, you can get a GIF, the original size, or upscale it to high definition. I found that the upscaler did seem to get stuck a bunch of times, so maybe it's not working correctly right now. We can also add in multiple characters into the same scene and have them interact with each

7:54 other. In this prompt, I copied in the full description of the female Jedi and I also copied in the full description of the Sith. And I have them talking to each other. So, the male Sith says, "Tell me Jedi, how many died while you meditated?" And the female Jedi says, "And how many screamed while you searched for meaning in their suffering?" You can't specifically tell each character the script that they're going to be saying inside the video. Tell me, Jedi, how many died while you meditated and how many screamed while you searched for meaning in their suffering? And using this same technique, we can create the two of them interacting together, for example, in a lightsaber battle.

8:44 So, this battle isn't exactly as dynamic as a real movie. This video clip does start off a little slow. There's not that much motion at the start, but as we get later on into the video and they start swinging their lightsabers. It's a pretty epic looking battle. I mean, there's sparks flying everywhere. The sound effects also sound realistic to me. So, what about when it comes to directing the camera inside the AI videos? You do have quite a bit of control over how the camera motion moves. To start off, let's go with a shot of a Sith temple. We're going to prompt for a slow crane shot behind a cloaked figure walking up the steps to an ancient Sith temple in Star Wars. The cameras flies above and slowly tilts down to reveal the Sith temple. Red

9:30 lightning inside a thunderstorm. So, what happened here? That wasn't exactly the video that I prompted for. The Sith is actually walking down the steps instead of walking up the step like I asked for inside the prompt. Just like any other AI image generator or AI video, there is some randomness in VO3, which means every time you run the same prompt, you won't get the same results. So, sometimes it's a good idea to run the prompt multiple times. Here's another video generated of me running the exact same prompt again. And this time, we'll see him walking up the steps like I asked

10:17 for. There's some pretty cool camera movement effects you can create. Here's an example where I asked for the video to start with a close-up shot on a gloved finger hovering over a blaster. And then the camera is going to tilt up and reveal a bounty hunter with a helmet on. This was one of my favorite video generations. I can't get over how good this looks. In the prompt, I specifically told the camera to quickly tilt up, and you see the movement flashing from the blaster to the helmet of the Dante Hunter. Here's more of an advanced tip. When you prompt for specific subjects in the video, a lot of the times that can override the camera movement and the camera angle that you ask for. For example, I want to generate

11:04 an over theshoulder shot of a rebel flying inside a spaceship. I want it to be looking outside the cockpit window where we can see other fighter jets flying by. Here's the overthe-shoulder view that I'm actually looking for. The camera is placed behind the pilot. We can see his shoulder and the back of his head, but we're looking outside the window. If I run that prompt inside of V3, though, you'll see that the video actually starts with a medium body shot showing the pilot's face and body before switching to an over-the-shoulder shot. That's not what I asked for. The reason is because we included the words the rebel pilot inside the prompt. And

11:50 so Veil 3 wants to show that subject in the video. However, if I remove that phrase, rebel pilot, and just say over the shoulder shot inside a cockpit of a spaceship, here's what we get. Because I didn't add the subject of rebel fighter inside the video, it prioritizes the camera angle that I'm asking for and gives me a perfect overthe-shoulder shot for the entire video. Here's another clip I generated where I asked her to start from the feet of stormtroopers marching through a town and then pan up to reveal the body and helmets of those stormtroopers. What are the limitations of what it can do as far as camera

12:36 motion is concerned? If you try to add in too much movement, the quality of the videos will degrade. In a prompt, if I ask for it to start with a shot of a Star Wars droid making some beeping noises and then pan towards the Jedi character that we described earlier, what's going to happen is that there's just too much going on inside this prompt for it to animate it properly. Let's take a look. I know, I know, quit your whining. While the emotion does look pretty good, the voice gets mixed up. It doesn't read the words that I asked her to say properly. I know. I know. Qui, quit your whining. Here's another example where I asked for

13:21 the Sith to reach out with his palm, say some words, and then use the force to send a bounty hunter flying through the air. This will be the last time you fail me. This video doesn't make any sense at all. It looks like the Sith does some kind of a 360 flip and then the bounty hunter appears out of nowhere. So, if you want to create more dynamic scenes like this, it's better to animate it in smaller chunks separately. So, I'll create a video of the Sith reaching out with his palm and then a separate one for the bounty hunter. This will be the last time you fail me. Think about the exact textures,

14:07 patterns, and materials that you want inside your videos. Here's a video I generated of our Jedi earlier, and she's riding on top of a speeder. But the chrome metallic silver look on the speeder, it doesn't fit a Star Wars movie. So instead, in this prompt, I specifically told the AI video generator the textures of the speeder that I want should be rusty, made out of old metal scraps, and have some plants growing on the side of it. And we get a scene that's much more visually aligned overall. So far we've covered text to video, but we can also do frames to video in which case we upload a reference image first and then Google will create a video for us based on that reference image. You might be thinking

14:53 if the text to video is so good, why do we still need image to video? The reason is because text to video is limited a lot of the times in its ability to generate the specific characters or scenes you want. For example, this is a character from Star Wars called Jar Jar Binks. If you're not familiar with him, he's basically an alien that looks like this guy. Kind of a pretty well-known character. But if I try to generate him inside Google V3 using only the text box, it gives me a character that doesn't look like Jar Jar Binks at all. Twist in the minds. You are just too blind to see. And that's where frames to video comes in. So let's go to frames to video and we'll have an option to upload

15:40 some images. So we can upload our image or generate an image directly inside 3. So let's generate an image and enter a prompt. Jar Binks from Star Wars as a Sith. He's wearing a black cloak and holding a red lightsaber cinematic film. Then we can run this prompt and create the images and it'll generate a character image that looks much much closer to Jar Jar. We can go ahead and use this image to generate a video of it. And inside the prompt, we can ask for anything. For example, the camera zooms zooms in as he waves his

16:28 lightsaber. Let's try and generate that and it'll create a decent looking video from the reference image that we uploaded. Now, when it comes to using the image frames to video feature, I don't recommend it as much as just trying to use text to video. So, let me explain. There are some decent features you can get from image to video. For example, if we upload a reference image, let's say we'll put in this photo of a spaceship. We do have the option of adding in some camera movements, but there's a bunch of options here. You can dialing in on the character. You can tilt the camera down. Um, there's a bunch of them. I'm not

17:13 sure why they're not all displaying right now properly, but let's go with orbiting left around the subject. and let's put in the prompt the camera circles around the spaceship. Now, if we look inside the settings, you can see that I have selected the Google V3 model for the video generation. However, if we try to actually run this prompt, you'll see this pop up which says switching you to a compatible model for this feature. This is because a lot of the imagetovideo features only work with the Ve2 model. For example, adding in the camera motions, which means you won't be able to use the most advanced video model for

18:00 this feature. But let's generate it anyways and see what it looks like. The camera movements do work pretty well. The only thing is that, like I said, the video model isn't going to be the newest V3 model. Depending on the scene that you need, I think a lot of these will work perfectly fine. The camera movements do look pretty smooth and follow what you ask for. But if you want to use camera motions with the new Veil 3 model, you'll need to directly add them into the prompt by yourself. Here's a example where I told the camera to pull back while the Jedi walks forward and waves her white lightsaber. Here's another example where I prompted for the camera to circle around the subject. You do have a decent amount of

18:46 control over the camera motion in just the prompt itself. Here's one of the camera tilting down while this hunter jumps down from the ledge. Circling back to what I said previously, I do think the text to video works better than using a reference image inside Veo 2. So, if possible, try to create the videos using just the text prompt if you can. So, I've got this reference image of the bounty hunter, and I told him to jump down from the ledge onto the ground. It's kind of a smaller hop, not exactly a dynamic jump from a tall ledge onto the ground. So, instead, I directly used just a text prompt where I described the visuals of the scene. I wanted to start with a

19:32 front view of a bounty hunter from Star Wars. He's wearing a rusty helmet and rusty armor and the camera tilts down as he jumps down from a tall ledge into a rock pit and lands on the ground. It should be a bright sunlit rocky environment with brown rocks and some sand just like the image that I showed earlier. And let's see what the video looks like if we compare it to using a reference image. The motion is way more dynamic and I think the sound effects in general are better. Also, here's another limitation. If you use reference images, you can't actually create the lifelike talking characters.

20:19 So, I started with this reference image of a woman sitting on a jade throne. And I prompted for her to stand up, pick up a lightsaber, and say the words, "May the force be with you." But one problem with using uh image reference is that when you try to animate a talking character, it can make the character move their mouth around like they're talking, but they won't actually say the words you want them to. You do see these subtitles on the screen. This is something that they need to fix cuz we don't need those. However, if you have an image of a character that you really, really want to use, for example, this Asian woman sitting on a jade throne, I mean, I think this looks really, really amazing. You can generate the video inside VA3 and use a lip sync tool on another platform. Uh so I created this

21:05 video of the female Jedi using the image frames to video. She doesn't say anything inside here, but we're going to go to an AI voice generator. I'm using 11 Labs here. And we can enter a speech that we wanted to say. May the force be with you. There's a bunch of voices you can pick from here. I'm selecting Jen. Then we can generate the AI speech. May the force be with you. We have the voice and the video from Google Veil. We can use a tool like Pix for example which does have AI lip sync. Hit this speech button right here. And inside here I've uploaded the video from Google V3 and also the audio file we just generated using 11 Labs. And we can create the lip

21:53 sync. This isn't the ideal solution. I'm sure Google at some point will update their Veil 3 model to also allow you to create talking characters using reference images, but for now, this is kind of what you got to do. And this is what our lipsync character looks like. May the force be with you. Again, let's try to replicate this scene using just text to video. Here's the prompt I'll use. An Asian female Jedi from Star Wars. She has two small ram horns on her head and is wearing a white sleeveless dress with gold ornaments. She sits on a circular jade throne in Saturn Asian Palace. She stands up, pulls out a green lightsaber. She says, "May the force be with you. Star Wars cinematic film." I tried to describe the image I showed earlier as closely as possible. Let's see what the

22:40 video looks like. May the force be with you. All right. So, let's be honest here. That's not a very visually appealing video. She does stand up and swing the lightsaber around and say the words I prompted for exactly, but I don't think this is a cinematic looking shot. I mean, the colors, the horns on her head look kind of weird. And the colors just look really, really bland. So, when it comes to writing the text prompts, sometimes it'll generate a really great looking video with a small amount of effort, but sometimes you need to describe the exact cinematic shot that you want, including the colors and the environment. So, let's update this

23:25 prompt a little bit. This time, I'm going to describe the visual look of the shot a bit more. So, again, we're going with an Asian female Jedi from Star Wars. She's wearing a white sleeveless dress with gold ornaments. She sits on a glowing dark jade throne inside a dark Asian palace. She stands up and pulls out a green lightsaber. Muted dark colors, high contrast between the dark green jade throne and the darker palace with intricate Asian designs and dragon statues behind the Jedi. Now, if we generate the video, I think we'll get something that looks a lot more visually aesthetic and cinematic. May the force be with you. If we compare the two video clips I

24:12 generated, I mean, the before and after is enormous. So, you really need to describe the visual aesthetic you're going for inside the prompt more closely to get the best looking shots. When you're using image frames to video, it does also limit your creativity quite a bit. I'm starting with this image frame of a woman who's holding a lightsaber. And in the prompt, I'm saying the camera pans left and tilts up to reveal a giant AT&T vehicle from Star Wars cinematic film. If we generate this, here's what happens. Okay, so you see that? You see how it suddenly switches from the initial image frame to this like weird animation of her and like a black AT&T

24:57 in the background. But instead, if we just use the text prompt by itself and describe the cinematic shot that we want, a close-up shot on a female Jedi from Star Wars with a white lightsaber on a dusty battlefield. The camera pans left and tilts up to reveal a giant AT&T vehicle from Star Wars. Dark orange dust and color toss. If we generate this without any reference image, the AT&T robot doesn't look exactly like it would inside a Star Wars movie, but you can see that the amount of creativity that you have and the control you have over the cinematic camera movements is way higher when using just the text prompt. There is another feature that I think is

25:44 worth mentioning inside VO3 and that is this ingredients to video feature. Now what this does is let you combine multiple characters or ingredients inside the same scene. So for example, if I go to ingredients to video down here, I can upload three images inside the video generator. And so I've got a couple of different characters. Uh, I have this this woman with blue skin, got this guy, uh, a bounty hunter with red skin, and I'm also going to go with this landscape image. And in the prompt, we can go with the woman with blue skin and the man with red skin walk together on a rocky landscape. using this image def

26:30 feature again. Uh it's not going to let you use the newest Veil 3 model and instead it'll force you to use an older Veo2 model. The quality isn't going to be as high. Uh it's also not going to give you the sound effects. This one was actually particularly bad. So I just ran this prompt again, except this time at the end I specifically told it to give it muted colors. um and also use the words cinematic film and we get something that's a bit more reasonable. So, this is a cool way that you can create videos with multiple of your own characters inside of them. Now, the characters aren't always going to match your reference images exactly. Um, this

27:18 guy with the blue skin doesn't look quite like the reference image that I used, but I do think this is worth mentioning. Now, there are some more features inside Google Flow. You can use add to scene inside of a video. And what this will do is actually let you extend the video if you want. So, here was a clip I generated of the woman uh and the AT&T robot. So, uh we can jump to which is supposed to generate a different angle of the scene. It doesn't actually work that well based on what I've seen. We can extend the video though. The camera passed to an army of storm troopers marching. And again, uh this is

28:04 limited to the Veil 2 quality video. Let's see what the extended clip looks like. And uh the camera does pan to this droid army marching behind. So it's okay. The quality will be a little bit lower in the extended video um because it's using a lower quality model, but it works okay for what it is. Here's a video clip I generated of the female Jedi standing up uh wielding a lightsaber. And this time, let's try the jump to feature, which which should give me a jump cut to a new video clip of the scene from a different camera angle. So,

28:51 The Jedi walks down the hallway uh with her green lightsaber. And let's change this to follow. So we see her from behind. Now let's generate that and see what happens. Okay, so this basically just extended the video. Um, the jump to feature is kind of hit or miss. Sometimes it just extends the video. Um, but I did also try it on this clip of the bounty hunter where we pan from his blaster to his helmet and then I used the jump to feature to generate him walking away from a spaceship. So,

29:37 sometimes it does work. So, I got this email from someone asking me if I could help them create a lawn 1.5 hour AI movie, and they want to create something epic, like a Braveheart type of movie. Now, this sounds pretty cool. So, here's what I responded with. Oh, wow. For a 1.5 hour film, that's something you need a team for. I specialize in creating AI videos, but you need sound design, maybe voice actors, cinematographers, video editors. While AI is very powerful for high quality work, you still definitely need people with traditional film making experience. I actually wrote a longer email with more details trying to explain that you don't want to overestimate what AI can do. And you're going to love this. Here's the oneliner

30:23 response I got. Hi, Veil 3 has all that. Thank you. So, I was a little irritated when I first read this because I don't think this person has experience, you know, working with AI and film. But, it got me thinking and I want to know what you think. Do you agree or do you disagree with me? Do you think that we could create a full Star Wars type of movie using the capabilities that VO3 has? Let me know. I'm really curious to know what you think. As far as the cost goes, it is expensive. It's like 125 a month and then after 3 months that's going to go up to 250. Now, whether the cost is worth it, that's up to you. What do I think? I think nice things cost money and you pay for the quality that you get. Now, there's tons of other awesome AI video generators out there with awesome features like lip sync,

31:10 making longer video clips, things like that. I did a complete breakdown between the other best AI video generators that I've seen and created a same animation using each of them to see which one is the best. If you want to see that video, go watch this tutorial right here.