WEBVTT 00:00:00.000 --> 00:00:02.640 Have you ever had that feeling where you're using 00:00:02.640 --> 00:00:05.139 one of these, you know, incredibly powerful AI 00:00:05.139 --> 00:00:07.780 image tools, and it feels less like you're designing 00:00:07.780 --> 00:00:10.259 something and more like you're just playing a 00:00:10.259 --> 00:00:12.480 slot machine? Oh, absolutely. You type in this 00:00:12.480 --> 00:00:14.960 perfect prompt, you hit generate, and what comes 00:00:14.960 --> 00:00:17.859 back is just completely random. It's pure chance, 00:00:18.100 --> 00:00:20.739 and the frustration, it really peaks when you 00:00:20.739 --> 00:00:22.839 finally get that one great character. You know, 00:00:22.960 --> 00:00:25.960 the lighting's perfect, the mood is there, and 00:00:25.960 --> 00:00:28.160 then you just ask it to... Say, move the camera. 00:00:28.320 --> 00:00:30.219 And it all falls apart. It all falls apart. Suddenly, 00:00:30.280 --> 00:00:31.519 your character looks like a totally different 00:00:31.519 --> 00:00:33.679 person. That's that AI slop everyone talks about, 00:00:33.840 --> 00:00:36.359 that inconsistency that just kills a project. 00:00:37.799 --> 00:00:40.479 Welcome back to the Deep Dive. Today, we're on 00:00:40.479 --> 00:00:42.899 a mission, really. We want to guide you through 00:00:42.899 --> 00:00:46.020 a very specific workflow to take back directorial 00:00:46.020 --> 00:00:49.179 control over AI generation. We are moving beyond 00:00:49.179 --> 00:00:52.329 luck. The goal is to turn one single perfect 00:00:52.329 --> 00:00:55.630 image into an entire cinematic sequence. That's 00:00:55.630 --> 00:00:57.469 it, exactly. We're moving you out of the passenger 00:00:57.469 --> 00:00:59.810 seat and into the director's chair. We're going 00:00:59.810 --> 00:01:02.630 to break down this workflow using Nano Banana 00:01:02.630 --> 00:01:05.629 Pro for the images. It's powered by Gemini 3 00:01:05.629 --> 00:01:08.750 and then VO 3 .1, faster the video part. So we'll 00:01:08.750 --> 00:01:11.010 cover getting that foundation right, mastering 00:01:11.010 --> 00:01:13.530 the language of cinematography, and then the 00:01:13.530 --> 00:01:16.260 secret to getting controlled. predictable animation. 00:01:16.500 --> 00:01:18.200 Let's just dive right into the core problem. 00:01:18.379 --> 00:01:21.000 Okay, let's unpack this. Why are most of these 00:01:21.000 --> 00:01:25.540 AI tools so stubborn? The source has kept using 00:01:25.540 --> 00:01:28.659 the word inconsistency. Inconsistency is the 00:01:28.659 --> 00:01:31.500 root of all evil here. So when you create a character 00:01:31.500 --> 00:01:34.579 with some unique detail, right, like a robotic 00:01:34.579 --> 00:01:37.420 eye or a specific scar. And then you ask for 00:01:37.420 --> 00:01:40.700 a wide shot. Most models just forget those details. 00:01:41.459 --> 00:01:43.620 They resample the whole scene. It treats your 00:01:43.620 --> 00:01:45.980 new prompt like a brand new request. So it's 00:01:45.980 --> 00:01:47.760 not even that it's forgetting. It's that it's 00:01:47.760 --> 00:01:49.879 literally starting a whole new drawing from scratch. 00:01:50.040 --> 00:01:52.019 It's ignoring the face, the clothes, everything. 00:01:52.420 --> 00:01:54.099 Exactly. I mean, you watch a movie, the main 00:01:54.099 --> 00:01:56.060 actor doesn't suddenly have a different nose 00:01:56.060 --> 00:01:58.819 in the middle of a scene. We need that same level 00:01:58.819 --> 00:02:01.400 of consistency. And that's where this reference 00:02:01.400 --> 00:02:03.719 workflow comes in. Wait, so is this reference 00:02:03.719 --> 00:02:06.760 workflow just a fancy term for image to image? 00:02:06.900 --> 00:02:08.439 How is this different? How do we know it's not 00:02:08.439 --> 00:02:09.960 just going to copy the lighting and forget the 00:02:09.960 --> 00:02:11.900 character? That's the key question. A lot of 00:02:11.900 --> 00:02:14.319 those older systems, they just use the reference 00:02:14.319 --> 00:02:17.819 as a style guide. But Nano Banana Pro, using 00:02:17.819 --> 00:02:20.439 Gemini 3's architecture, is built to hold onto 00:02:20.439 --> 00:02:23.539 those key features. OK. So when you upload that 00:02:23.539 --> 00:02:26.240 reference image, your master blueprint, and you 00:02:26.240 --> 00:02:29.120 ask for a change, the model is forced to anchor 00:02:29.120 --> 00:02:32.199 the character's geometry. That robotic eye, for 00:02:32.199 --> 00:02:34.219 example, is locked in. It doesn't get washed 00:02:34.219 --> 00:02:36.340 away. So this reference workflow, it's basically 00:02:36.340 --> 00:02:38.599 the blueprint that forces consistency across 00:02:38.599 --> 00:02:40.939 all the other shots you generate. It's the only 00:02:40.939 --> 00:02:43.000 way to lock down that character's identity. And 00:02:43.000 --> 00:02:45.639 the whole thing hinges on that single foundation. 00:02:45.740 --> 00:02:48.419 Yeah. Which brings us to the most important step, 00:02:49.159 --> 00:02:52.319 creating the foundation image. If you rush this 00:02:52.319 --> 00:02:54.599 part, the whole thing just falls apart later. 00:02:54.879 --> 00:02:56.979 Right. This is the image that defines everything. 00:02:57.229 --> 00:02:59.849 The character, the lighting, the mood. All of 00:02:59.849 --> 00:03:01.729 it, the textures, everything. And the example 00:03:01.729 --> 00:03:03.229 in the sources, they didn't just use a simple 00:03:03.229 --> 00:03:06.629 portrait. They chose this really complex cyberpunk 00:03:06.629 --> 00:03:10.409 street food chef in a rainy Neo -Tokyo, which 00:03:10.409 --> 00:03:13.770 seems designed to be difficult. It's intentionally 00:03:13.770 --> 00:03:16.409 complex. That's where the real learning is. You 00:03:16.409 --> 00:03:18.729 have to deal with steam and puddles and neon 00:03:18.729 --> 00:03:21.729 lights. You can't just type cyberpunk chef and 00:03:21.729 --> 00:03:24.210 hope for the best. You need precision. Surgical 00:03:24.210 --> 00:03:26.539 precision. on the subject, the environment, and 00:03:26.539 --> 00:03:29.020 the camera. The sample prompt they gave is basically 00:03:29.020 --> 00:03:31.840 a technical checklist. It starts with the subject. 00:03:32.520 --> 00:03:35.569 Old Japanese chef. gray beard, and then the key 00:03:35.569 --> 00:03:38.830 detail. Robotic cybernetic left eye with a red 00:03:38.830 --> 00:03:41.610 lens, stained apron. That red lens, that's the 00:03:41.610 --> 00:03:43.689 anchor. That's the non -negotiable detail that 00:03:43.689 --> 00:03:45.430 AI has to hold on to. And then the environment 00:03:45.430 --> 00:03:49.310 gets just as specific. Ramen stall, heavy steam, 00:03:49.770 --> 00:03:52.909 rainy Neo -Tokyo, pink and blue neon signs blurred 00:03:52.909 --> 00:03:56.229 in the distance. If you don't lock that in, the 00:03:56.229 --> 00:03:57.969 lighting will be all over the place in your other 00:03:57.969 --> 00:04:01.189 shots. And finally, the camera. Medium shot, 00:04:01.409 --> 00:04:04.289 eye level, cinematic lighting. Photorealistic. 00:04:04.729 --> 00:04:07.550 Once you get that perfect image, that one file 00:04:07.550 --> 00:04:09.430 will become your North Star. So that's the one 00:04:09.430 --> 00:04:11.270 you keep coming back to. It's the one you upload 00:04:11.270 --> 00:04:13.650 to the reference section every single time. I 00:04:13.650 --> 00:04:17.449 still wrestle with prompt drift myself. I once 00:04:17.449 --> 00:04:20.170 spent three hours trying to get consistent shots 00:04:20.170 --> 00:04:22.589 because I was too lazy to perfect my first prompt. 00:04:22.670 --> 00:04:25.389 An entire afternoon wasted. Exactly. I learned 00:04:25.389 --> 00:04:28.050 my lesson. You have to invest the time upfront. 00:04:28.279 --> 00:04:31.779 in that foundation image. So extreme specificity 00:04:31.779 --> 00:04:34.360 on subject, environment, and camera angle is 00:04:34.360 --> 00:04:36.600 what prevents all those inconsistencies down 00:04:36.600 --> 00:04:38.680 the line. OK, this is where it gets really interesting. 00:04:39.240 --> 00:04:41.420 We're shifting from being a prompt engineer to 00:04:41.420 --> 00:04:44.000 being a director. We have our chef. He's consistent. 00:04:44.279 --> 00:04:46.480 Now we need to build a shot library around him. 00:04:46.579 --> 00:04:47.920 And to do that, we have to learn to speak the 00:04:47.920 --> 00:04:49.800 language of cameras, because these tools, like 00:04:49.800 --> 00:04:52.879 Nano Banana Pro, they understand cinematic terms 00:04:52.879 --> 00:04:54.959 better than just vague descriptions. We need 00:04:54.959 --> 00:04:57.339 to start big. The extreme wide shot. This isn't 00:04:57.339 --> 00:04:59.810 just about showing the scenery, is it? No. It's 00:04:59.810 --> 00:05:01.990 an establishing shot. It tells a story. It shows 00:05:01.990 --> 00:05:05.750 the chef as this small human detail against these 00:05:05.750 --> 00:05:08.470 massive dark skyscrapers. It sets the stakes. 00:05:08.850 --> 00:05:10.490 Then you bring in the drama with a low angle 00:05:10.490 --> 00:05:12.870 shot. Right. You want the chef to look powerful, 00:05:12.970 --> 00:05:15.310 like a master of his craft. You look up at him. 00:05:15.470 --> 00:05:17.550 You prompt for looking up from the counter level. 00:05:17.910 --> 00:05:20.569 And the sources said the AI kept that red cybernetic 00:05:20.569 --> 00:05:24.339 eye perfectly, even from below. The anchor held. 00:05:24.720 --> 00:05:26.560 And you can flip that for a high angle. That's 00:05:26.560 --> 00:05:28.879 more for vulnerability or for showing information, 00:05:29.040 --> 00:05:30.759 right? Looking down at his hands as he prepares 00:05:30.759 --> 00:05:33.579 the food. And my personal favorite for creating 00:05:33.579 --> 00:05:35.959 tension is the Dutch angle. The tilted shot. 00:05:36.199 --> 00:05:38.139 Yeah, the canted shot. You tell the camera to 00:05:38.139 --> 00:05:40.980 tilt 30 degrees and it immediately puts the viewer 00:05:40.980 --> 00:05:43.199 on edge. It signals that something is unstable. 00:05:43.500 --> 00:05:46.620 You can even add heavy rain falling diagonally 00:05:46.620 --> 00:05:49.459 to really push that dynamic energy. The AI knows 00:05:49.459 --> 00:05:52.180 exactly what Dutch angle means. Whoa. I mean, 00:05:52.259 --> 00:05:54.779 just imagine scaling this. If you can generate 00:05:54.779 --> 00:05:57.800 10 or 15 perfectly consistent shots like this, 00:05:58.079 --> 00:06:00.699 all ready to be dropped into an editor, that's 00:06:00.699 --> 00:06:03.139 real cinematic production power. It is. Then 00:06:03.139 --> 00:06:05.199 you need connection. So you use the over the 00:06:05.199 --> 00:06:07.860 shoulder shot, you prompt for a blurry customer 00:06:07.860 --> 00:06:10.360 in the foreground, and it pulls the viewer right 00:06:10.360 --> 00:06:12.819 into that interaction as the chef hands over 00:06:12.819 --> 00:06:16.360 the ramen. And finally, the really intimate detail 00:06:16.360 --> 00:06:19.680 shot, the macro shot. Yeah, the extreme close 00:06:19.680 --> 00:06:21.560 -up. For the cybernetic eye, you don't just say 00:06:21.560 --> 00:06:24.420 macro. You have to get technical. Increase the 00:06:24.420 --> 00:06:26.920 depth of field prompt. Emphasize the texture 00:06:26.920 --> 00:06:29.939 of the lens, the reflection of the neon. So the 00:06:29.939 --> 00:06:32.000 big takeaway is you're building this library, 00:06:32.379 --> 00:06:35.399 this shot library of 10, maybe 15 consistent 00:06:35.399 --> 00:06:38.740 angles all anchored to that one North Star image. 00:06:39.079 --> 00:06:41.600 Then you're ready for motion. Exactly. And if 00:06:41.600 --> 00:06:44.000 you really want to convey that emotional intensity 00:06:44.000 --> 00:06:47.180 or instability, the Dutch angle is the classic 00:06:47.180 --> 00:06:49.639 cinematic tool for the job. Now we make the move. 00:06:49.819 --> 00:06:53.439 We switch over to VO 3 .1 fast because it's quick 00:06:53.439 --> 00:06:55.790 and it, well. It understands physics pretty well. 00:06:56.009 --> 00:06:57.730 So there are two methods here. The simple one 00:06:57.730 --> 00:07:00.550 is method A, the start frame. Right. You give 00:07:00.550 --> 00:07:02.529 via one of your images and a really simple prompt 00:07:02.529 --> 00:07:04.949 like steam rises or rain falls. It's good for 00:07:04.949 --> 00:07:06.949 adding a bit of atmosphere, just pushing the 00:07:06.949 --> 00:07:09.649 sled a little. But method B is the real secret 00:07:09.649 --> 00:07:12.370 sauce for anything complex. We call it the start 00:07:12.370 --> 00:07:15.389 and end frame or laying the train track. OK, 00:07:15.550 --> 00:07:18.410 explain that. Well, if you just tell the AI chef 00:07:18.410 --> 00:07:21.720 walks away, the model has to guess. It starts 00:07:21.720 --> 00:07:24.360 to improvise the physics, the trajectory, and 00:07:24.360 --> 00:07:26.120 that's when you get that weird, unpredictable 00:07:26.120 --> 00:07:29.139 morphing. The AI slop. The slop comes back. But 00:07:29.139 --> 00:07:31.120 if you give it a perfect start frame our chef 00:07:31.120 --> 00:07:34.019 standing, and a perfect end frame our chef walking 00:07:34.019 --> 00:07:37.160 into the distance, both already generated consistently 00:07:37.160 --> 00:07:39.879 in NanoBanana Pro, then I have no choice. It 00:07:39.879 --> 00:07:41.680 has no choice but to connect the dots. It's like 00:07:41.680 --> 00:07:44.100 a train on a fixed track. It guarantees you have 00:07:44.100 --> 00:07:46.519 control over big camera moves, complex actions, 00:07:46.819 --> 00:07:48.680 anything like that. So the two -frame method 00:07:48.680 --> 00:07:51.319 is so vital because it forces the AI to connect 00:07:51.319 --> 00:07:54.139 two consistent states, which stops it from just 00:07:54.139 --> 00:07:57.279 making things up and morphing. Let's talk about 00:07:57.279 --> 00:08:00.600 a hack here for efficiency. The source is mentioned 00:08:00.600 --> 00:08:03.839 using an LLM, like Claude or GPT, to actually 00:08:03.839 --> 00:08:06.540 write your prompts for you. This is a huge time 00:08:06.540 --> 00:08:09.279 saver. Because trying to describe a subtle camera 00:08:09.279 --> 00:08:11.540 move in just the right words can be really hard 00:08:11.540 --> 00:08:13.939 for a person. So you upload your two frames, 00:08:14.240 --> 00:08:17.220 say, the macro shot of the eye. and medium shot 00:08:17.220 --> 00:08:20.319 of the face. And you just ask the LLM to write 00:08:20.319 --> 00:08:23.199 a concise prompt that explains the camera move 00:08:23.199 --> 00:08:25.240 needed to get from A to B. And you just copy 00:08:25.240 --> 00:08:27.279 and paste that directly into Vio. It takes all 00:08:27.279 --> 00:08:29.579 the guesswork out of it. And it lets you do really 00:08:29.579 --> 00:08:32.919 professional tricks like a rack focus. You generate 00:08:32.919 --> 00:08:35.220 image A, where the chef is sharp and the background 00:08:35.220 --> 00:08:37.799 is blurry, and image B, where the chef is blurry 00:08:37.799 --> 00:08:39.580 and the neon signs are sharp. And the prompt 00:08:39.580 --> 00:08:43.019 is just... The prompt is rack focus from the 00:08:43.019 --> 00:08:45.240 foreground character to the background neon sign. 00:08:45.839 --> 00:08:48.039 And it works perfectly because both of those 00:08:48.039 --> 00:08:50.419 states, A and B, came from the same foundation 00:08:50.419 --> 00:08:53.320 image. Even with this whole system though, there 00:08:53.320 --> 00:08:55.600 have to be common mistakes. What are the big 00:08:55.600 --> 00:08:57.720 pitfalls people need to watch out for? Number 00:08:57.720 --> 00:09:00.480 one is forgetting the foundation. You get excited, 00:09:00.820 --> 00:09:02.440 you try to generate a new shot, but you forget 00:09:02.440 --> 00:09:04.480 to put your North Star image in the reference 00:09:04.480 --> 00:09:06.820 slot. And the consistency breaks. Instantly. 00:09:07.559 --> 00:09:10.320 The fix is just discipline. Always, always use 00:09:10.320 --> 00:09:12.639 the foundation image as your anchor. Okay, what's 00:09:12.639 --> 00:09:15.039 mistake number two? Being lazy with prompts, 00:09:15.200 --> 00:09:17.399 just typing something like different angle or 00:09:17.399 --> 00:09:20.259 move camera a bit, it just confuses the model. 00:09:20.379 --> 00:09:22.220 You have to be specific. You have to use the 00:09:22.220 --> 00:09:25.000 professional language. Low angle, Broca, depth 00:09:25.000 --> 00:09:27.559 of field. You have to talk to the tool like a 00:09:27.559 --> 00:09:29.879 director talking to a cinematographer. Mistake 00:09:29.879 --> 00:09:33.240 three, the morphing hands. Hands are still a 00:09:33.240 --> 00:09:35.340 huge challenge for AI, aren't they? They are. 00:09:35.659 --> 00:09:37.919 If your chef is doing anything complex, like 00:09:37.919 --> 00:09:40.539 chopping or stirring, you absolutely have to 00:09:40.539 --> 00:09:43.100 use the start and end frame method for that specific 00:09:43.100 --> 00:09:46.100 action. It's the only way to minimize that weird 00:09:46.100 --> 00:09:48.059 digital distortion. And the last one, which you 00:09:48.059 --> 00:09:50.320 said is the most common reason a video feels 00:09:50.320 --> 00:09:53.940 amateur. Ignoring lighting consistency. If your 00:09:53.940 --> 00:09:56.720 wide shot is at night, your close -up can't suddenly 00:09:56.720 --> 00:09:58.980 look like it's daytime. It breaks the whole illusion. 00:09:59.220 --> 00:10:01.580 And the fix is just repetition. Just repetition. 00:10:01.700 --> 00:10:04.240 You have to repeat the exact atmospheric keywords 00:10:04.240 --> 00:10:07.860 in every single prompt. For our chef, that was 00:10:07.860 --> 00:10:10.799 neon lights, rainy night, blue and pink hues 00:10:10.799 --> 00:10:14.179 every single time, no matter the camera angle. 00:10:14.299 --> 00:10:16.399 That lack of repetition is what makes so much 00:10:16.399 --> 00:10:19.340 AI video feel so disjointed. It's the biggest 00:10:19.340 --> 00:10:21.860 giveaway of amateur work. You have to lock that 00:10:21.860 --> 00:10:23.919 atmosphere down. So when you put it all together, 00:10:23.980 --> 00:10:25.600 what does this actually mean? The sources seem 00:10:25.600 --> 00:10:29.009 to be saying that the era of AI video looking 00:10:29.009 --> 00:10:32.950 like a, you know, weird dream is coming to an 00:10:32.950 --> 00:10:36.309 end. It is. We are moving into an era of AI directing. 00:10:37.450 --> 00:10:39.769 Control is finally starting to outweigh chance. 00:10:40.190 --> 00:10:41.809 The recap is pretty straightforward, then. It 00:10:41.809 --> 00:10:45.750 is. One perfect foundation image to lock in your 00:10:45.750 --> 00:10:49.259 character. Use that to build a solid shot library 00:10:49.259 --> 00:10:52.519 of different cinematic angles and then animate 00:10:52.519 --> 00:10:54.940 it all with the start and end frame method for 00:10:54.940 --> 00:10:57.139 total control over the motion. It's a much more 00:10:57.139 --> 00:10:59.139 deliberate process than just hitting generate 00:10:59.139 --> 00:11:01.460 over and over, but the payoff is moving from 00:11:01.460 --> 00:11:04.480 random clips to actual controlled storytelling. 00:11:04.639 --> 00:11:07.139 Your story told exactly how you see it in your 00:11:07.139 --> 00:11:09.320 head. With that cinematic consistency completely 00:11:09.320 --> 00:11:11.639 locked in. So the next step for you, the listener, 00:11:11.779 --> 00:11:13.639 is to actually try this. Just do the first part. 00:11:13.980 --> 00:11:17.139 Open up the tool and spend some real time making 00:11:17.139 --> 00:11:20.679 one perfect character, your North Star. And then 00:11:20.679 --> 00:11:23.820 just use that one file to generate three consistent 00:11:23.820 --> 00:11:28.000 angles, a wide, a medium, and a close up. And 00:11:28.000 --> 00:11:30.620 just feel that shift in control. It's a powerful 00:11:30.620 --> 00:11:32.360 feeling. And here's a final thought to leave 00:11:32.360 --> 00:11:34.779 you with. Now that you understand the cinematic 00:11:34.779 --> 00:11:37.519 language that a low angle means power, a Dutch 00:11:37.519 --> 00:11:40.419 angle means tension. Try applying that to the 00:11:40.419 --> 00:11:42.879 real world. Think about it when you're just taking 00:11:42.879 --> 00:11:45.139 a photo with your phone. It really does change 00:11:45.139 --> 00:11:47.480 how you see everything around you. Until next 00:11:47.480 --> 00:11:49.639 time. Thank you for joining us on this deep dive.