WEBVTT 00:00:00.000 --> 00:00:03.040 You're staring at a flat 10 second phone video 00:00:03.040 --> 00:00:06.179 of a beach. Maybe you shot it on vacation and 00:00:06.179 --> 00:00:08.339 you've tried every single editing app available. 00:00:08.599 --> 00:00:10.679 Right. And it still looks completely amateur. 00:00:10.839 --> 00:00:14.500 Exactly. It just feels flat. But what if what 00:00:14.500 --> 00:00:17.820 if one simple text prompt could change all of 00:00:17.820 --> 00:00:20.660 that? Yeah. What if you could turn that exact 00:00:20.660 --> 00:00:25.300 clip into a sweeping cinematic drone shot and 00:00:25.300 --> 00:00:28.140 the whole process took under 60 seconds? Welcome 00:00:28.140 --> 00:00:30.730 to the deep dive. I have been spending a lot 00:00:30.730 --> 00:00:33.490 of time recently thinking about how video creation 00:00:33.490 --> 00:00:36.149 is evolving. It is moving incredibly fast right 00:00:36.149 --> 00:00:38.750 now. It really is. And our mission today is very 00:00:38.750 --> 00:00:41.549 clear. we are unpacking the true capabilities 00:00:41.549 --> 00:00:45.229 of Gemini Omni. Yes. Because I think there is 00:00:45.229 --> 00:00:47.070 a huge misconception out there about what this 00:00:47.070 --> 00:00:49.369 tool actually does. Oh, absolutely. There really 00:00:49.369 --> 00:00:51.289 is. Most people, you know, they just use Omni 00:00:51.289 --> 00:00:54.070 to make AI avatars. Right. They create a digital 00:00:54.070 --> 00:00:56.350 clone of a person that just speaks directly to 00:00:56.350 --> 00:00:59.549 the camera. And I mean, that is fine. Yeah. It 00:00:59.549 --> 00:01:02.450 is a neat trick. It is. But by stopping there, 00:01:02.570 --> 00:01:05.189 they are missing the entire bigger picture. They 00:01:05.189 --> 00:01:08.010 are completely missing five hidden workflows 00:01:08.010 --> 00:01:11.390 that actually revolutionize real video post -production. 00:01:11.829 --> 00:01:14.349 Right. And that is exactly what we are exploring 00:01:14.349 --> 00:01:17.409 today. We are going to look at how you can take 00:01:17.409 --> 00:01:19.870 the mundane footage already sitting on your phone 00:01:19.870 --> 00:01:23.500 and entirely manipulate its reality. Yeah. We're 00:01:23.500 --> 00:01:26.000 talking about adding impossible camera movements 00:01:26.000 --> 00:01:29.140 to static shots. Which is mind blowing. We will 00:01:29.140 --> 00:01:31.180 look at translating your natural speech into 00:01:31.180 --> 00:01:33.719 entirely new languages without dubbing, generating 00:01:33.719 --> 00:01:36.459 full explainer videos from nothing but a single 00:01:36.459 --> 00:01:38.939 thought. Right. And finally locking three -dimensional 00:01:38.939 --> 00:01:42.900 text right into a moving physical scene. It genuinely 00:01:42.900 --> 00:01:45.340 is a complete paradigm shift for creators. I 00:01:45.340 --> 00:01:47.659 mean, it takes the supercomputer in your pocket 00:01:47.659 --> 00:01:50.799 and turns it into a high -end iterative post 00:01:50.799 --> 00:01:53.810 -production studio. Before we could perform any 00:01:53.810 --> 00:01:56.349 of this video magic, we need to talk about this 00:01:56.349 --> 00:01:58.709 setup. Yeah, the workspace matters. It really 00:01:58.709 --> 00:02:01.969 does. We have to understand where we are actually 00:02:01.969 --> 00:02:04.430 doing the work. Yeah. If you just open the standard 00:02:04.430 --> 00:02:06.469 mobile app, you're going to hit a wall. That 00:02:06.469 --> 00:02:09.330 is such a crucial distinction to make right away. 00:02:10.289 --> 00:02:12.590 Gemini Omni essentially lives inside two very 00:02:12.590 --> 00:02:15.370 separate environments. The Gemini app is built 00:02:15.370 --> 00:02:18.370 for quick, single -step edits. You type a prompt 00:02:18.370 --> 00:02:21.509 and you get a fast result. But the catch is that 00:02:21.509 --> 00:02:23.849 it wipes your slate completely clean every single 00:02:23.849 --> 00:02:26.330 time. It is essentially just a temporary scratch 00:02:26.330 --> 00:02:29.069 pad. It is ephemeral. Exactly. It does not save 00:02:29.069 --> 00:02:32.569 your history, which means you cannot build a 00:02:32.569 --> 00:02:35.069 sequence of complex edits over time. No, you 00:02:35.069 --> 00:02:37.830 cannot. And that brings us to Google Flow. Flow 00:02:37.830 --> 00:02:41.110 is the workspace where the serious work actually 00:02:41.110 --> 00:02:43.590 happens. Yes. Google Flow organizes all your 00:02:43.590 --> 00:02:46.030 video generations into specific projects. It 00:02:46.030 --> 00:02:48.750 actually saves your history. So you can select 00:02:48.750 --> 00:02:51.509 a generated output and use it as the new base 00:02:51.509 --> 00:02:54.669 for your next prompt. Every single edit seamlessly 00:02:54.669 --> 00:02:57.090 layers on top of the last one. It is like stacking 00:02:57.090 --> 00:03:00.280 Lego blocks of data. Oh, I love that. You were 00:03:00.280 --> 00:03:03.199 building your final video piece by piece, rather 00:03:03.199 --> 00:03:05.460 than hoping the machine spits out a masterpiece 00:03:05.460 --> 00:03:08.639 on the very first try. Right. Let us look at 00:03:08.639 --> 00:03:11.400 a specific example. Say you upload that 10 -second 00:03:11.400 --> 00:03:15.360 beach clip into Google Flow. You type, add a 00:03:15.360 --> 00:03:17.979 large crowd on the beach behind me. The system 00:03:17.979 --> 00:03:20.460 processes that instruction beautifully. It really 00:03:20.460 --> 00:03:22.759 does. It keeps your original shot completely 00:03:22.759 --> 00:03:27.159 intact. Your face, your posture, your position 00:03:27.159 --> 00:03:29.800 in the frame. Those do not change at all. But 00:03:29.800 --> 00:03:33.840 the AI selectively alters the background environment 00:03:33.840 --> 00:03:36.580 around you. It inserts people naturally into 00:03:36.580 --> 00:03:38.560 the depth of the scene. I have to admit, I still 00:03:38.560 --> 00:03:40.699 wrestle with prompt drift myself. Oh, really? 00:03:40.860 --> 00:03:43.759 Yeah. I am constantly fighting the urge to cram 00:03:43.759 --> 00:03:45.879 too many instructions into the very first prompt. 00:03:46.199 --> 00:03:48.039 Yes. You just want to say, add a crowd and make 00:03:48.039 --> 00:03:50.229 it sunset and make me wear sunglasses. all at 00:03:50.229 --> 00:03:52.629 once. Yeah, and that is honestly the biggest 00:03:52.629 --> 00:03:55.490 mistake new users make. You must keep the first 00:03:55.490 --> 00:03:58.469 prompt focused on one single isolated change. 00:03:58.930 --> 00:04:01.430 Asking for multiple massive edits at once confuses 00:04:01.430 --> 00:04:03.830 the underlying model. Right. It makes it incredibly 00:04:03.830 --> 00:04:06.129 hard to identify what went wrong if the output 00:04:06.129 --> 00:04:08.490 looks weird. Why not just use the Gemini app 00:04:08.490 --> 00:04:10.539 since it is a faster entry point. Because if 00:04:10.539 --> 00:04:13.219 you use the app, you lose that layered progress 00:04:13.219 --> 00:04:15.759 entirely. You cannot iterate on a successful 00:04:15.759 --> 00:04:17.819 first step. You always have to start from zero 00:04:17.819 --> 00:04:21.079 again. Flow lets you keep your wins. Right. Flow 00:04:21.079 --> 00:04:25.040 saves progress. The app wipes your slate clean. 00:04:25.899 --> 00:04:28.600 Beat. So once you successfully get that crowd 00:04:28.600 --> 00:04:30.379 generated in the background, you move forward. 00:04:30.500 --> 00:04:33.720 Yeah. You select that newly generated video inside 00:04:33.720 --> 00:04:36.540 Flow as your new source file. Then you layer 00:04:36.540 --> 00:04:38.680 in the next specific change. Maybe you ask it 00:04:38.680 --> 00:04:41.379 to change the time of day to golden hour. It 00:04:41.379 --> 00:04:43.339 edits the already generated version. Exactly. 00:04:43.500 --> 00:04:45.519 It does not start over from your original phone 00:04:45.519 --> 00:04:48.199 footage. This is exactly what makes iterations 00:04:48.199 --> 00:04:53.069 so powerful here. Yes. But of course, It is not 00:04:53.069 --> 00:04:55.430 always perfect. Oh no, definitely not. Sometimes 00:04:55.430 --> 00:04:57.329 the system just gets it fundamentally wrong. 00:04:57.689 --> 00:05:00.189 It might alter an object at the wrong moment 00:05:00.189 --> 00:05:02.769 or glitch out a part of your face that you wanted 00:05:02.769 --> 00:05:05.430 kept completely intact. When that happens, do 00:05:05.430 --> 00:05:08.230 not try to fix the broken video. Right? I see 00:05:08.230 --> 00:05:11.550 so many people prompting the AI to fix the weird 00:05:11.550 --> 00:05:14.310 hand. Iterating on a bad generation usually just 00:05:14.310 --> 00:05:16.709 compounds the errors. Yeah, it gets messy fast. 00:05:16.790 --> 00:05:18.730 You just have to go back to the previous clean 00:05:18.730 --> 00:05:21.430 clip, write a clearer prompt, and try again. 00:05:21.519 --> 00:05:23.680 A clearer prompt on a clean source always wins. 00:05:24.720 --> 00:05:27.800 So flow lets you fix the environment without 00:05:27.800 --> 00:05:29.680 breaking the subject. Right. Right. But what 00:05:29.680 --> 00:05:32.000 if the environment is actually fine and the problem 00:05:32.000 --> 00:05:34.899 is how you shot it? Oh, this is good. You are 00:05:34.899 --> 00:05:39.560 stuck with a static boring eye level tripod shot. 00:05:39.680 --> 00:05:42.699 Yeah. Let us pivot and talk about altering the 00:05:42.699 --> 00:05:45.699 actual camera itself. Yes. Because here is where 00:05:45.699 --> 00:05:48.199 it gets really interesting. Gemini Omni has this 00:05:48.199 --> 00:05:51.980 amazing ability to completely reinterpret how 00:05:51.980 --> 00:05:55.279 a shot was filmed. It changes the physical reality 00:05:55.279 --> 00:05:57.920 of the camera after the fact. Just think about 00:05:57.920 --> 00:05:59.879 the physics of that for a second. Yeah. You have 00:05:59.879 --> 00:06:03.519 a flat static clip recorded from eye level. You 00:06:03.519 --> 00:06:06.000 prompt Omni to zoom out and turn it into an aerial 00:06:06.000 --> 00:06:08.850 drone shot. Right. And to do that, it has to 00:06:08.850 --> 00:06:10.889 essentially rebuild the entire physical scene 00:06:10.889 --> 00:06:12.870 from a perspective that never existed in the 00:06:12.870 --> 00:06:15.589 real world. The first couple of seconds of the 00:06:15.589 --> 00:06:19.050 generated clip might look a little unstable as 00:06:19.050 --> 00:06:20.850 the system tries to establish that brand new 00:06:20.850 --> 00:06:24.350 perspective. But then... The movement smooths 00:06:24.350 --> 00:06:27.569 out into this beautiful, sweeping cinematic shot. 00:06:27.850 --> 00:06:29.870 It opens up massive possibilities. I mean, if 00:06:29.870 --> 00:06:32.350 you are an indie filmmaker or a solo creator, 00:06:32.930 --> 00:06:35.149 you do not need an expensive drone or a heavy 00:06:35.149 --> 00:06:37.850 gimbal system anymore. Right. But the system 00:06:37.850 --> 00:06:40.209 can also be a little too eager to help sometimes. 00:06:40.370 --> 00:06:42.910 It might hallucinate contextual props into your 00:06:42.910 --> 00:06:45.350 scene to justify the new camera angle. Yeah, 00:06:45.430 --> 00:06:47.889 let us clarify that term quickly. Hallucinate. 00:06:48.110 --> 00:06:51.670 in AI terms, just means creating fake details 00:06:51.670 --> 00:06:54.050 to make a scene logically consistent. Exactly. 00:06:54.170 --> 00:06:56.810 For example, if you ask for a sweeping drone 00:06:56.810 --> 00:07:00.050 shot of yourself on the beach, the AI might actually 00:07:00.050 --> 00:07:02.329 generate a plastic drone controller in your empty 00:07:02.329 --> 00:07:04.649 hands. Because it thinks, well, if there is a 00:07:04.649 --> 00:07:06.149 drone flying around them, they must be the one 00:07:06.149 --> 00:07:08.170 flying it. Right. It is a fascinating glimpse 00:07:08.170 --> 00:07:10.790 into how the machine understands human context, 00:07:11.129 --> 00:07:13.410 not just pixels. And you can control that virtual 00:07:13.410 --> 00:07:16.009 camera with extreme precision, right? You do 00:07:16.009 --> 00:07:18.560 not just have to type pan left, and hope for 00:07:18.560 --> 00:07:20.639 the best. Exactly. You can use what is called 00:07:20.639 --> 00:07:23.480 the arrow technique to dictate exact flight paths. 00:07:23.879 --> 00:07:26.240 You take a still frame of your video, just a 00:07:26.240 --> 00:07:29.509 simple screenshot. and you literally draw arrows 00:07:29.509 --> 00:07:32.310 on that image to show the exact curved path you 00:07:32.310 --> 00:07:35.069 want the camera to take. Oh, wow. Then you upload 00:07:35.069 --> 00:07:37.430 that marked up image alongside your original 00:07:37.430 --> 00:07:41.209 clip. Whoa. Imagine the AI understanding a flat 00:07:41.209 --> 00:07:45.350 2D image so well it can reconstruct a full 3D 00:07:45.350 --> 00:07:47.850 drone flight path through it. It is wild. It 00:07:47.850 --> 00:07:50.750 builds a virtual 3D dome over your scene, maps 00:07:50.750 --> 00:07:53.790 the flat image onto it, and flies a digital camera 00:07:53.790 --> 00:07:56.730 along your drawn line. How does the system know 00:07:56.730 --> 00:07:58.790 exactly exactly where you want this virtual drone 00:07:58.790 --> 00:08:01.529 to fly. You prompt it to trace the path shown 00:08:01.529 --> 00:08:04.050 in the reference image. You tell it to maintain 00:08:04.050 --> 00:08:06.829 a forward -facing perspective, and you explicitly 00:08:06.829 --> 00:08:09.529 tell it to remove the drawn arrows from the final 00:08:09.529 --> 00:08:11.509 output. So it literally just follows the arrows 00:08:11.509 --> 00:08:13.930 you drew on the image. It gives the system every 00:08:13.930 --> 00:08:16.790 single geometric constraint it needs to succeed. 00:08:16.870 --> 00:08:18.930 Right. The more constrained your instruction, 00:08:19.189 --> 00:08:21.189 the better and more consistent the final result 00:08:21.189 --> 00:08:23.589 will be across multiple generations. OK, so we 00:08:23.589 --> 00:08:25.329 have manipulated the background elements. Mm 00:08:25.329 --> 00:08:28.139 -hmm. entirely rebuilt the camera movements. 00:08:29.019 --> 00:08:33.460 Now, let us talk about manipulating the actual 00:08:33.460 --> 00:08:36.740 subject speaking in the video. Reaching a global 00:08:36.740 --> 00:08:39.740 audience. Yes. This is a huge pain point for 00:08:39.740 --> 00:08:42.820 creators right now. Rerecording the same video 00:08:42.820 --> 00:08:46.080 in multiple languages takes massive effort. Oh, 00:08:46.159 --> 00:08:48.220 absolutely. Hiring voice actors, dubbing the 00:08:48.220 --> 00:08:50.679 audio. I mean, it is easily the most time consuming 00:08:50.679 --> 00:08:54.019 part of a global content workflow. Gemini Omni 00:08:54.019 --> 00:08:56.720 handles this problem through its dedicated avatar 00:08:56.720 --> 00:08:59.399 feature. You can deliver your exact message in 00:08:59.399 --> 00:09:01.580 a completely different language, and you never 00:09:01.580 --> 00:09:03.519 have to record a second take yourself. What you 00:09:03.519 --> 00:09:05.779 do is set up a hyper -realistic version of yourself 00:09:05.779 --> 00:09:08.440 in the app first. You give it some baseline footage 00:09:08.440 --> 00:09:11.779 so it learns your face, then you bring that custom 00:09:11.779 --> 00:09:14.720 avatar right into Google Flow, and you simply 00:09:14.720 --> 00:09:17.419 type out the message in your new target language. 00:09:17.820 --> 00:09:20.679 The system has been thoroughly tested on several 00:09:20.679 --> 00:09:23.940 common languages. French, Spanish, Portuguese, 00:09:24.059 --> 00:09:27.500 and German all produce incredibly reliable, natural 00:09:27.500 --> 00:09:29.679 sounding results. And researchers have even tested 00:09:29.679 --> 00:09:32.240 it on much less conventional options just to 00:09:32.240 --> 00:09:34.190 push the boundaries of the model. People have 00:09:34.190 --> 00:09:36.710 generated outputs in Latin and even American 00:09:36.710 --> 00:09:38.730 Sign Language. Though I imagine those outputs 00:09:38.730 --> 00:09:41.169 are much harder to verify without a native speaker. 00:09:41.309 --> 00:09:43.649 Definitely. Still, the underlying capability 00:09:43.649 --> 00:09:46.049 is simply staggering to think about. You can 00:09:46.049 --> 00:09:48.950 run one single marketing message through five 00:09:48.950 --> 00:09:51.529 different languages, back to back. And because 00:09:51.529 --> 00:09:54.409 you are in flow, each language runs as a completely 00:09:54.409 --> 00:09:56.470 separate generation. You never have to touch 00:09:56.470 --> 00:09:59.149 your original audio or video recording again. 00:09:59.330 --> 00:10:01.830 Does it just slap a dubbed audio track over the 00:10:01.830 --> 00:10:04.950 original video? Not at all. It actively rebuilds 00:10:04.950 --> 00:10:08.029 the visual data of your lower face. It perfectly 00:10:08.029 --> 00:10:10.669 matches the new syllables to your mouth movements 00:10:10.669 --> 00:10:14.110 using pixel -level reconstruction. Ah, it actually 00:10:14.110 --> 00:10:16.370 alters your facial expressions and natural lip 00:10:16.370 --> 00:10:19.009 sync. Two -sex silence. It genuinely creates 00:10:19.009 --> 00:10:21.350 a seamless illusion for the international viewer. 00:10:21.909 --> 00:10:25.169 You avoid that uncanny valley effect of old dubbed 00:10:25.169 --> 00:10:27.210 movies. Right, where the mouth movements are 00:10:27.210 --> 00:10:29.970 completely wrong. Exactly. The translated message 00:10:29.970 --> 00:10:34.190 feels entirely authentic and na - So avatars 00:10:34.190 --> 00:10:35.850 are great for translating what you've already 00:10:35.850 --> 00:10:38.669 said. But what if you do not even have a video 00:10:38.669 --> 00:10:42.750 yet? What if you need the AI to build an educational 00:10:42.750 --> 00:10:46.409 video entirely from scratch? Most explainer workflows 00:10:46.409 --> 00:10:49.309 require a massive amount of heavy lifting up 00:10:49.309 --> 00:10:51.750 front. Yes they do. You need a written script, 00:10:52.009 --> 00:10:53.950 you need to record a clean voiceover, and you 00:10:53.950 --> 00:10:57.230 need to source a huge stack of relevant b -roll 00:10:57.230 --> 00:10:59.950 footage. It is a lot of work. It is. But Gemini 00:10:59.950 --> 00:11:03.480 Omni lets you skip all of that manual preparation 00:11:03.480 --> 00:11:05.799 entirely. You do not feed it a script at all. 00:11:05.860 --> 00:11:08.700 You just give it a single focused topic to explain. 00:11:08.960 --> 00:11:11.179 Let us say you ask it to explain how rockets 00:11:11.179 --> 00:11:13.879 work. All right. You tell it to include an avatar 00:11:13.879 --> 00:11:16.299 presenter in the corner of the screen. That is 00:11:16.299 --> 00:11:18.879 literally all the instruction the system needs 00:11:18.879 --> 00:11:21.299 to begin generating. It builds a beautifully 00:11:21.299 --> 00:11:23.980 structured video explanation completely on its 00:11:23.980 --> 00:11:26.899 own. Yeah. It draws from its deep training data 00:11:27.289 --> 00:11:30.470 on scientific subjects. It automatically creates 00:11:30.470 --> 00:11:33.029 scenes showing the action and reaction of the 00:11:33.029 --> 00:11:36.070 launch sequence. Right. It generates clear, accurate 00:11:36.070 --> 00:11:38.210 animations of fuel combustion and high pressure 00:11:38.210 --> 00:11:41.429 gas. It shows how the resulting thrust pushes 00:11:41.429 --> 00:11:45.470 the heavy rocket upward. The final output genuinely 00:11:45.470 --> 00:11:48.210 feels like a finished, polished piece of media. 00:11:48.490 --> 00:11:51.509 It does not feel like a disjointed draft. And 00:11:51.509 --> 00:11:54.629 it does all this from one incredibly short prompt. 00:11:54.870 --> 00:11:57.350 A great habit here is to keep that very first 00:11:57.350 --> 00:12:00.570 prompt broad. Right. Let the system build the 00:12:00.570 --> 00:12:03.250 initial structural foundation for you. Then, 00:12:03.389 --> 00:12:05.490 because you are in flow, you review the output 00:12:05.490 --> 00:12:08.549 and refine it. You only add more depth to specific 00:12:08.549 --> 00:12:11.009 scenes where it is actually needed. Exactly. 00:12:11.190 --> 00:12:13.190 Do I need to feed it a detailed script first 00:12:13.190 --> 00:12:16.000 for the explainer? No. You just provide the core 00:12:16.000 --> 00:12:18.840 topic and your preferred visual style. The system 00:12:18.840 --> 00:12:21.059 automatically handles the narrative pacing and 00:12:21.059 --> 00:12:23.539 the scene transitions for you. No. It structures 00:12:23.539 --> 00:12:25.879 the whole visual breakdown from one short prompt, 00:12:26.220 --> 00:12:29.009 beat. Now, speaking of building scenes, there 00:12:29.009 --> 00:12:31.389 is an incredible bonus hack we need to discuss 00:12:31.389 --> 00:12:33.750 here. Oh, yes. It involves altering location 00:12:33.750 --> 00:12:36.149 data and moving footage. Let us set the scene 00:12:36.149 --> 00:12:38.970 for this. OK. Say you have video filmed from 00:12:38.970 --> 00:12:41.710 inside a moving car. You were just driving down 00:12:41.710 --> 00:12:44.850 a very boring suburban street. But you want to 00:12:44.850 --> 00:12:47.169 completely change the location outside your window 00:12:47.169 --> 00:12:50.029 to make it look like Tokyo at night. In traditional 00:12:50.029 --> 00:12:52.909 editing, this requires incredibly tedious rotoscoping. 00:12:53.029 --> 00:12:55.629 You would have to manually mask out the windows 00:12:55.629 --> 00:12:58.789 frame by frame. But in Omni, you just take a 00:12:58.789 --> 00:13:01.470 screenshot from Google Maps of the new city. 00:13:01.929 --> 00:13:04.990 you upload that map image right alongside your 00:13:04.990 --> 00:13:07.470 original driving clip. You prompt the system 00:13:07.470 --> 00:13:09.909 to change the environment outside the windshield 00:13:09.909 --> 00:13:12.570 using the map as a reference. Yeah. But you tell 00:13:12.570 --> 00:13:15.110 it to keep the car interior exactly the same. 00:13:15.169 --> 00:13:17.850 The model is doing something called depth segmentation. 00:13:18.009 --> 00:13:19.710 Which means separating the foreground from the 00:13:19.710 --> 00:13:22.759 background in the image. Exactly. It is not just 00:13:22.759 --> 00:13:26.159 looking at a flat image. It literally draws an 00:13:26.159 --> 00:13:29.700 invisible 3D boundary between the foreground, 00:13:30.100 --> 00:13:32.940 your steering wheel and dashboard, and the background 00:13:32.940 --> 00:13:36.120 outside the glass. So it replaces the outside 00:13:36.120 --> 00:13:38.500 layer with the new city while protecting the 00:13:38.500 --> 00:13:41.019 inside layer perfectly. Yes. It even keeps the 00:13:41.019 --> 00:13:43.440 original window stickers and dashboard reflections 00:13:43.440 --> 00:13:46.919 completely intact. It is a stunning display of 00:13:46.919 --> 00:13:49.700 spatial awareness by the model. It treats the 00:13:49.700 --> 00:13:52.440 car window like a digital green screen, projecting 00:13:52.440 --> 00:13:54.799 the new data exclusively into that back layer. 00:13:55.299 --> 00:13:58.259 OK. Let's take a quick moment here. Sponsor. 00:13:58.399 --> 00:14:02.139 Minerals sponsor, read placeholder. We are back. 00:14:02.240 --> 00:14:04.539 We have completely altered backgrounds, cameras, 00:14:04.740 --> 00:14:06.860 languages, and locations today. We have covered 00:14:06.860 --> 00:14:09.580 a lot. We have. But to add the final layer of 00:14:09.580 --> 00:14:12.620 polish to an explainer or a product demo, you 00:14:12.620 --> 00:14:16.399 usually need text on the screen. But we are not 00:14:16.399 --> 00:14:19.299 talking about flat, boring text overlays here. 00:14:19.600 --> 00:14:22.799 No. Standard video editors just slap text on 00:14:22.799 --> 00:14:25.340 top of the footage. It lacks parallax. Meaning 00:14:25.340 --> 00:14:27.120 objects moving at different speeds depending 00:14:27.120 --> 00:14:29.879 on their distance? Right. It does not move naturally 00:14:29.879 --> 00:14:32.440 with the underlying physical scene, and it certainly 00:14:32.440 --> 00:14:34.899 does not attach itself to real objects in the 00:14:34.899 --> 00:14:37.320 frame. It constantly breaks the illusion of reality 00:14:37.320 --> 00:14:39.720 for the viewer. Gemini Omni changes this entirely. 00:14:39.879 --> 00:14:42.559 It renders text directly into the three -dimensional 00:14:42.559 --> 00:14:45.139 space of your video. It is basically treating 00:14:45.139 --> 00:14:47.679 the physical object like it has digital sticky 00:14:47.679 --> 00:14:49.740 notes glued to it. When your physical camera 00:14:49.740 --> 00:14:52.480 moves, the text stays locked in place. Now let's 00:14:52.480 --> 00:14:55.100 look at a close -up video of a blooming orchid. 00:14:55.759 --> 00:14:58.600 You prompt the system to label the different 00:14:58.600 --> 00:15:01.440 parts of the flower. You ask it to use an AI 00:15:01.440 --> 00:15:04.620 -style text aesthetic for the labels. You instruct 00:15:04.620 --> 00:15:07.299 it to keep each label securely attached to its 00:15:07.299 --> 00:15:09.899 corresponding petal. The AI actually understands 00:15:09.899 --> 00:15:12.059 the spherical geometry of the petal, not just 00:15:12.059 --> 00:15:15.279 the pixels on the screen. It places a distinct 00:15:15.279 --> 00:15:18.340 text label onto each individual element and locks 00:15:18.340 --> 00:15:21.220 them firmly into the 3D space. As your camera 00:15:21.220 --> 00:15:25.039 slowly pans around the flower, the text tracks 00:15:25.039 --> 00:15:28.000 perfectly. Yeah. The labels move with the object 00:15:28.000 --> 00:15:30.279 naturally, rather than drifting loosely around 00:15:30.279 --> 00:15:32.779 the frame. This specific feature is remarkably 00:15:32.779 --> 00:15:35.340 effective for educational content or dynamic 00:15:35.340 --> 00:15:37.759 product demonstrations online. Right. You can 00:15:37.759 --> 00:15:40.159 call out specific features directly on a physical 00:15:40.159 --> 00:15:42.139 product while you just handle the item normally 00:15:42.139 --> 00:15:44.820 on camera. It adds an interactive, high -production 00:15:44.820 --> 00:15:47.600 feel to simple phone footage, and it requires 00:15:47.600 --> 00:15:50.659 absolutely zero manual post -production tracking 00:15:50.659 --> 00:15:53.440 work from you. But, you know, there are rules 00:15:53.440 --> 00:15:56.210 to this. What happens to those 3D text labels 00:15:56.210 --> 00:15:58.909 if the camera shakes? The system needs clear 00:15:58.909 --> 00:16:01.190 visual anchors to track the physical objects. 00:16:01.590 --> 00:16:04.289 If the clip is shaky or poorly lit, it loses 00:16:04.289 --> 00:16:06.590 those anchors quickly, and the text labels will 00:16:06.590 --> 00:16:08.450 immediately start sliding off their targets. 00:16:08.690 --> 00:16:11.750 Got it. Shaky clips cause the 3D labels to drift 00:16:11.750 --> 00:16:15.850 and misalign. Beats. Exactly. You must film the 00:16:15.850 --> 00:16:18.330 object steadily and ensure good lighting. That 00:16:18.330 --> 00:16:20.970 is a core best practice. There are a few other 00:16:20.970 --> 00:16:23.669 critical best practices we must cover if you 00:16:23.669 --> 00:16:26.610 want Omni to work consistently for you. First, 00:16:27.029 --> 00:16:29.230 keep your source video clip strictly under 10 00:16:29.230 --> 00:16:31.990 seconds long. A 10 -second clip with one clear 00:16:31.990 --> 00:16:35.309 subject is the ideal canvas. The system thrives 00:16:35.309 --> 00:16:37.029 when the visual information it has to process 00:16:37.029 --> 00:16:39.690 is limited and highly focused. If there's too 00:16:39.690 --> 00:16:42.269 much visual competition in the frame, It struggles 00:16:42.269 --> 00:16:44.870 to isolate the correct element. Yeah. Second, 00:16:45.149 --> 00:16:47.490 you must use specific time markers in your written 00:16:47.490 --> 00:16:49.429 prompts. Right. If a transformation needs to 00:16:49.429 --> 00:16:52.049 happen mid -clip, tell it exactly when. Say, 00:16:52.289 --> 00:16:53.909 change the background at the three -second mark. 00:16:54.250 --> 00:16:56.409 Without that clear reference, the system just 00:16:56.409 --> 00:16:58.909 guesses the timing. And its guess is usually 00:16:58.909 --> 00:17:01.690 completely wrong for your specific edit. It usually 00:17:01.690 --> 00:17:04.369 is. A time marker removes the guesswork entirely 00:17:04.369 --> 00:17:06.849 and gives the system a fixed point to work toward. 00:17:07.329 --> 00:17:09.809 And finally, we have to repeat the golden rule 00:17:09.809 --> 00:17:12.509 of this entire workflow. Yes. If a generation 00:17:12.509 --> 00:17:15.349 is fundamentally broken, do not try to fix it. 00:17:15.609 --> 00:17:17.630 Go straight back to the original clip and restart 00:17:17.630 --> 00:17:19.829 your process. Rewrite your prompt to be much 00:17:19.829 --> 00:17:22.450 more specific. Yeah. Iterating on a broken video 00:17:22.450 --> 00:17:25.890 just compounds the initial errors. So if we connect 00:17:25.890 --> 00:17:28.009 this to the bigger picture, what does this all 00:17:28.009 --> 00:17:31.910 mean for you as a creator? Let us synthesize 00:17:31.910 --> 00:17:34.910 the entire deep dive right here. OK. The critical 00:17:34.910 --> 00:17:37.930 insight today is that the gap between a usable 00:17:37.930 --> 00:17:40.710 video and a great one is not about writing a 00:17:40.710 --> 00:17:43.509 longer, more complex prompt. The secret is always 00:17:43.509 --> 00:17:46.089 just one more round of iteration and flow. Right. 00:17:46.289 --> 00:17:48.690 And that is exactly why you cannot just rely 00:17:48.690 --> 00:17:51.970 on the mobile app. You have to embrace the layered, 00:17:52.369 --> 00:17:54.990 non -linear approach of the workspace. Omnia 00:17:54.990 --> 00:17:56.809 is so much more than a simple avatar generator. 00:17:57.130 --> 00:17:59.630 It is a complete post -production studio sitting 00:17:59.630 --> 00:18:02.589 in your pocket. It unlocks the hidden potential 00:18:02.589 --> 00:18:05.250 of the mundane footage you already have. You 00:18:05.250 --> 00:18:07.170 just have to change how you approach the editing 00:18:07.170 --> 00:18:10.170 process. Start small, pick just one specific 00:18:10.170 --> 00:18:12.789 use case we discussed today, run it through Google 00:18:12.789 --> 00:18:15.509 Flow, stack a few edits, and carefully observe 00:18:15.509 --> 00:18:18.569 what comes back. The learning curve is surprisingly 00:18:18.569 --> 00:18:21.490 short. The results will honestly tell you more 00:18:21.490 --> 00:18:24.950 than any tutorial ever could. But before we go, 00:18:25.309 --> 00:18:27.430 exploring all of this does leave us with a rather 00:18:27.430 --> 00:18:30.390 profound final thought to mull over. It really 00:18:30.390 --> 00:18:33.230 forces us to question the nature of digital video 00:18:33.230 --> 00:18:36.089 itself moving forward. We started today by looking 00:18:36.089 --> 00:18:39.250 at a flat amateur phone video of a beach. A clip 00:18:39.250 --> 00:18:42.509 that felt undeniably real, just poorly shot. 00:18:42.670 --> 00:18:45.509 Right. But if AI can now retroactively add a 00:18:45.509 --> 00:18:48.210 flawless drone perspective, if it can completely 00:18:48.210 --> 00:18:50.069 alter the weather outside a moving car window 00:18:50.069 --> 00:18:52.710 using nothing but a map screenshot. How long 00:18:52.710 --> 00:18:55.029 until we stop trusting the reality of casual, 00:18:55.269 --> 00:18:58.150 everyday phone footage altogether? It is a fascinating 00:18:58.150 --> 00:19:00.630 and slightly terrifying question for the future 00:19:00.630 --> 00:19:03.589 of digital media. And terrifying is exactly the 00:19:03.589 --> 00:19:05.269 word for it. Thank you for joining us on this 00:19:05.269 --> 00:19:08.710 Deep Drive. We will see you next time. OTO music.