WEBVTT 00:00:00.000 --> 00:00:02.620 Have you ever tried to get an AI to make a group 00:00:02.620 --> 00:00:05.080 photo? Oh, it's a nightmare. It's an absolute 00:00:05.080 --> 00:00:07.179 disaster. For years, you'd ask for something 00:00:07.179 --> 00:00:09.400 simple. You know, five friends at a picnic, and 00:00:09.400 --> 00:00:11.980 you'd get, well, you'd get a thumb where it shouldn't 00:00:11.980 --> 00:00:14.640 be. Or that one blurry face in the back that 00:00:14.640 --> 00:00:16.899 looks like it's melting. Exactly. Or just that 00:00:16.899 --> 00:00:18.940 awful flat lighting that makes everyone look 00:00:18.940 --> 00:00:23.179 like a plastic doll. AI images were. They were 00:00:23.179 --> 00:00:26.280 fun toys. A novelty. A novelty, for sure. But 00:00:26.280 --> 00:00:28.859 what if that all just changed almost overnight? 00:00:29.489 --> 00:00:32.450 Welcome to the deep dive. Today, we are digging 00:00:32.450 --> 00:00:35.369 into a massive update to Google's image generator. 00:00:36.030 --> 00:00:39.170 Officially, it's called Gemini 3, maybe Imogen 00:00:39.170 --> 00:00:41.850 3. But the community has its own names. It really 00:00:41.850 --> 00:00:44.570 does. Some are calling it Nano Banana Pro. Which 00:00:44.570 --> 00:00:47.960 is, I mean, it's just jargon. But it gets at 00:00:47.960 --> 00:00:50.000 the feeling here. It does. It shows this isn't 00:00:50.000 --> 00:00:51.979 just a small step. The source material we're 00:00:51.979 --> 00:00:54.659 looking at calls it a giant leap. So our mission 00:00:54.659 --> 00:00:56.859 today is to unpack what makes this so different. 00:00:57.299 --> 00:00:59.380 We'll show you how to use it for free right now, 00:00:59.380 --> 00:01:01.700 and we're even going to share some prompts to 00:01:01.700 --> 00:01:04.599 get you those stunning results. Yeah, we're going 00:01:04.599 --> 00:01:06.980 to get past the plastic faces and straight to 00:01:06.980 --> 00:01:10.120 photorealism. So where do we start? I think we 00:01:10.120 --> 00:01:12.859 have to start with the physics. The huge jump 00:01:12.859 --> 00:01:16.750 in realism and why that plastic look. is finally 00:01:16.750 --> 00:01:19.030 a thing of the past. OK. Then we'll get into 00:01:19.030 --> 00:01:22.390 the AI's memory, which is really where the magic 00:01:22.390 --> 00:01:25.010 is for creative work. And finally, we'll cover 00:01:25.010 --> 00:01:26.870 how you can actually use this professionally. 00:01:26.950 --> 00:01:29.930 Let's do it. So the fundamental problem with 00:01:29.930 --> 00:01:32.569 the old models, it was always about texture. 00:01:33.049 --> 00:01:34.629 They could see shapes. They could see colors. 00:01:34.989 --> 00:01:37.450 But they didn't understand reality. If you ask 00:01:37.450 --> 00:01:39.689 for five friends on a beach, you just got this 00:01:39.689 --> 00:01:42.390 flat, cartoonish image. It was never convincing. 00:01:42.510 --> 00:01:45.799 It just felt fake. Totally fake. Yeah. The breakthrough 00:01:45.799 --> 00:01:47.840 here and what the sources are all pointing to 00:01:47.840 --> 00:01:51.159 is that Gemini 3 gets nuance. It understands 00:01:51.159 --> 00:01:54.060 that human skin has pores, imperfections. It's 00:01:54.060 --> 00:01:56.560 not smooth plastic. Exactly. It knows wood has 00:01:56.560 --> 00:01:59.379 grain. It processes how light actually works 00:01:59.379 --> 00:02:01.959 in the real world. And that's the biggest giveaway 00:02:01.959 --> 00:02:04.439 for an AI image bad lighting. And this update 00:02:04.439 --> 00:02:06.200 seems to fix that. It pretty much eliminates 00:02:06.200 --> 00:02:09.180 it. OK, so let's unpack the tech behind this. 00:02:09.719 --> 00:02:12.439 The sources are highlighting, I think, four major 00:02:12.439 --> 00:02:14.599 improvements that get rid of that flatness. Yeah. 00:02:14.639 --> 00:02:17.120 And the first one is the big one. Which is that. 00:02:17.259 --> 00:02:20.599 It actually thinks. It actually thinks. Before, 00:02:21.240 --> 00:02:23.340 a model would just start drawing pixels immediately. 00:02:24.139 --> 00:02:28.419 This one, it pauses. It actually takes a moment 00:02:28.419 --> 00:02:30.539 to plan the composition. Like an artist doing 00:02:30.539 --> 00:02:34.000 a quick sketch first. Precisely. A wire frame, 00:02:34.379 --> 00:02:37.759 a thumbnail sketch. it plans before it renders. 00:02:37.879 --> 00:02:39.819 And that leads directly to the second point, 00:02:39.919 --> 00:02:43.659 which is just the realism. That smooth, plasticky 00:02:43.659 --> 00:02:46.460 skin is gone. The photos look like they were 00:02:46.460 --> 00:02:49.599 taken with a high -end camera. A DSLR, for sure. 00:02:50.060 --> 00:02:51.539 It's a different quality altogether. Then there's 00:02:51.539 --> 00:02:53.960 the text handling. This is huge. Oh, this has 00:02:53.960 --> 00:02:56.060 been a problem for years. You'd get a great image, 00:02:56.159 --> 00:02:59.099 but the sign in the background would be just 00:02:59.099 --> 00:03:00.860 gibberish. Now, if you want a sign that says 00:03:00.860 --> 00:03:02.740 coffee shop, it actually spells coffee shop. 00:03:02.759 --> 00:03:04.840 Which is... critical for any kind of commercial 00:03:04.840 --> 00:03:08.120 work. And that brings us to the last point. Consistency. 00:03:08.280 --> 00:03:10.960 The holy grail. You can actually keep the same 00:03:10.960 --> 00:03:14.219 character, the same face, across a bunch of different 00:03:14.219 --> 00:03:16.400 images. That used to be almost impossible. It 00:03:16.400 --> 00:03:19.300 was. And that planning phase you mentioned, that's 00:03:19.300 --> 00:03:21.740 the engine behind all of this. It's what allows 00:03:21.740 --> 00:03:25.180 for the realism, the good text, everything. Even 00:03:25.180 --> 00:03:29.199 the AI now thinks. How does that planning phase 00:03:29.199 --> 00:03:31.639 fundamentally alter the quality compared to the 00:03:31.639 --> 00:03:34.900 older models? It allows for complex cinematic 00:03:34.900 --> 00:03:38.639 lighting and detailed realism, which were previously 00:03:38.639 --> 00:03:42.099 impossible. The planning step lets the AI build 00:03:42.099 --> 00:03:44.639 the physics of the scene first. Exactly. It plans 00:03:44.639 --> 00:03:47.379 the physics before it draws the picture. So how 00:03:47.379 --> 00:03:49.360 do we actually get our hands on this? Is it, 00:03:49.360 --> 00:03:51.500 you know, some special software? Not at all. 00:03:51.520 --> 00:03:53.340 It's available right now in the browser. Just 00:03:53.340 --> 00:03:55.300 on Google Gemini? Yep. You just have to look 00:03:55.300 --> 00:03:57.580 for the Gemini advanced model, or sometimes you'll 00:03:57.580 --> 00:03:59.780 see a little label that says thinking with three. 00:04:00.280 --> 00:04:01.719 That's how you know you're using the new engine. 00:04:01.979 --> 00:04:04.120 And it's free. There are limits for free users, 00:04:04.300 --> 00:04:07.060 usually about three to five high quality images 00:04:07.060 --> 00:04:09.639 a day. And there's a catch, right? The watermark. 00:04:09.939 --> 00:04:12.819 Right. All the images have an invisible digital 00:04:12.819 --> 00:04:15.699 watermark. You can't see it, but it's embedded 00:04:15.699 --> 00:04:18.420 in the file's data. Which is for traceability, 00:04:18.459 --> 00:04:21.319 I assume. to fight deepfakes. That's the idea. 00:04:21.720 --> 00:04:24.139 It's Google's way of creating a digital signature, 00:04:24.439 --> 00:04:27.300 so you can always check if an image was AI -generated. 00:04:27.779 --> 00:04:30.139 For serious users, that's actually a good thing. 00:04:30.259 --> 00:04:33.740 It builds trust. Okay, that makes sense. Now, 00:04:33.899 --> 00:04:35.899 let's talk about the feature that everyone is 00:04:35.899 --> 00:04:40.199 buzzing about, this. continuous context editing 00:04:40.199 --> 00:04:42.180 this is the one I mean we've all been there I 00:04:42.180 --> 00:04:44.360 I still wrestle with prompt drift myself you 00:04:44.360 --> 00:04:46.100 get the perfect character the perfect face and 00:04:46.100 --> 00:04:48.620 then you want to change one tiny thing exactly 00:04:48.620 --> 00:04:50.879 you'd say okay now make the cat wear a hat and 00:04:50.879 --> 00:04:52.879 poof he would generate a completely different 00:04:52.879 --> 00:04:56.000 cat you'd lose the one you liked the AI had zero 00:04:56.000 --> 00:04:59.339 memory zero this new model it remembers what 00:04:59.339 --> 00:05:01.759 it just created It holds onto the context from 00:05:01.759 --> 00:05:04.220 one prompt to the next. For storyboarding or 00:05:04.220 --> 00:05:07.199 any sequential work, it's a game changer. I love 00:05:07.199 --> 00:05:09.439 the coffee shop experiment they walk through 00:05:09.439 --> 00:05:11.779 in the source material. It shows this perfectly. 00:05:11.959 --> 00:05:14.300 Yeah, that's a great example. So the first prompt 00:05:14.300 --> 00:05:17.620 sets the scene. Something like, create a realistic 00:05:17.620 --> 00:05:20.839 cinematic photo of a young male barista in a 00:05:20.839 --> 00:05:25.939 green apron. Warm lighting. 16 .9 aspect ratio, 00:05:26.060 --> 00:05:28.220 and you get a great image. Right, your starting 00:05:28.220 --> 00:05:30.680 point. But then the magic word is just next. 00:05:30.860 --> 00:05:32.779 You just type next. And what happens? You get 00:05:32.779 --> 00:05:35.100 a new angle of the same barista and the same 00:05:35.100 --> 00:05:38.060 cafe. The face is the same, the apron, the lighting. 00:05:38.240 --> 00:05:40.779 It's all consistent. It feels like a second shot 00:05:40.779 --> 00:05:42.639 from the same photo shoot. It's incredible. Then 00:05:42.639 --> 00:05:44.819 you can change the action. You prompt, now show 00:05:44.819 --> 00:05:46.839 him serving coffee to a customer. Yeah. And it 00:05:46.839 --> 00:05:49.060 keeps the barista identical, but just changes 00:05:49.060 --> 00:05:51.100 what he's doing. We've never had that level of 00:05:51.100 --> 00:05:53.540 control. So beyond just making a few sequential 00:05:53.540 --> 00:05:56.899 shots, what is the core implication of retaining 00:05:56.899 --> 00:05:59.959 that same character across, say, hundreds of 00:05:59.959 --> 00:06:02.819 requests? Businesses can now create reliable 00:06:02.819 --> 00:06:05.819 brand mascots and models without needing expensive 00:06:05.819 --> 00:06:08.660 photo shoots. You're creating a persistent digital 00:06:08.660 --> 00:06:11.579 asset you can use over and over. Exactly. OK, 00:06:11.579 --> 00:06:14.040 let's move from the how to the what, specifically 00:06:14.040 --> 00:06:17.139 what to type. If you're still just prompting 00:06:17.139 --> 00:06:19.459 woman in Japan street, you're going to get flat. 00:06:19.600 --> 00:06:22.579 generic results. You need the ProPrompt Blueprint. 00:06:23.439 --> 00:06:25.860 The sources are really clear on this. The best 00:06:25.860 --> 00:06:28.959 prompts use terms from photography and physics. 00:06:29.160 --> 00:06:31.800 Let's do a comparison. Take a travel blogger. 00:06:32.079 --> 00:06:34.720 Bad prompt is just subject and location. Right. 00:06:35.019 --> 00:06:38.160 A ProPrompt gets specific. It uses phrases like 00:06:38.160 --> 00:06:40.980 candid realistic medium shot. That controls the 00:06:40.980 --> 00:06:43.920 framing. And lighting. It would specify golden 00:06:43.920 --> 00:06:46.379 hour lighting. Exactly. And my favorite one? 00:06:47.660 --> 00:06:51.480 Background is slightly blurry. Okay, that instantly 00:06:51.480 --> 00:06:53.839 adds depth and makes it look professional. You're 00:06:53.839 --> 00:06:56.420 basically telling the AI which camera lens to 00:06:56.420 --> 00:06:59.259 use. Okay, what about a food ad instead of just 00:06:59.259 --> 00:07:01.459 picture of a burger? You use sensory details, 00:07:01.600 --> 00:07:03.920 professional food photography. You tell the cheese 00:07:03.920 --> 00:07:06.240 is melting down the side. You can get that specific. 00:07:06.399 --> 00:07:09.259 Oh, yeah, and scheme is rising from the patty. 00:07:09.379 --> 00:07:12.680 Then you control the lighting. Dark, moody background 00:07:12.680 --> 00:07:15.579 with a spotlight. Whoa, imagine generating hundreds 00:07:15.579 --> 00:07:18.410 of perfect product shots an hour. where you can 00:07:18.410 --> 00:07:20.730 control the exact moment the cheese melts, that's 00:07:20.730 --> 00:07:23.189 profound. It's a massive shift. And you see the 00:07:23.189 --> 00:07:25.189 same thing with a sneaker ad. The prop isn't 00:07:25.189 --> 00:07:28.829 just sneaker, it's dynamic action shot. And the 00:07:28.829 --> 00:07:32.089 camera angle. Low angle view, ground level. And 00:07:32.089 --> 00:07:35.050 this is the craziest part. Water droplets are 00:07:35.050 --> 00:07:38.009 frozen in midair. You're directing physics. So 00:07:38.009 --> 00:07:39.790 let's talk about where this is already being 00:07:39.790 --> 00:07:43.750 used. The sources list a few key business applications. 00:07:44.949 --> 00:07:46.779 Real estate agents are a big one. Right. They 00:07:46.779 --> 00:07:49.199 can upload a photo of an empty room and just 00:07:49.199 --> 00:07:51.959 tell the AI to furnish it, add a modern gray 00:07:51.959 --> 00:07:54.459 sofa, but keep the windows and walls the same. 00:07:54.660 --> 00:07:57.139 It's instant digital staging. Saves a ton of 00:07:57.139 --> 00:07:59.660 money. And for e -commerce, you can take a simple 00:07:59.660 --> 00:08:02.199 photo of your product, say a candle on your kitchen 00:08:02.199 --> 00:08:05.300 table. You prompt, place this candle on a wooden 00:08:05.300 --> 00:08:08.540 spa shelf, soft lighting, and suddenly your basic 00:08:08.540 --> 00:08:10.730 photo looks like a high -end ad. And the last 00:08:10.730 --> 00:08:12.850 one was for event planners. Yeah, this one's 00:08:12.850 --> 00:08:15.110 cool. They can upload a napkin sketch of a wedding 00:08:15.110 --> 00:08:18.149 layout. A literal napkin sketch. A photo of it. 00:08:18.529 --> 00:08:20.769 And prompt, turn this sketch into a realistic 00:08:20.769 --> 00:08:23.110 photo of a wedding reception with flowers and 00:08:23.110 --> 00:08:26.810 gold lights. It goes from a vague idea to a photorealistic 00:08:26.810 --> 00:08:29.709 concept in seconds. OK, so considering the quality 00:08:29.709 --> 00:08:31.529 we're talking about, especially for those food 00:08:31.529 --> 00:08:35.830 ads, is this tool already replacing entry -level 00:08:35.830 --> 00:08:38.830 commercial photographers? For initial mock -ups 00:08:38.830 --> 00:08:41.610 and marketing visuals, yes, the speed and cost 00:08:41.610 --> 00:08:44.389 savings are enormous. So it's not killing the 00:08:44.389 --> 00:08:46.570 creative vision of the artist, but it's automating 00:08:46.570 --> 00:08:49.850 the technical execution. Exactly. It's democratizing 00:08:49.850 --> 00:08:52.909 the ability to create really persuasive visual 00:08:52.909 --> 00:08:57.090 content. Okay, so based on all the tests in the 00:08:57.090 --> 00:09:00.100 source material, there are four. Let's call them 00:09:00.100 --> 00:09:03.059 secrets, to getting truly professional results. 00:09:03.139 --> 00:09:05.960 A master tips. Master tips. And tip number one 00:09:05.960 --> 00:09:08.539 is the face first rule. Okay, what's that? If 00:09:08.539 --> 00:09:10.960 you're combining things, like a specific person 00:09:10.960 --> 00:09:13.200 with a car and a dog, always upload the photo 00:09:13.200 --> 00:09:16.539 of the person first. The AI gives the most attention 00:09:16.539 --> 00:09:18.759 to the initial input. So if you lead with the 00:09:18.759 --> 00:09:20.919 car, the car will be perfect, but the face might 00:09:20.919 --> 00:09:23.480 be messy. Exactly. The AI needs you to tell it 00:09:23.480 --> 00:09:25.740 what the most complex and important part of the 00:09:25.740 --> 00:09:28.299 image is. That makes sense. What's number two? 00:09:28.379 --> 00:09:30.240 Let's specify the aspect ratio. The shape of 00:09:30.240 --> 00:09:34.820 the picture. Yep. 16 .9 for wide video, 9 .16 00:09:34.820 --> 00:09:38.320 for a tall social media story, 1 .1 for a square 00:09:38.320 --> 00:09:41.320 post. If you don't tell it, it defaults to a 00:09:41.320 --> 00:09:44.139 square and can crop your image badly. Simple 00:09:44.139 --> 00:09:47.799 but crucial. Okay, tip three. Trust the thinking 00:09:47.799 --> 00:09:50.220 process. When it says analyzing or thinking, 00:09:50.860 --> 00:09:53.659 don't cancel it. Be patient. You have to be. 00:09:54.169 --> 00:09:56.110 Pause is the planning phase we talked about. 00:09:56.250 --> 00:09:58.149 It's where the quality comes from. If you interrupt 00:09:58.149 --> 00:10:00.690 it, you get a rush job. Got it. And the last 00:10:00.690 --> 00:10:04.049 one. Know when to use specialized tools. Google 00:10:04.049 --> 00:10:07.399 is. pretty strict on safety. If you need super 00:10:07.399 --> 00:10:10.259 high resolution, like 4K, or you're creating 00:10:10.259 --> 00:10:12.460 content that's a bit more edgy, you might need 00:10:12.460 --> 00:10:14.559 a paid tool like Higgs Field AI or something 00:10:14.559 --> 00:10:17.279 in Adobe Photoshop. That's a great list. So if 00:10:17.279 --> 00:10:19.259 a listener only takes one tip away from this 00:10:19.259 --> 00:10:21.220 whole deep dive, which one should they prioritize? 00:10:21.340 --> 00:10:23.279 What's the number one thing to do? I would say 00:10:23.279 --> 00:10:26.279 tip three. Wait for the AI to finish its analysis. 00:10:26.960 --> 00:10:29.200 That pause is the secret ingredient for quality 00:10:29.200 --> 00:10:31.139 because you're letting it run its physics simulation. 00:10:31.309 --> 00:10:32.610 See, I thought you were going to say the face 00:10:32.610 --> 00:10:34.389 -first rule. That seems like it fixes the most 00:10:34.389 --> 00:10:37.169 common problem, you know, messy hands and faces. 00:10:37.429 --> 00:10:40.970 And it is an easy fix, for sure. But if you cancel 00:10:40.970 --> 00:10:43.330 the thinking phase, even a perfect face -first 00:10:43.330 --> 00:10:45.590 image is still going to have flat lighting and 00:10:45.590 --> 00:10:49.210 bad texture. The planning is foundational. It 00:10:49.210 --> 00:10:51.590 sets the quality for everything else. So prioritize 00:10:51.590 --> 00:10:54.129 the process over the subject. Precisely. So we've 00:10:54.129 --> 00:10:57.190 covered a lot of ground. I think the big idea 00:10:57.190 --> 00:11:01.289 here is that this update, Gemini 3, It transforms 00:11:01.289 --> 00:11:04.029 the image generator from a fun little toy into 00:11:04.029 --> 00:11:07.070 a professional artist inside your computer. And 00:11:07.070 --> 00:11:09.110 that transformation comes down to two things. 00:11:09.269 --> 00:11:11.629 Two things. Hyperrealism, which it gets through 00:11:11.629 --> 00:11:14.470 that planning phase, and continuous memory, which 00:11:14.470 --> 00:11:17.090 gives you consistent characters. So our challenge 00:11:17.090 --> 00:11:19.490 to you, the listener, is pretty simple. Go to 00:11:19.490 --> 00:11:21.789 Google Gemini today, think of a senior dream 00:11:21.789 --> 00:11:24.350 kitchen, a funny picture of your pet, and apply 00:11:24.350 --> 00:11:27.049 the prompt tips we shared. Yeah, focus on the 00:11:27.049 --> 00:11:29.850 lighting, the texture, the camera angle. See 00:11:29.850 --> 00:11:31.690 what you can create. The only way to really keep 00:11:31.690 --> 00:11:33.389 up with this stuff is to get your hands dirty 00:11:33.389 --> 00:11:35.149 and start playing with it. And we have to think 00:11:35.149 --> 00:11:37.590 about what this means. I mean, what if, in a 00:11:37.590 --> 00:11:40.269 year, the line between a real photo and an AI 00:11:40.269 --> 00:11:42.870 image becomes completely invisible, even to experts? 00:11:43.250 --> 00:11:45.759 That's the path we're on. It is. We're heading 00:11:45.759 --> 00:11:48.559 into a world where visual truth is going to be 00:11:48.559 --> 00:11:51.360 a lot harder to define. Keep testing those boundaries. 00:11:51.720 --> 00:11:54.059 Keep playing with the tech. Until next time, 00:11:54.360 --> 00:11:55.179 keep digging deeper.