WEBVTT 00:00:00.000 --> 00:00:03.220 Identity in the digital age is kind of, well, 00:00:03.520 --> 00:00:06.879 elusive, Pete. You know, you can prompt an AI 00:00:06.879 --> 00:00:09.759 to generate a beautiful striking face once, but 00:00:09.759 --> 00:00:12.720 maintaining that specific soul... across time, 00:00:12.820 --> 00:00:16.179 across different angles, and lighting, that has 00:00:16.179 --> 00:00:17.839 always been the hardest puzzle to solve. Oh, 00:00:17.839 --> 00:00:19.839 it really is. I mean, there's nothing more frustrating 00:00:19.839 --> 00:00:22.359 than dialing in this brilliant AI actor, and 00:00:22.359 --> 00:00:25.100 then suddenly they just, they change their entire 00:00:25.100 --> 00:00:27.500 bone structure in the very next frame. They ask 00:00:27.500 --> 00:00:30.239 them to look left, and their jawline completely 00:00:30.239 --> 00:00:33.500 morphs. It drives creators crazy. It really does. 00:00:34.259 --> 00:00:36.079 Welcome to the Deep Dive, everyone. We are so 00:00:36.079 --> 00:00:38.219 glad you're here with us. Today, we're exploring 00:00:38.219 --> 00:00:40.479 a massive new beta update. It's called Google 00:00:40.479 --> 00:00:43.340 Flow. Yeah, highly anticipated. Definitely. And 00:00:43.340 --> 00:00:45.200 we're looking specifically at its characters, 00:00:45.719 --> 00:00:48.179 voice profiles, and personal avatars. So if you're 00:00:48.179 --> 00:00:50.079 listening to this, whether you're a solo creator, 00:00:50.320 --> 00:00:52.340 building a brand, or just endlessly curious about 00:00:52.340 --> 00:00:54.880 the mechanics of AI, this is your shortcut to 00:00:54.880 --> 00:00:57.899 understanding the death of prompt drift. It's 00:00:57.899 --> 00:01:00.119 a fundamental shift, really, in how we interact 00:01:00.119 --> 00:01:02.200 with these models. We're finally moving away 00:01:02.200 --> 00:01:06.260 from the random slot machine of AI generation. 00:01:06.439 --> 00:01:09.739 We're moving toward genuine locked -in consistency. 00:01:10.079 --> 00:01:12.079 OK, let's unpack this. Before we talk about the 00:01:12.079 --> 00:01:14.099 outputs, we really need to understand the architecture. 00:01:14.879 --> 00:01:17.000 Flow isn't just a standalone image generator, 00:01:17.280 --> 00:01:19.980 right? It connects directly to Google Gemini. 00:01:20.219 --> 00:01:22.640 Exactly. And the underlying connection is the 00:01:22.640 --> 00:01:25.980 core of everything. When you type a vague idea 00:01:25.980 --> 00:01:28.180 into Flow, you aren't actually talking directly 00:01:28.180 --> 00:01:31.040 to the image generator. Gemini acts as this intermediate 00:01:31.040 --> 00:01:34.000 director. It processes your plain English. Then 00:01:34.000 --> 00:01:37.340 it writes highly technical, hyper detailed prompt 00:01:37.340 --> 00:01:40.519 matrices for the image models to actually render. 00:01:40.819 --> 00:01:42.599 And that structural setup solves the biggest 00:01:42.599 --> 00:01:45.650 problem in AI video right now. Amnesia. Yes, 00:01:45.989 --> 00:01:48.709 total amnesia. Previously, the AI literally forgot 00:01:48.709 --> 00:01:50.609 who your character was the moment the frame ended. 00:01:50.989 --> 00:01:54.810 It was like recasting an entirely new actor for 00:01:54.810 --> 00:01:57.829 every single camera angle. But Flow acts like 00:01:57.829 --> 00:02:00.969 an ironclad contract for one specific digital 00:02:00.969 --> 00:02:04.689 actor. This is obviously huge for UGC. And by 00:02:04.689 --> 00:02:07.530 that, I mean videos made by regular people to 00:02:07.530 --> 00:02:10.110 promote products online. Oh, massive. Think about 00:02:10.110 --> 00:02:12.210 the broader landscape for a moment. I mean, Runway 00:02:12.210 --> 00:02:14.469 is already famous in the industry for performance 00:02:14.469 --> 00:02:16.830 capture, taking an existing video and stylizing 00:02:16.830 --> 00:02:18.909 it. Yeah, that's everywhere right now. And Sora 00:02:18.909 --> 00:02:21.610 showed off those sweeping cinematic clips with 00:02:21.610 --> 00:02:24.830 some reusable characters. But Flow is doing something 00:02:24.830 --> 00:02:27.289 structurally different. It brings the face, the 00:02:27.289 --> 00:02:30.810 voice, and the personal avatar system into one 00:02:30.810 --> 00:02:33.990 complete unified workflow. It is worth noting 00:02:33.990 --> 00:02:36.330 though this is a premium beta feature. You do 00:02:36.330 --> 00:02:38.349 need a Gemini advanced subscription to access 00:02:38.349 --> 00:02:40.569 it. And it's currently rolling out geographically 00:02:40.569 --> 00:02:42.629 so you know not everyone has it today. True. 00:02:42.669 --> 00:02:45.389 But I want to push on something here. Why is 00:02:45.389 --> 00:02:48.289 this specifically a game changer for solo creators 00:02:48.289 --> 00:02:51.289 versus massive studios? Well, it really comes 00:02:51.289 --> 00:02:53.509 down to scale and resources. Massive studios 00:02:53.509 --> 00:02:56.449 have entire animation budgets. They have whole 00:02:56.449 --> 00:02:58.330 departments of technical directors completely 00:02:58.330 --> 00:03:01.330 dedicated to tracking facial geometry and keeping 00:03:01.330 --> 00:03:04.689 a character consistent frame by frame. A solo 00:03:04.689 --> 00:03:07.310 creator simply doesn't have the time or the compute 00:03:07.310 --> 00:03:10.710 power to manually fix a warping jawline in post 00:03:10.710 --> 00:03:13.509 -production. Yeah, absolutely not. Flow solves 00:03:13.509 --> 00:03:16.030 this lack of consistency natively. You don't 00:03:16.030 --> 00:03:18.569 need a Pixar level budget to keep your digital 00:03:18.569 --> 00:03:21.319 actor looking the exact same across a hundred 00:03:21.319 --> 00:03:23.780 different videos. So it essentially gives one 00:03:23.780 --> 00:03:27.580 person a full consistent digital acting troupe. 00:03:27.900 --> 00:03:30.120 Precisely. It democratizes a level of continuity 00:03:30.120 --> 00:03:32.419 that used to cost millions. Okay, so we understand 00:03:32.419 --> 00:03:35.300 the fundamental problem this solves. But building 00:03:35.300 --> 00:03:38.300 a consistent digital human means you need a solid 00:03:38.300 --> 00:03:40.419 foundation. You have to sculpt them. And Flow 00:03:40.419 --> 00:03:43.099 gives you three distinct methods to create a 00:03:43.099 --> 00:03:45.159 character's physical appearance. Yeah. The first 00:03:45.159 --> 00:03:47.800 is using templates. This is mostly for rapid 00:03:47.800 --> 00:03:51.080 prototyping. You select a broad archetype, like 00:03:51.080 --> 00:03:53.780 the eccentric. The system automatically generates 00:03:53.780 --> 00:03:56.259 the underlying prompt matrix. You have very little 00:03:56.259 --> 00:03:58.080 control here, but it gets a face on the screen 00:03:58.080 --> 00:04:00.620 immediately. Which brings us to the second method, 00:04:00.879 --> 00:04:03.889 where the real power lies. writing text prompts. 00:04:04.210 --> 00:04:06.530 This gives you absolute control over the generation. 00:04:06.550 --> 00:04:08.789 It does. And when you do this, you actually have 00:04:08.789 --> 00:04:11.469 to choose between two specific rendering models. 00:04:11.889 --> 00:04:15.610 You have Nano Banana Pro, which is hyper -realistic. 00:04:15.930 --> 00:04:17.930 It handles complex lighting and skin textures 00:04:17.930 --> 00:04:20.149 perfectly. Then there is Nano Banana 2. Right, 00:04:20.290 --> 00:04:22.910 and Nano Banana 2 is much faster to run, but 00:04:22.910 --> 00:04:26.470 it fundamentally leans toward... stylized, illustrative, 00:04:26.870 --> 00:04:30.209 or artistic aesthetics. The way it interprets 00:04:30.209 --> 00:04:34.470 a text prompt prioritizes broad creative strokes 00:04:34.470 --> 00:04:37.949 over microscopic pores. Makes sense. And the 00:04:37.949 --> 00:04:40.149 third method is uploading an image directly. 00:04:40.389 --> 00:04:42.769 You provide a portrait, and the model maps that 00:04:42.769 --> 00:04:45.170 face as your baseline. But there's a very strict 00:04:45.170 --> 00:04:47.430 referencing rule here, isn't there? Oh, incredibly 00:04:47.430 --> 00:04:49.930 strict. You must establish one clean face image 00:04:49.930 --> 00:04:53.170 first, no complex backgrounds, no crazy lighting. 00:04:53.410 --> 00:04:56.029 Right. If the age or the expression feels even 00:04:56.029 --> 00:04:58.410 slightly wrong, you have to use the what you 00:04:58.410 --> 00:05:00.930 want to change interface to correct it. You absolutely 00:05:00.930 --> 00:05:03.230 do not move forward if the baseline is wrong. 00:05:03.649 --> 00:05:06.370 Beat. I have to admit, I still wrestle with prompt 00:05:06.370 --> 00:05:09.139 drift myself. Just the other day, I had this 00:05:09.139 --> 00:05:11.259 great character. I added a coffee cup to the 00:05:11.259 --> 00:05:13.740 scene, and suddenly my character aged 20 years. 00:05:13.819 --> 00:05:16.519 Oh, wow. Yeah, that is a classic latent space 00:05:16.519 --> 00:05:19.259 problem. The model associates the concept of 00:05:19.259 --> 00:05:21.680 a coffee cup with the training data it learned 00:05:21.680 --> 00:05:25.160 from. Often, images of people holding coffee 00:05:25.160 --> 00:05:27.699 cups in stock photos are older professionals 00:05:27.699 --> 00:05:30.939 reading morning papers. So the model accidentally 00:05:30.939 --> 00:05:34.060 pulls those older demographic features into your 00:05:34.060 --> 00:05:36.259 character's face. That makes total sense. The 00:05:36.259 --> 00:05:38.910 prop actually contains the facial data. Exactly. 00:05:39.230 --> 00:05:41.689 And what's fascinating here is the system's strict 00:05:41.689 --> 00:05:44.490 limit on reference images to combat that exact 00:05:44.490 --> 00:05:47.550 contamination. Once your main face is perfect, 00:05:47.930 --> 00:05:50.029 you are allowed to add a second reference set 00:05:50.029 --> 00:05:52.389 for side and back views. You keep the clothing 00:05:52.389 --> 00:05:54.730 prompts incredibly basic, like Navy sweater, 00:05:55.110 --> 00:05:57.470 but the hard limit is exactly one main reference 00:05:57.470 --> 00:06:00.410 and one extra set of alternative angles. It's 00:06:00.410 --> 00:06:02.829 so tempting to just dump 20 photos of a character 00:06:02.829 --> 00:06:05.269 in the system, assuming more data makes the AI 00:06:05.269 --> 00:06:08.860 smarter. But you can't. What happens if you try 00:06:08.860 --> 00:06:11.019 to force 10 reference images to make it smarter? 00:06:11.279 --> 00:06:13.740 The model architecture just isn't built to blend 00:06:13.740 --> 00:06:16.540 that many distinct 2D inputs into a cohesive 00:06:16.540 --> 00:06:20.120 3D map. Oh, I see. If you overwhelm it with conflicting 00:06:20.120 --> 00:06:22.879 visual data, different lighting, slight changes 00:06:22.879 --> 00:06:25.639 in focal length, the model's attention mechanism 00:06:25.639 --> 00:06:28.459 gets confused about which core features to prioritize. 00:06:28.860 --> 00:06:31.420 It actually dilutes the facial identity instead 00:06:31.420 --> 00:06:34.810 of reinforcing it. Less is more. Feeding in too 00:06:34.810 --> 00:06:37.129 many images actually breaks the system. Exactly. 00:06:37.250 --> 00:06:39.730 It forces you to rely on mathematical precision 00:06:39.730 --> 00:06:42.649 rather than visual volume. But having a pixel 00:06:42.649 --> 00:06:45.310 -perfect face is completely useless for a video 00:06:45.310 --> 00:06:47.350 series if the illusion shatters the second they 00:06:47.350 --> 00:06:50.529 speak. That brings us to the audio engine. Yeah, 00:06:50.550 --> 00:06:52.970 this is where Google has instituted a very strict 00:06:52.970 --> 00:06:55.649 audio rule. You cannot upload an outside audio 00:06:55.649 --> 00:06:58.410 file to clone a voice. The entire vocal identity 00:06:58.410 --> 00:07:01.329 must be generated inside Flow's native ecosystem. 00:07:01.629 --> 00:07:03.870 You can select the built -in voice template to 00:07:03.870 --> 00:07:06.990 start. Or you can use that template as a base 00:07:06.990 --> 00:07:10.470 to engineer a custom voice. Customization requires 00:07:10.470 --> 00:07:13.350 three things. A name, a description, and sample 00:07:13.350 --> 00:07:15.610 dialogue. But the description isn't just about 00:07:15.610 --> 00:07:18.649 accents, is it? No, not at all. You have to define 00:07:18.649 --> 00:07:21.569 the acoustic parameters. You must explicitly 00:07:21.569 --> 00:07:24.670 detail the emotion, the pitch, and the speed. 00:07:24.810 --> 00:07:27.870 Wow. The system needs those behavioral cues to 00:07:27.870 --> 00:07:30.870 map the audio wave. For instance, you might input 00:07:30.870 --> 00:07:34.730 sad, low -pitched, fast -paced. Then you provide 00:07:34.730 --> 00:07:36.990 a sample script so the engine can render a preview. 00:07:37.670 --> 00:07:40.269 Imagine the pressure for a creator in this moment. 00:07:40.649 --> 00:07:42.269 You have to listen to that preview and get it 00:07:42.269 --> 00:07:44.230 absolutely perfect before locking it in, because 00:07:44.230 --> 00:07:46.329 there is a massive limitation here. Yeah, there 00:07:46.329 --> 00:07:49.209 is. Once you click add to character, that voice 00:07:49.209 --> 00:07:51.709 profile is permanently fused to the actor. You 00:07:51.709 --> 00:07:53.649 cannot go back and tweak the pitch. You cannot 00:07:53.649 --> 00:07:55.689 edit it. You can only delete the entire voice 00:07:55.689 --> 00:07:58.370 and start over. It seems rigid, but if we connect 00:07:58.370 --> 00:08:00.310 this to the bigger picture, it makes perfect 00:08:00.310 --> 00:08:03.350 sense. Google is actively preventing the injection 00:08:03.350 --> 00:08:06.730 of deep fake audio. By restricting outside MP3 00:08:06.730 --> 00:08:09.089 uploads, they stop you from cloning a real politician 00:08:09.089 --> 00:08:11.930 or celebrity. You're forced to use their internal 00:08:11.930 --> 00:08:14.509 text to speech tools, which keeps the whole process 00:08:14.509 --> 00:08:17.610 inside a secure, monitored environment. But practically 00:08:17.610 --> 00:08:20.310 speaking, if I notice the voice is slightly too 00:08:20.310 --> 00:08:24.000 fast after saving, what's my move? You have absolutely 00:08:24.000 --> 00:08:27.040 no edit button. You must delete that specific 00:08:27.040 --> 00:08:29.699 profile entirely from the character sheet, open 00:08:29.699 --> 00:08:32.759 a brand new custom voice matrix, rewrite your 00:08:32.759 --> 00:08:34.940 speed parameters, and render it from scratch. 00:08:35.080 --> 00:08:37.120 You have to scrap it entirely and build a brand 00:08:37.120 --> 00:08:39.539 new voice profile. Exactly. It forces creators 00:08:39.539 --> 00:08:42.340 to be incredibly intentional before finalizing 00:08:42.340 --> 00:08:44.820 a digital identity. Let's take a quick break. 00:08:45.320 --> 00:08:48.720 Mid -roll sponsor read. Okay, we're back. We 00:08:48.720 --> 00:08:50.799 have a face and we have a voice. They are locked. 00:08:51.559 --> 00:08:53.960 Now we move to the most powerful, yet probably 00:08:53.960 --> 00:08:55.799 the most misunderstood feature of this entire 00:08:55.799 --> 00:08:58.899 update, the digital soul. Yes, the character 00:08:58.899 --> 00:09:02.000 info box. This is where users consistently make 00:09:02.000 --> 00:09:04.500 a massive mistake. Yeah, they assume it's another 00:09:04.500 --> 00:09:06.820 prompt box, so they start typing physical descriptions. 00:09:07.200 --> 00:09:09.080 Brown hair, blue eyes. You should absolutely 00:09:09.080 --> 00:09:11.580 not do that. The image model already knows what 00:09:11.580 --> 00:09:13.539 the character looks like. Here's where it gets 00:09:13.539 --> 00:09:16.960 really interesting. You treat this text box like 00:09:16.960 --> 00:09:19.879 giving a real Hollywood actor a psychological 00:09:19.879 --> 00:09:22.179 motivation sheet. Exactly. You aren't typing 00:09:22.179 --> 00:09:24.919 visuals. You are typing behavioral structures. 00:09:25.440 --> 00:09:29.139 Quirks. mannerisms, speaking behavior, their 00:09:29.139 --> 00:09:31.740 emotional baseline. You define how they exist 00:09:31.740 --> 00:09:35.220 in a space. You type calm mentor. You describe 00:09:35.220 --> 00:09:37.440 that they speak with deliberate pauses, that 00:09:37.440 --> 00:09:39.940 they smile gently, and that they naturally use 00:09:39.940 --> 00:09:42.480 open hand gestures when explaining things. Or 00:09:42.480 --> 00:09:44.480 conversely, you might build a sarcastic creator 00:09:44.480 --> 00:09:47.200 who constantly smirks, breaks eye contact, and 00:09:47.200 --> 00:09:50.440 rolls their eyes. Right. And the flow agent processes 00:09:50.440 --> 00:09:53.039 this psychological data in three very specific 00:09:53.039 --> 00:09:56.019 ways. First is behavior inheritance. If you put 00:09:56.019 --> 00:09:57.960 smiles gently in that box, you never have to 00:09:57.960 --> 00:10:00.139 type smile gently into your daily video prompts 00:10:00.139 --> 00:10:02.639 ever again. The character just naturally defaults 00:10:02.639 --> 00:10:05.259 to it. The second mechanism is generation guidance. 00:10:05.940 --> 00:10:08.860 The AI acts as a shadow director. It actively 00:10:08.860 --> 00:10:11.539 guides the video rendering model to ensure the 00:10:11.539 --> 00:10:13.980 micro expressions match that saved emotional 00:10:13.980 --> 00:10:17.220 baseline. A sarcastic character will naturally 00:10:17.220 --> 00:10:19.720 carry tension in their jaw, even when silent. 00:10:19.960 --> 00:10:22.860 Which is incredible. And the third way is dialogue 00:10:22.860 --> 00:10:25.759 consistency. The pacing of their custom voice 00:10:25.759 --> 00:10:28.940 automatically adjusts to match that mood. Does 00:10:28.940 --> 00:10:31.139 this mean the AI automatically controls body 00:10:31.139 --> 00:10:33.700 language during a video? Yes, it really does. 00:10:34.059 --> 00:10:36.059 The flow agent reads the personality vectors 00:10:36.059 --> 00:10:38.820 before it renders a single frame. It literally 00:10:38.820 --> 00:10:41.409 translates text traits like... speaking with 00:10:41.409 --> 00:10:44.129 clear authority into the actual posture, the 00:10:44.129 --> 00:10:46.750 physical micro -movements, and the spatial awareness 00:10:46.750 --> 00:10:48.929 of the character on screen. Right. It directs 00:10:48.929 --> 00:10:51.929 the character's acting based purely on that psychological 00:10:51.929 --> 00:10:54.190 text box. It bridges the gap between an animated 00:10:54.190 --> 00:10:56.450 puppet and a true digital human. So the actor 00:10:56.450 --> 00:10:59.629 is fully prepped. The face, voice, and soul are 00:10:59.629 --> 00:11:01.590 integrated. How do we actually call them to set 00:11:01.590 --> 00:11:04.210 and start shooting? Flow uses a remarkably simple 00:11:04.210 --> 00:11:05.970 trigger system. You just type the at symbol, 00:11:06.110 --> 00:11:09.289 followed by their name, like add John. And that 00:11:09.289 --> 00:11:12.929 single tag acts as an entire data package. It 00:11:12.929 --> 00:11:15.750 instantly pulls the facial map, the audio profile, 00:11:15.830 --> 00:11:18.350 and the psychological behaviors straight into 00:11:18.350 --> 00:11:21.789 your active prompt. Whoa. Imagine scaling an 00:11:21.789 --> 00:11:24.769 entire multi -platform video campaign with just 00:11:24.769 --> 00:11:27.450 one at symbol. You don't have to rewrite 50 lines 00:11:27.450 --> 00:11:29.269 of character description for every TikTok or 00:11:29.269 --> 00:11:31.730 YouTube short. It's a huge time saver. Beat. 00:11:32.350 --> 00:11:34.610 It is best practice, though, to test this in 00:11:34.610 --> 00:11:37.879 a static image first. My type. at John sitting 00:11:37.879 --> 00:11:40.899 in a coffee shop. You can change outfits, lighting, 00:11:41.220 --> 00:11:43.620 and scenes endlessly while the core identity 00:11:43.620 --> 00:11:45.759 stays totally locked. And when you're ready to 00:11:45.759 --> 00:11:48.600 transition to motion, the syntax is very specific. 00:11:49.100 --> 00:11:51.519 In the main generation box, you type the at character 00:11:51.519 --> 00:11:53.639 name, followed by a physical action, followed 00:11:53.639 --> 00:11:56.039 by the actual spoken script inside quotation 00:11:56.039 --> 00:11:57.980 marks. But as you scale this, you might suddenly 00:11:57.980 --> 00:12:01.220 run into a violate our policies era. This is 00:12:01.220 --> 00:12:03.120 the safety system kicking in, right? Because 00:12:03.120 --> 00:12:05.360 Flow is constantly scanning outputs for real 00:12:05.360 --> 00:12:07.759 internet faces. Yeah, the security filter is 00:12:07.759 --> 00:12:10.360 exceptionally strict. It is specifically engineered 00:12:10.360 --> 00:12:12.899 to stop deepfakes at the generation level. Just 00:12:12.899 --> 00:12:16.620 to define that clearly, a deepfake is AI -generated 00:12:16.620 --> 00:12:19.820 media that digitally replaces a real person's 00:12:19.820 --> 00:12:23.409 likeness. Right. And the system is scanning biometric 00:12:23.409 --> 00:12:26.909 ratios. Even if your prompt is completely innocent, 00:12:27.429 --> 00:12:30.070 the filter will permanently block the video if 00:12:30.070 --> 00:12:33.049 your digital actor's bone structure mathematically 00:12:33.049 --> 00:12:36.129 aligns too closely with a real photograph scraped 00:12:36.129 --> 00:12:37.769 from the internet. There are two ways around 00:12:37.769 --> 00:12:39.830 this. Option one is what the platform highly 00:12:39.830 --> 00:12:42.490 recommends. You just use a pure AI face from 00:12:42.490 --> 00:12:45.309 the start. Let Flow generate a completely original 00:12:45.309 --> 00:12:48.309 face from a text prompt. The system auto approves 00:12:48.309 --> 00:12:51.049 it down the line because it has the digital provenance, 00:12:51.350 --> 00:12:53.669 proving it is a virtual human. Option two is 00:12:53.669 --> 00:12:56.009 trickier. It's for when you use your own real 00:12:56.009 --> 00:12:58.570 picture as the baseline. The system will likely 00:12:58.570 --> 00:13:00.990 flag it. When it does, you have to click the 00:13:00.990 --> 00:13:03.769 flag icon on the error message. You explicitly 00:13:03.769 --> 00:13:05.929 state that it's an AI character based on your 00:13:05.929 --> 00:13:08.169 own likeness, and then you wait for manual review 00:13:08.169 --> 00:13:10.889 by their technical team. But hold on, if I generated 00:13:10.889 --> 00:13:13.730 the character myself entirely inside their system, 00:13:14.149 --> 00:13:16.450 why does it still flag my character? Because 00:13:16.450 --> 00:13:18.769 the automated scanners are dealing with finite 00:13:18.769 --> 00:13:21.769 mathematical probabilities. They can't always 00:13:21.769 --> 00:13:24.649 distinguish between a highly realistic generated 00:13:24.649 --> 00:13:27.490 face and a copyrighted photograph of a stranger. 00:13:27.590 --> 00:13:30.049 Oh, I see. If the lighting, the texture and the 00:13:30.049 --> 00:13:32.389 geometry cross a certain threshold of realism, 00:13:33.169 --> 00:13:35.549 the system triggers a false positive just to 00:13:35.549 --> 00:13:38.740 be safe. The safety filter panics. if your creation 00:13:38.740 --> 00:13:41.120 looks slightly too human or familiar. Exactly. 00:13:41.340 --> 00:13:43.700 It's forced to err on the side of extreme caution 00:13:43.700 --> 00:13:46.980 to protect real people. Up until now, we've been 00:13:46.980 --> 00:13:49.960 building virtual people from scratch. But what 00:13:49.960 --> 00:13:52.539 if you want to bypass the creation phase entirely? 00:13:53.039 --> 00:13:55.179 What if you want to digitize your actual self? 00:13:55.320 --> 00:13:58.539 Well, Flow has introduced a beta AI avatar feature 00:13:58.539 --> 00:14:01.120 to do exactly this. And this is where the hardware 00:14:01.120 --> 00:14:03.620 requirement becomes fascinating. To build a true 00:14:03.620 --> 00:14:06.259 avatar, Gemini doesn't just need a flat photo, 00:14:06.360 --> 00:14:09.330 it needs spatial data. Yeah, it stands your real 00:14:09.330 --> 00:14:11.950 facial geometry, it analyzes your natural micro 00:14:11.950 --> 00:14:14.190 -expressions, the asymmetry of your mouth when 00:14:14.190 --> 00:14:16.509 you talk, and your baseline speaking rhythm. 00:14:16.750 --> 00:14:19.870 Because it needs that rich data, the setup is 00:14:19.870 --> 00:14:22.309 strictly mobile. You cannot do this with a standard 00:14:22.309 --> 00:14:24.610 desktop webcam. You have to use the Gemini app 00:14:24.610 --> 00:14:27.230 on your phone. Right. You log in, tap your profile, 00:14:27.350 --> 00:14:30.370 and select Avatar New. You have to agree to extensive 00:14:30.370 --> 00:14:33.549 microphone and camera terms. Then you hold your 00:14:33.549 --> 00:14:36.419 phone perfectly at eye level. And you read a 00:14:36.419 --> 00:14:38.460 short, specific script on the screen out loud. 00:14:39.139 --> 00:14:42.259 This calibrates the phonetic tracking. It records 00:14:42.259 --> 00:14:44.500 how your specific vocal cords handle different 00:14:44.500 --> 00:14:47.539 vowel sounds. After the mobile app processes 00:14:47.539 --> 00:14:50.179 that heavy data, you switch back to your desktop 00:14:50.179 --> 00:14:52.940 and check the fully rendered avatar in the Flow 00:14:52.940 --> 00:14:55.659 workspace. It is still clearly in beta, so there 00:14:55.659 --> 00:14:57.940 are technical limitations. It struggles heavily 00:14:57.940 --> 00:15:00.039 with fast lip movements. The rendering can miss 00:15:00.039 --> 00:15:03.870 very small, nuanced micro -expressions. But honestly, 00:15:04.129 --> 00:15:06.090 the privacy rules surrounding this feature are 00:15:06.090 --> 00:15:08.370 just as interesting as the tech. This raises 00:15:08.370 --> 00:15:10.269 an important question about the ownership of 00:15:10.269 --> 00:15:12.669 digital identity. Google has engineered this 00:15:12.669 --> 00:15:15.409 to be completely account locked. No one else 00:15:15.409 --> 00:15:18.809 on the platform can search for access or utilize 00:15:18.809 --> 00:15:21.529 your avatar template. And they've implemented 00:15:21.529 --> 00:15:25.389 a strict non -transferable protocol. But what 00:15:25.389 --> 00:15:28.889 if I want to edit a video in Premiere? Can I 00:15:28.889 --> 00:15:31.450 export my digital twin to use in another editing 00:15:31.450 --> 00:15:34.320 software? Absolutely not. You cannot download 00:15:34.320 --> 00:15:37.399 the underlying 3D mesh. You cannot transfer the 00:15:37.399 --> 00:15:40.799 raw avatar profile to any outside project, game 00:15:40.799 --> 00:15:43.919 engine, or third party workspace. The generation 00:15:43.919 --> 00:15:46.519 capability is entirely geofenced within your 00:15:46.519 --> 00:15:48.700 specific Google login. No exporting allowed. 00:15:48.899 --> 00:15:51.539 Your digital clone is locked inside your private 00:15:51.539 --> 00:15:54.000 Google account. It operates as a highly secure 00:15:54.000 --> 00:15:56.440 walled garden to prevent your identity from being 00:15:56.440 --> 00:15:58.820 hijacked. So let's bring this all together. Google 00:15:58.820 --> 00:16:01.529 Flow fundamentally changes the medium. It isn't 00:16:01.529 --> 00:16:04.169 just an image generator anymore. It is a completely 00:16:04.169 --> 00:16:07.750 unified production studio. It turns AI from a 00:16:07.750 --> 00:16:10.970 random, unpredictable slot machine of faces into 00:16:10.970 --> 00:16:13.950 a predictable, directable camera. Yeah, it really 00:16:13.950 --> 00:16:16.070 does. Now, we have to stay grounded. It is not 00:16:16.070 --> 00:16:18.269 flawless yet. The commercial rights regarding 00:16:18.269 --> 00:16:21.409 generated likenesses are still murky legal territory. 00:16:21.610 --> 00:16:24.129 There are plenty of beta bugs. But for a solo 00:16:24.129 --> 00:16:26.789 creator trying to build a narrative universe, 00:16:27.509 --> 00:16:30.840 this is a massive leap forward. It saves hundreds 00:16:30.840 --> 00:16:33.340 of hours of manual prompting and post -production 00:16:33.340 --> 00:16:36.039 corrections. So what does this all mean? First 00:16:36.039 --> 00:16:37.860 of all, thank you for taking this Deem Dive with 00:16:37.860 --> 00:16:41.299 us. If you do have access to the beta, you really 00:16:41.299 --> 00:16:44.279 should go test that at symbol function. Seeing 00:16:44.279 --> 00:16:47.279 an entire character, voice, and personality summon 00:16:47.279 --> 00:16:49.480 instantly changes how you think about workflow. 00:16:49.580 --> 00:16:52.879 That's pretty wild. But stepping back, it leaves 00:16:52.879 --> 00:16:54.600 us with something much deeper to think about. 00:16:54.750 --> 00:16:57.850 If we can now hard code our specific quirks, 00:16:58.070 --> 00:17:00.769 our mannerisms, and our exact vocal cadences 00:17:00.769 --> 00:17:04.170 into a digital twin that never gets tired, that 00:17:04.170 --> 00:17:06.990 never ages, that never forgets its lines, what 00:17:06.990 --> 00:17:09.309 happens to the concept of authenticity? What 00:17:09.309 --> 00:17:11.029 happens when our audiences can no longer tell 00:17:11.029 --> 00:17:13.250 if it's really us talking to them or just our 00:17:13.250 --> 00:17:15.569 very well -documented ghost in the machine? Two 00:17:15.569 --> 00:17:18.230 -Sec Silence. Until next time, keep exploring.