WEBVTT 00:00:00.000 --> 00:00:02.660 You know, NanoBanana is this incredibly powerful 00:00:02.660 --> 00:00:06.179 tool. It feels like it should be able to, say, 00:00:06.440 --> 00:00:08.740 add a realistic backpack to your photo or maybe 00:00:08.740 --> 00:00:11.339 just wipe away a trash can seamlessly. Right. 00:00:11.419 --> 00:00:13.919 Exactly. But then you try it and getting the 00:00:13.919 --> 00:00:15.859 shadows right or making sure the background fill 00:00:15.859 --> 00:00:18.960 looks natural. It often feels impossible. You 00:00:18.960 --> 00:00:21.679 can spend hours just tweaking one little thing. 00:00:21.920 --> 00:00:25.019 That disconnect is definitely real. So today 00:00:25.019 --> 00:00:27.260 we're moving past the really basic prompts. We 00:00:27.260 --> 00:00:30.960 want to unpack 10 proven methods, think of them 00:00:30.960 --> 00:00:33.280 like formulas, that can turn that frustration 00:00:33.280 --> 00:00:36.219 into professional -grade results consistently. 00:00:36.439 --> 00:00:38.600 Yeah, this is our deep dive into Google's Nano 00:00:38.600 --> 00:00:41.719 Banana. And the mission is clear. Let's skip 00:00:41.719 --> 00:00:43.560 the endless trial and error. We're going to show 00:00:43.560 --> 00:00:45.439 you the specific kinds of commands that actually 00:00:45.439 --> 00:00:47.759 work, the ones that lead to high -quality realistic 00:00:47.759 --> 00:00:50.420 images. OK, but before we even get into the fun 00:00:50.420 --> 00:00:52.780 stuff, the prompts themselves, we have to cover 00:00:52.780 --> 00:00:56.359 step zero. Accessing Nano Manana the right way, 00:00:56.659 --> 00:00:58.420 it actually matters more than you might think. 00:00:58.659 --> 00:01:01.380 Oh, absolutely. It's a critical first step. So 00:01:01.380 --> 00:01:03.799 many people get, frankly, bad results just because 00:01:03.799 --> 00:01:05.500 they're using the wrong platform, and then they 00:01:05.500 --> 00:01:09.079 blame the AI. The data strongly suggests use 00:01:09.079 --> 00:01:11.920 Google AI Studio, not the general Gemini interface. 00:01:12.500 --> 00:01:14.519 Wait, really? So just using Gemini might mean 00:01:14.519 --> 00:01:16.319 you're getting worse results before you even 00:01:16.319 --> 00:01:19.680 type a prompt? For consistent, you know, professional 00:01:19.680 --> 00:01:22.000 -level output, yeah. The workflow makes a huge 00:01:22.000 --> 00:01:24.840 difference. In Gemini, you have to manually pick 00:01:24.840 --> 00:01:27.500 the image tool every single time. Forget once. 00:01:27.659 --> 00:01:29.819 You just get text back about the image. Ah, so 00:01:29.819 --> 00:01:31.859 it loses the context. Exactly right. AI Studio 00:01:31.859 --> 00:01:34.379 is just smarter about it. The Nano Banana button 00:01:34.379 --> 00:01:37.239 is obvious, and crucially, if you used the image 00:01:37.239 --> 00:01:40.340 tool last time, it remembers. It defaults to 00:01:40.340 --> 00:01:43.219 it. Saves time, saves errors. Okay, so a quick 00:01:43.219 --> 00:01:45.099 tip for you listening. Always just glance at 00:01:45.099 --> 00:01:46.980 those run settings. Make sure Nano Banana is 00:01:46.980 --> 00:01:48.780 actually selected. It'll save you some headache. 00:01:48.980 --> 00:01:51.099 Alright, let's get to the recipes. Starting with 00:01:51.099 --> 00:01:54.280 adding and removing things smoothly. Method one. 00:01:54.900 --> 00:01:57.420 Seamless insertion. Adding something new, making 00:01:57.420 --> 00:01:59.739 it look like it belongs. The big insight here, 00:01:59.760 --> 00:02:02.060 we call it the shadow rule. Because if I just 00:02:02.060 --> 00:02:03.920 tell it, add a green backpack to someone walking 00:02:03.920 --> 00:02:07.859 in a forest, it often looks, well, flat, like 00:02:07.859 --> 00:02:10.240 a sticker, right? Totally. The prompt has to 00:02:10.240 --> 00:02:12.879 force the AI to deal with the existing light 00:02:12.879 --> 00:02:15.879 and texture. So in our example, yeah, we say 00:02:15.879 --> 00:02:18.860 add a moss green hiking backpack that is a bit 00:02:18.860 --> 00:02:22.659 old and used. That helps. But the real key is 00:02:22.659 --> 00:02:25.379 the command about light. Create a natural shadow 00:02:25.379 --> 00:02:27.939 from the backpack onto the woman's back and jacket 00:02:27.939 --> 00:02:30.159 matching the sunlight that is coming from the 00:02:30.159 --> 00:02:33.000 top left. OK, how does specifying the light source 00:02:33.000 --> 00:02:35.460 like top left sunlight, how does that help it 00:02:35.460 --> 00:02:38.060 avoid looking fake, looking pasted on? It forces 00:02:38.060 --> 00:02:41.620 coherence. It gives the AI concrete data, points 00:02:41.620 --> 00:02:43.680 the direction of the type of light so it can 00:02:43.680 --> 00:02:46.199 calculate and integrate the new object's shadow 00:02:46.199 --> 00:02:48.960 and even subtle texture interactions correctly. 00:02:49.159 --> 00:02:51.360 It grounds the new object in the scene's reality. 00:02:51.759 --> 00:02:54.539 Makes sense. Method two is the flip side. removing 00:02:54.539 --> 00:02:57.020 unwanted stuff. Cleaning up a photo, getting 00:02:57.020 --> 00:02:59.259 rid of a distracting sign, maybe footprints on 00:02:59.259 --> 00:03:01.479 a nice clean beach. And the key insight you mentioned 00:03:01.479 --> 00:03:04.219 is, don't just remove, rebuild. You have to tell 00:03:04.219 --> 00:03:06.680 the AI what should be there after the thing is 00:03:06.680 --> 00:03:09.219 gone. Right. So don't just say, remove all the 00:03:09.219 --> 00:03:11.599 footprints. You need more. Add commands like, 00:03:11.979 --> 00:03:14.080 make the surface of the sand look smooth and 00:03:14.080 --> 00:03:17.039 natural. And crucially, keep the original color 00:03:17.039 --> 00:03:19.759 and texture of the sand, and also keep the waves. 00:03:19.960 --> 00:03:22.560 way back there. So if we only say remove the 00:03:22.560 --> 00:03:24.819 footprints, what's the risk? What does the AI 00:03:24.819 --> 00:03:28.580 do with that blank space? Well, it guesses. It 00:03:28.580 --> 00:03:31.360 might fill it with some random texture or something 00:03:31.360 --> 00:03:34.680 inconsistent that just screams fake edit. Okay, 00:03:34.900 --> 00:03:36.800 adding and removing down. Let's move to more 00:03:36.800 --> 00:03:40.710 precise edits. Focus changes and... Style transfers. 00:03:41.310 --> 00:03:43.650 Method three is about changing just one specific 00:03:43.650 --> 00:03:45.909 part, like a sofa color, without messing up the 00:03:45.909 --> 00:03:47.830 whole room. Yeah, this is where we fight against 00:03:47.830 --> 00:03:50.009 prompt drift. That annoying thing where the AI 00:03:50.009 --> 00:03:51.849 thinks you want to redo the whole image vibe, 00:03:51.930 --> 00:03:54.150 where you just want a yellow sofa instead of 00:03:54.150 --> 00:03:57.370 gray. I have to admit, I still wrestle with prompt 00:03:57.370 --> 00:03:59.289 drift myself sometimes when I'm trying to change 00:03:59.289 --> 00:04:01.969 just one small thing. It's like the AI sees the 00:04:01.969 --> 00:04:05.370 whole picture as one unit. It does. So the formula 00:04:05.370 --> 00:04:07.889 here is all about constraint. We say, change 00:04:07.889 --> 00:04:11.449 the gray sofa to deep mustard yellow, okay? But 00:04:11.449 --> 00:04:14.210 then the critical part is adding positive constraints. 00:04:14.830 --> 00:04:17.629 Keep the original corduroy fabric texture as 00:04:17.629 --> 00:04:19.470 well as the wrinkles and shadows. And there's 00:04:19.470 --> 00:04:21.750 a negative command too, right? Yes, absolutely 00:04:21.750 --> 00:04:23.430 vital. Everything else in the room must stay 00:04:23.430 --> 00:04:25.470 exactly the same. The don't touch anything else 00:04:25.470 --> 00:04:27.769 command is just non -negotiable if you want to 00:04:27.769 --> 00:04:30.209 clean edit. Why is defining the original texture, 00:04:30.350 --> 00:04:32.790 like corduroy, so important? Isn't just naming 00:04:32.790 --> 00:04:36.009 the new color enough? Not really, because defining 00:04:36.009 --> 00:04:39.089 texture, corduroy, velvet, whatever, forces the 00:04:39.089 --> 00:04:42.410 AI to retain depth and realism. It prevents the 00:04:42.410 --> 00:04:45.370 object from looking like some flat kind of soulless 00:04:45.370 --> 00:04:47.689 3D model replacement. It keeps the character. 00:04:47.889 --> 00:04:49.490 That's a really helpful way to think about it. 00:04:49.769 --> 00:04:52.129 OK, method four, style transfer. And this one 00:04:52.129 --> 00:04:54.930 has a pretty blunt rule. Go famous or go home. 00:04:55.129 --> 00:04:57.360 Uh -huh, pretty much. If you want your photo 00:04:57.360 --> 00:04:59.699 turned into a specific art style, the AI needs 00:04:59.699 --> 00:05:02.000 loads of training data on that style. Think The 00:05:02.000 --> 00:05:05.540 Simpsons, Van Gogh, maybe Liechtenstein pop art. 00:05:05.839 --> 00:05:08.720 Don't try like obscure indie comic artists. It 00:05:08.720 --> 00:05:10.740 won't know them well enough. Yeah, I learned 00:05:10.740 --> 00:05:12.459 that one the hard way. Tried to get a selfie 00:05:12.459 --> 00:05:16.100 in the style of some lesser known abstract expressionist. 00:05:16.160 --> 00:05:19.240 It just kind of smeared the photo a bit. Didn't 00:05:19.240 --> 00:05:21.879 capture the style at all. Exactly. because the 00:05:21.879 --> 00:05:24.000 training data is just too thin for niche styles. 00:05:24.660 --> 00:05:27.160 So you need to name the famous style, describe 00:05:27.160 --> 00:05:29.860 its key visual elements like classic yellow skin, 00:05:30.379 --> 00:05:33.660 big round eyes for the Simpsons, and also tell 00:05:33.660 --> 00:05:36.319 it what to keep, like maybe keep my hairstyle 00:05:36.319 --> 00:05:39.420 and smile. So based on the data, what usually 00:05:39.420 --> 00:05:41.600 happens if someone tries a style transfer with 00:05:41.600 --> 00:05:44.160 an obscure art style the AI barely recognizes? 00:05:44.560 --> 00:05:46.319 It typically just gives you back the original 00:05:46.319 --> 00:05:49.199 photo with maybe a few generic lines or simple 00:05:49.199 --> 00:05:51.579 filter effects overlaid. It fails to capture 00:05:51.579 --> 00:05:54.420 the actual essence, the brushwork, the color 00:05:54.420 --> 00:05:56.500 palette of the style. Okay, let's shift to blending 00:05:56.500 --> 00:05:59.279 things together. Method five is about combining 00:05:59.279 --> 00:06:01.620 elements from different photos, like maybe taking 00:06:01.620 --> 00:06:03.360 a jacket from one picture and putting it on a 00:06:03.360 --> 00:06:06.019 person in another. Virtual try -on, essentially. 00:06:06.399 --> 00:06:08.579 Here, you absolutely have to define the source 00:06:08.579 --> 00:06:12.060 and the destination clearly. And crucially, state 00:06:12.060 --> 00:06:14.920 what must not change in the main image. Things 00:06:14.920 --> 00:06:17.639 like keep the woman's face, hairstyle, pose, 00:06:17.819 --> 00:06:21.019 and blue jeans exactly the same. And for realism, 00:06:21.480 --> 00:06:23.899 the key command seems to be about lighting again, 00:06:24.199 --> 00:06:26.660 that instruction. The lighting on the new jacket 00:06:26.660 --> 00:06:29.040 must match the lighting of the environment in 00:06:29.040 --> 00:06:31.600 image one. So thinking about that, what's the 00:06:31.600 --> 00:06:34.620 detailed command that really stops the new clothing 00:06:34.620 --> 00:06:37.160 item from looking, you know, totally out of place, 00:06:37.240 --> 00:06:40.100 like it doesn't belong? It's asking the AI specifically 00:06:40.100 --> 00:06:43.000 to match the new item's lighting to the existing 00:06:43.000 --> 00:06:45.160 environment's lighting. It forces integration. 00:06:45.680 --> 00:06:48.540 Right. OK, moving quickly, method six, creating 00:06:48.540 --> 00:06:50.779 new angles. This needs more technical language, 00:06:50.819 --> 00:06:52.300 doesn't it? You can't just say, show me the side. 00:06:52.540 --> 00:06:54.540 Exactly, too vague. So if you have a front view 00:06:54.540 --> 00:06:57.339 of a vintage car, you need to be specific. A 00:06:57.339 --> 00:07:00.180 low angle from the front left side, a 34 view. 00:07:00.779 --> 00:07:02.819 And you can add environmental interaction too, 00:07:03.319 --> 00:07:05.180 like the sunset light should be reflected on 00:07:05.180 --> 00:07:07.500 the hood and windshield. Beyond just specifying 00:07:07.500 --> 00:07:09.699 the angle. What's that small detail that makes 00:07:09.699 --> 00:07:12.240 the car photo feel more dynamic, almost like 00:07:12.240 --> 00:07:14.560 it's ready to drive off? It's asking for the 00:07:14.560 --> 00:07:16.399 front left wheel to be turned slightly towards 00:07:16.399 --> 00:07:18.540 the camera. That little bit of implied action 00:07:18.540 --> 00:07:21.720 really brings it to life. Mid -rule sponsor break. 00:07:24.300 --> 00:07:26.540 Welcome back to our deep dive on nano banana 00:07:26.540 --> 00:07:29.680 formulas. Before the break, we were digging into 00:07:29.680 --> 00:07:32.240 how light texture and perspective commands are 00:07:32.240 --> 00:07:34.740 crucial. Now let's look at some really high value 00:07:34.740 --> 00:07:38.040 practical uses. Yeah, method seven. Creating 00:07:38.040 --> 00:07:40.519 professional product mock -ups. This is huge 00:07:40.519 --> 00:07:43.019 for e -commerce, for client work. Being able 00:07:43.019 --> 00:07:45.300 to visualize products quickly. And the trick 00:07:45.300 --> 00:07:48.980 here seems to be twofold. Create a context, a 00:07:48.980 --> 00:07:51.500 sort of lifestyle story, and make sure the logo 00:07:51.500 --> 00:07:53.720 or design looks like it's actually on the product, 00:07:53.819 --> 00:07:56.079 not just stuck there, like placing a logo on 00:07:56.079 --> 00:07:58.180 a tote pack. Right, you set the scene. Carried 00:07:58.180 --> 00:08:00.339 on the shoulder of a woman, walking through an 00:08:00.339 --> 00:08:02.399 outdoor farmer's market. That gives context. 00:08:02.740 --> 00:08:05.399 But the quality command is the killer. The logo 00:08:05.399 --> 00:08:07.720 must look like it is printed on the fabric, not 00:08:07.720 --> 00:08:10.639 pasted on. That forces the AI to think about 00:08:10.639 --> 00:08:13.019 the fabric's texture, how it folds, the lighting. 00:08:13.279 --> 00:08:15.939 That's massive. I remember trying product mockups 00:08:15.939 --> 00:08:18.560 ages ago and just getting a logo to not look 00:08:18.560 --> 00:08:21.480 like a flat sticker. It took forever, so many 00:08:21.480 --> 00:08:24.019 attempts. You're saying that one command helps 00:08:24.019 --> 00:08:26.500 fix that? It really tackles the biggest hurdle 00:08:26.500 --> 00:08:28.660 in making digital mockups look convincing, yeah. 00:08:28.839 --> 00:08:32.309 Whoa! Just imagine scaling that, being able to 00:08:32.309 --> 00:08:34.649 mock up thousands of different designs onto products 00:08:34.649 --> 00:08:37.950 for a whole online store, like instantly. That's 00:08:37.950 --> 00:08:41.129 a serious force multiplier. It really is. OK, 00:08:41.350 --> 00:08:43.830 method eight is about turning maybe a casual 00:08:43.830 --> 00:08:47.730 selfie into a proper professional headshot. And 00:08:47.730 --> 00:08:50.110 this really hinges on defining the purpose and 00:08:50.110 --> 00:08:53.000 specifying good lighting. So you state the purpose 00:08:53.000 --> 00:08:55.440 clearly for a LinkedIn profile. You describe 00:08:55.440 --> 00:08:58.179 the attire Navy blue blazer, neatly combed hair, 00:08:58.500 --> 00:09:01.100 set the scene light gray office wall, softly 00:09:01.100 --> 00:09:04.360 blurred, and then nail the lighting. The lighting 00:09:04.360 --> 00:09:06.019 should look like natural light from a window 00:09:06.019 --> 00:09:08.120 hitting him from the right side. That natural 00:09:08.120 --> 00:09:10.279 light specification, that's the game changer. 00:09:10.440 --> 00:09:13.220 It immediately creates depth, looks more professional 00:09:13.220 --> 00:09:15.840 than just flat, even lighting. So what really 00:09:15.840 --> 00:09:18.500 separates that flat, maybe poorly lit snapshot 00:09:18.500 --> 00:09:20.279 from the professional look you get with Method 00:09:20.279 --> 00:09:22.740 8? It's that specified natural light source. 00:09:22.960 --> 00:09:25.820 It creates dimension, shape, and just feels more 00:09:25.820 --> 00:09:28.919 polished. OK, method nine, which you called time 00:09:28.919 --> 00:09:31.299 travel, taking modern folks and putting them 00:09:31.299 --> 00:09:35.039 into historical eras like the 1920s. This feels 00:09:35.039 --> 00:09:37.620 tricky, balancing the style change with keeping 00:09:37.620 --> 00:09:40.480 the person recognizable. It is a balancing act. 00:09:40.840 --> 00:09:43.559 The key is separating identity features from 00:09:43.559 --> 00:09:45.679 stylistic elements. You have to explicitly tell 00:09:45.679 --> 00:09:47.940 the AI what facial features to keep so you still 00:09:47.940 --> 00:09:49.919 recognize them, but then demand it change the 00:09:49.919 --> 00:09:52.960 aesthetic. Right, so alongside describing the 00:09:52.960 --> 00:09:55.059 beaded flapper dress or the pinstripe suit and 00:09:55.059 --> 00:09:57.480 maybe adding a sepia tone, the really clever 00:09:57.480 --> 00:10:00.139 command is, keep their facial features but change 00:10:00.139 --> 00:10:03.139 their hairstyles and makeup to fit the era. Exactly. 00:10:03.480 --> 00:10:06.019 And finally, method 10, completely changing the 00:10:06.019 --> 00:10:09.100 background. Moving someone, say a child, from 00:10:09.100 --> 00:10:11.559 a boring room to maybe sitting on a soft, fluffy 00:10:11.559 --> 00:10:14.379 cloud in a night sky. This really tests how well 00:10:14.379 --> 00:10:16.639 the AI handles light interaction. Yeah, you're 00:10:16.639 --> 00:10:19.519 going beyond just cutting and pasting. You define 00:10:19.519 --> 00:10:23.200 the mood, magical dreamlike atmosphere. But the 00:10:23.200 --> 00:10:26.419 truly advanced bit is commanding. The light from 00:10:26.419 --> 00:10:28.980 the moon and stars should gently light up the 00:10:28.980 --> 00:10:31.980 child's hair and clothes. That's a complex step 00:10:31.980 --> 00:10:34.580 for the AI. When you swap a background like that, 00:10:34.879 --> 00:10:37.200 why is asking the new light source to affect 00:10:37.200 --> 00:10:39.779 the original subject considered such an advanced 00:10:39.779 --> 00:10:42.659 command? Because it forces the AI to essentially 00:10:42.659 --> 00:10:45.360 run a second lighting pass on the original subject, 00:10:45.879 --> 00:10:48.000 blending the new background light with whatever 00:10:48.000 --> 00:10:50.259 light was already there. It has to harmonize 00:10:50.259 --> 00:10:52.360 two different light environments onto one object. 00:10:52.580 --> 00:10:54.620 Makes it much more believable. Okay, before we 00:10:54.620 --> 00:10:56.419 wrap up with the big takeaway, let's quickly 00:10:56.419 --> 00:10:58.980 hit those three pitfalls again. Things to avoid 00:10:58.980 --> 00:11:01.639 even if you use these formulas. Right. First, 00:11:02.100 --> 00:11:04.279 don't use obscure art styles. The AI just doesn't 00:11:04.279 --> 00:11:06.720 know them well enough. Stick to famous ones. 00:11:07.000 --> 00:11:10.580 Second, avoid super long complex mega prompts. 00:11:11.039 --> 00:11:13.379 Break down big edits into smaller sequential 00:11:13.379 --> 00:11:17.480 steps. It works better. And third, stop using 00:11:17.480 --> 00:11:21.259 vague size words like big or small. Be specific. 00:11:21.539 --> 00:11:24.580 Use powerful adjectives like gigantic, massive, 00:11:24.700 --> 00:11:27.639 or tiny. General terms just seem to confuse the 00:11:27.639 --> 00:11:30.820 AI's sense of scale. OK, let's try and synthesize 00:11:30.820 --> 00:11:33.419 this. The big lesson seems to be that unlocking 00:11:33.419 --> 00:11:35.840 NanoBanana's real power isn't luck. It's about 00:11:35.840 --> 00:11:38.639 using these kinds of effective formulas. It takes 00:11:38.639 --> 00:11:41.860 a bit of precision and patience. Totally. If 00:11:41.860 --> 00:11:44.000 you get specific about texture, about shadows, 00:11:44.100 --> 00:11:46.860 the angle of light, and critically, use those 00:11:46.860 --> 00:11:48.940 strong negative commands telling the AI what 00:11:48.940 --> 00:11:51.240 not to change your results, jump from looking 00:11:51.240 --> 00:11:54.559 kind of basic to looking photorealistic. The 00:11:54.559 --> 00:11:56.600 underlying principle is achieving coherence, 00:11:56.740 --> 00:11:58.750 especially with light and texture. Absolutely. 00:11:58.929 --> 00:12:00.889 It's really about learning to speak the AI's 00:12:00.889 --> 00:12:02.590 language when it comes to visual details, giving 00:12:02.590 --> 00:12:04.610 it the specific data points it needs to create 00:12:04.610 --> 00:12:06.690 that harmony. Which leads to a final thought, 00:12:06.809 --> 00:12:08.730 maybe something for you, the listener, to chew 00:12:08.730 --> 00:12:11.830 on. Of all the detailed commands we talked about 00:12:11.830 --> 00:12:14.669 today, matching that shadow direction, preserving 00:12:14.669 --> 00:12:17.570 the corduroy texture, specifying the window light 00:12:17.570 --> 00:12:20.409 angle, which one do you think is the single most 00:12:20.409 --> 00:12:23.429 underrated, most powerful tool for getting truly 00:12:23.429 --> 00:12:25.669 photorealistic results with AI image editing? 00:12:25.870 --> 00:12:28.429 It's worth experimenting. Maybe start simple. 00:12:28.789 --> 00:12:31.210 Pick one photo you have. Try applying just one 00:12:31.210 --> 00:12:33.750 of these specific techniques tonight. That's 00:12:33.750 --> 00:12:35.610 honestly the fastest way to level up your skills 00:12:35.610 --> 00:12:37.250 and start getting those kinds of professional 00:12:37.250 --> 00:12:38.370 results we've been talking about.