WEBVTT 00:00:00.000 --> 00:00:02.560 Think about the pure chaos of a machine trying 00:00:02.560 --> 00:00:04.679 to dream. Oh, absolutely. It gets messy fast. 00:00:05.040 --> 00:00:07.219 You picture a cinematic masterpiece in your head. 00:00:07.280 --> 00:00:09.960 You type it out with perfect clarity. You hit 00:00:09.960 --> 00:00:13.259 generate. And what you get back is a total fever 00:00:13.259 --> 00:00:17.410 dream. Faces morph into strange... unrecognizable 00:00:17.410 --> 00:00:20.550 shapes. Right. And cars turn into these unidentifiable 00:00:20.550 --> 00:00:23.989 melting blobs. The action sequence makes absolutely 00:00:23.989 --> 00:00:26.710 no physical sense. It really does. The continuity 00:00:26.710 --> 00:00:29.030 is just entirely broken. The lighting shifts. 00:00:29.250 --> 00:00:32.350 The camera angles just defy gravity. But it honestly 00:00:32.350 --> 00:00:34.549 doesn't have to be that way anymore. I recently 00:00:34.549 --> 00:00:38.689 watched a truly seamless AI short film. It had 00:00:38.689 --> 00:00:41.609 perfect locked -in continuity. Wait, really? 00:00:41.729 --> 00:00:44.090 Perfect continuity? Yeah. It featured a moving 00:00:44.090 --> 00:00:46.590 armored convoy with real physical weight. It 00:00:46.590 --> 00:00:49.109 even had baked -in audio. The dialogue actually 00:00:49.109 --> 00:00:51.390 matched the lip movements perfectly. It feels 00:00:51.390 --> 00:00:53.630 like actual magic when it finally works. Welcome 00:00:53.630 --> 00:00:56.109 to another Deep Tech. Glad to be here. Today, 00:00:56.229 --> 00:00:58.509 we are exploring something very specific and 00:00:58.509 --> 00:01:01.890 very powerful. We are unpacking the reference 00:01:01.890 --> 00:01:05.879 -first AI short film workflow. Our mission is 00:01:05.879 --> 00:01:09.200 to guide you through this exact process. We want 00:01:09.200 --> 00:01:11.659 to bridge that gap between chaotic AI generation 00:01:11.659 --> 00:01:14.780 and true cinematic control. We are going to look 00:01:14.780 --> 00:01:17.340 at a few critical steps today. We will cover 00:01:17.340 --> 00:01:20.560 the beginner's mega prompt mistake. We will talk 00:01:20.560 --> 00:01:23.299 about building solid visual references. We will 00:01:23.299 --> 00:01:26.739 discuss generating the actual video files. Then 00:01:26.739 --> 00:01:29.079 we tackle maintaining that incredibly tricky 00:01:29.079 --> 00:01:32.019 continuity. Finally, we look at the simple editing 00:01:32.019 --> 00:01:33.980 phase. Let's unpack this core principle first. 00:01:34.400 --> 00:01:37.879 The idea of control versus chaos. Most beginners 00:01:37.879 --> 00:01:39.959 start entirely in the wrong place. They jump 00:01:39.959 --> 00:01:41.760 straight into the video generator. Yeah, and 00:01:41.760 --> 00:01:44.079 it is a guaranteed recipe for immediate failure. 00:01:44.340 --> 00:01:45.840 I think we need to understand the psychology 00:01:45.840 --> 00:01:48.790 here. We want the technology to be magic. So 00:01:48.790 --> 00:01:51.609 beginners write one giant, overly detailed prompt. 00:01:51.810 --> 00:01:53.989 Oh, the mega prompt. Exactly. They type something 00:01:53.989 --> 00:01:55.989 like serious soldier on a desert bridge. They 00:01:55.989 --> 00:01:58.730 add explosions, dramatic lighting, fast Hollywood 00:01:58.730 --> 00:02:01.469 pacing. They just expect the AI to handle the 00:02:01.469 --> 00:02:04.549 rest of the movie. But text alone cannot lock 00:02:04.549 --> 00:02:07.989 down complex visuals. Text is essentially a low 00:02:07.989 --> 00:02:11.050 bandwidth communication method. The AI engine 00:02:11.050 --> 00:02:13.409 is basically just guessing at the details. Because 00:02:13.409 --> 00:02:15.090 it doesn't actually know what you want. Right. 00:02:15.229 --> 00:02:17.550 It has to invent the character's facial structure. 00:02:17.729 --> 00:02:20.229 It guesses the outfit, the specific location, 00:02:20.389 --> 00:02:22.710 and the background props. It even has to guess 00:02:22.710 --> 00:02:25.169 the timing and the physics. I have a vulnerable 00:02:25.169 --> 00:02:27.409 admission to make here. I still wrestle with 00:02:27.409 --> 00:02:30.439 prompt drift myself. It's so tempting to just 00:02:30.439 --> 00:02:33.439 ask the AI to read my mind. We all do it. We 00:02:33.439 --> 00:02:35.560 want that immediate dopamine hit. We want the 00:02:35.560 --> 00:02:38.319 quick shortcut. But skipping the foundational 00:02:38.319 --> 00:02:41.379 work always costs you hours of time later. It's 00:02:41.379 --> 00:02:43.500 kind of like giving a chef a list of ingredients 00:02:43.500 --> 00:02:46.280 but no recipe. You just expect a Michelin star 00:02:46.280 --> 00:02:48.960 meal to magically appear. That is a perfect analogy. 00:02:49.360 --> 00:02:51.479 Let's talk about the mechanics of why that fails. 00:02:51.719 --> 00:02:54.680 The AI is a probabilistic engine, right? It starts 00:02:54.680 --> 00:02:57.479 with pure visual noise. It resolves that static 00:02:57.479 --> 00:03:00.400 into an image based on your text. Yes, step by 00:03:00.400 --> 00:03:03.060 step. Exactly. If your text just says serious 00:03:03.060 --> 00:03:05.740 soldier, there are millions of soldiers in its 00:03:05.740 --> 00:03:08.939 latent space. The AI picks one variation entirely 00:03:08.939 --> 00:03:12.020 at random. In the very next frame, it picks a 00:03:12.020 --> 00:03:14.319 slightly different variation. And that leads 00:03:14.319 --> 00:03:16.599 to the shifting visuals you mentioned. Clothes 00:03:16.599 --> 00:03:19.060 change color across different cliffs. Cars just 00:03:19.060 --> 00:03:22.020 mysteriously disappear from the road. The model 00:03:22.020 --> 00:03:24.740 simply forgets what it just drew. So by front 00:03:24.740 --> 00:03:26.740 -loading the visuals, we're basically taking 00:03:26.740 --> 00:03:29.020 all the dangerous guesswork away from the AI 00:03:29.020 --> 00:03:31.639 engine. Precisely. You lock the visual parameters 00:03:31.639 --> 00:03:34.560 down tight. The AI no longer has to invent every 00:03:34.560 --> 00:03:36.939 single pixel from scratch. Right. We define the 00:03:36.939 --> 00:03:39.000 world first so the AI doesn't have to. Exactly. 00:03:39.319 --> 00:03:42.139 Two -sec silence. So let's talk about how to 00:03:42.139 --> 00:03:44.569 actually do that. This brings us to step one 00:03:44.569 --> 00:03:47.569 of the workflow. You need to meticulously build 00:03:47.569 --> 00:03:50.250 your foundational assets. We are talking about 00:03:50.250 --> 00:03:53.550 the holy trinity of film. Characters, locations, 00:03:53.710 --> 00:03:56.729 and props. Yes, and you want to use tools specifically 00:03:56.729 --> 00:03:59.770 designed for cinematic results. The source material 00:03:59.770 --> 00:04:02.229 recommends using Higgs field inside Cinema Studio. 00:04:02.610 --> 00:04:05.289 Let's pause and clarify that environment. Cinema 00:04:05.289 --> 00:04:07.509 Studio is essentially a professional generative 00:04:07.509 --> 00:04:10.370 workspace. It handles the heavy lifting behind 00:04:10.370 --> 00:04:13.069 the scenes. You set the model to auto inside 00:04:13.069 --> 00:04:17.550 that workspace. You choose a 9 .16 aspect ratio 00:04:17.550 --> 00:04:20.290 for vertical video. And you set the output resolution 00:04:20.290 --> 00:04:25.009 to a crisp 4K. The auto model is a fascinating 00:04:25.009 --> 00:04:27.870 choice. It removes unnecessary decision -making 00:04:27.870 --> 00:04:30.129 from the user. You just describe what you want 00:04:30.129 --> 00:04:32.350 to see. Yeah, the system automatically routes 00:04:32.350 --> 00:04:34.649 the prompt to the most suitable underlying generator. 00:04:34.949 --> 00:04:37.490 It streamlines all the technical friction. Let's 00:04:37.490 --> 00:04:39.230 talk about building the actual characters next. 00:04:39.709 --> 00:04:41.550 In this short film example, we have two main 00:04:41.550 --> 00:04:43.839 guys. We have Ryder. who is our tactical commander. 00:04:44.120 --> 00:04:46.560 For Ryder, the workflow suggests using a reference 00:04:46.560 --> 00:04:49.459 image. You start with a stock -style photo of 00:04:49.459 --> 00:04:52.279 a serious commander. You want his realistic face 00:04:52.279 --> 00:04:55.160 clearly visible. You know the dark hair and the 00:04:55.160 --> 00:04:57.899 light stubble. You specify modern black tactical 00:04:57.899 --> 00:05:00.259 gear. Having that anchor image keeps the character 00:05:00.259 --> 00:05:03.300 incredibly consistent. The AI studies the facial 00:05:03.300 --> 00:05:05.879 geometry. But what about the other character? 00:05:06.000 --> 00:05:09.079 We have a snoper named Vance. For Vance, the 00:05:09.079 --> 00:05:11.259 approach is entirely different. You don't start 00:05:11.259 --> 00:05:13.740 with a downloaded reference image. You use an 00:05:13.740 --> 00:05:17.500 integrated tool called AI Cast instead. AI Cast 00:05:17.500 --> 00:05:20.180 is incredibly powerful for world building. Okay, 00:05:20.240 --> 00:05:22.920 let's define that term for clarity. What exactly 00:05:22.920 --> 00:05:25.980 is AI Cast? A preset menu to build character 00:05:25.980 --> 00:05:28.779 looks without typing long descriptions. That 00:05:28.779 --> 00:05:30.839 sounds incredibly useful for maintaining sanity. 00:05:31.339 --> 00:05:33.319 You just pick the genre from a drop -down, like 00:05:33.319 --> 00:05:35.399 action. Yeah, you even set a virtual production 00:05:35.399 --> 00:05:38.560 budget, like, say, $200 million. That budget 00:05:38.560 --> 00:05:41.160 setting tells the AI to aim for high -end aesthetics. 00:05:41.500 --> 00:05:44.639 You choose an archetype, like the sage. You pick 00:05:44.639 --> 00:05:47.300 a white male in his 40s. Right, and you can seamlessly 00:05:47.300 --> 00:05:50.240 add details, like facial scars or a rugged beard. 00:05:50.379 --> 00:05:52.459 It feels much more like casting a real human 00:05:52.459 --> 00:05:55.079 actor. Beat, that covers the characters. Now 00:05:55.079 --> 00:05:57.500 we need to establish the location. The main scene 00:05:57.500 --> 00:05:59.600 happens on a highly cinematic desert bridge. 00:05:59.779 --> 00:06:02.019 You use the cinematic locations mode for this 00:06:02.019 --> 00:06:05.040 specific task. You describe a wide, imposing 00:06:05.040 --> 00:06:07.779 desert bridge stretching over a deep canyon. 00:06:08.079 --> 00:06:11.540 You specify harsh noon sunlight. You add a dusty, 00:06:11.660 --> 00:06:14.310 atmospheric haze to the air. You generate that 00:06:14.310 --> 00:06:17.550 pristine, clean version of the bridge. But here 00:06:17.550 --> 00:06:19.850 is where the workflow gets really interesting. 00:06:20.170 --> 00:06:22.750 You also have to create a mathematically damaged 00:06:22.750 --> 00:06:25.910 version of it. This step is absolutely crucial 00:06:25.910 --> 00:06:28.790 for the later action scenes. You upload that 00:06:28.790 --> 00:06:31.589 clean image right back into the system. You explicitly 00:06:31.589 --> 00:06:35.769 ask the AI to add severe explosion damage. You 00:06:35.769 --> 00:06:38.879 want broken structural steel. deep burn marks, 00:06:39.019 --> 00:06:41.839 and lingering smoke, but you give the engine 00:06:41.839 --> 00:06:44.980 one non -negotiable instruction. Do not change 00:06:44.980 --> 00:06:47.459 the original bridge structure. Exactly. Whoa. 00:06:48.250 --> 00:06:50.389 Imagine generating the exact explosion damage 00:06:50.389 --> 00:06:53.009 before the explosion even happens. It genuinely 00:06:53.009 --> 00:06:55.329 feels like you are cheating time. You are building 00:06:55.329 --> 00:06:57.449 the future aftermath before the event occurs, 00:06:57.689 --> 00:07:00.149 but it is absolutely essential for the logic 00:07:00.149 --> 00:07:02.290 of the workflow. Why not just let the video model 00:07:02.290 --> 00:07:04.389 figure out what a destroyed bridge looks like 00:07:04.389 --> 00:07:06.389 when the bomb goes off? Well, if you let the 00:07:06.389 --> 00:07:08.470 video model invent the destruction dynamically, 00:07:08.850 --> 00:07:12.029 it loses context. It might completely redesign 00:07:12.029 --> 00:07:14.550 the entire canyon in the background. It might 00:07:14.550 --> 00:07:16.829 accidentally change the time of day. You lose 00:07:16.829 --> 00:07:19.540 the continuity. Exactly. Because the AI might 00:07:19.540 --> 00:07:21.300 generate a completely different bridge altogether. 00:07:21.600 --> 00:07:24.060 That makes sense. Finally, you need to establish 00:07:24.060 --> 00:07:27.019 your reusable props. In this specific case, it 00:07:27.019 --> 00:07:30.100 is a heavy armored convoy car. You want a clean, 00:07:30.100 --> 00:07:33.480 side -front view of this exact vehicle. You use 00:07:33.480 --> 00:07:35.660 a highly realistic model for this generation. 00:07:36.060 --> 00:07:38.000 The source suggests something sophisticated, 00:07:38.339 --> 00:07:42.480 like Soul Cinema 2K. You just want a clear, well 00:07:42.480 --> 00:07:45.899 -lit image of the physical object, heavy military 00:07:45.899 --> 00:07:48.970 design elements, reinforced bulletproof windows, 00:07:49.310 --> 00:07:52.730 rugged oversized tires. You save that image alongside 00:07:52.730 --> 00:07:55.269 your other visual references. Let's quickly review 00:07:55.269 --> 00:07:57.529 the assets we have gathered. We have Ryder, we 00:07:57.529 --> 00:07:59.810 have Vance, we have a pristine, clean bridge, 00:07:59.970 --> 00:08:02.269 we have a heavily damaged bridge, and we have 00:08:02.269 --> 00:08:05.129 our armored car. That structural logic is exactly 00:08:05.129 --> 00:08:08.529 what we need. Beat. We will get into the video 00:08:08.529 --> 00:08:10.569 generation phase right after this quick break. 00:08:11.170 --> 00:08:13.230 Sponsor. All right, so we have our visual bays 00:08:13.230 --> 00:08:15.649 officially locked. We are finally ready for step 00:08:15.649 --> 00:08:18.389 two. generating the actual video sequence. This 00:08:18.389 --> 00:08:20.949 is where we move from static pictures to fluid 00:08:20.949 --> 00:08:24.230 moving scenes. We use a very smart prompt structure 00:08:24.230 --> 00:08:27.110 for this phase. You open your dedicated AI movie 00:08:27.110 --> 00:08:29.910 generator. The source material uses a platform 00:08:29.910 --> 00:08:32.769 called C -Dance. You start by opening the director 00:08:32.769 --> 00:08:35.549 panel. You set the overarching genre to action. 00:08:35.950 --> 00:08:38.809 You choose smart for the integrated shot control. 00:08:39.029 --> 00:08:41.429 Smart control is infinitely better for beginners. 00:08:41.980 --> 00:08:44.019 It handles the complex camera angles and virtual 00:08:44.019 --> 00:08:46.519 movement automatically. Because manual camera 00:08:46.519 --> 00:08:50.200 control involves deep math. It's just too complex 00:08:50.200 --> 00:08:52.679 at first. Yeah, exactly. You set the clip duration 00:08:52.679 --> 00:08:55.379 to 10 or 15 seconds, and you critically turn 00:08:55.379 --> 00:08:58.320 the audio on in. That audio setting is a massive 00:08:58.320 --> 00:09:01.639 paradigm shift. Then you feed your visual references 00:09:01.639 --> 00:09:04.659 into the AI system. You upload the images of 00:09:04.659 --> 00:09:07.840 Ryder, Vance, the clean bridge, and the armored 00:09:07.840 --> 00:09:10.440 car. This shift in the workflow is huge. You 00:09:10.440 --> 00:09:12.519 are no longer desperately describing writer with 00:09:12.519 --> 00:09:14.860 text. You are actively selecting writer from 00:09:14.860 --> 00:09:17.100 your pre -built assets. Next, you have to carefully 00:09:17.100 --> 00:09:18.899 set the character emotions. You click the little 00:09:18.899 --> 00:09:21.419 parameter icon next to their faces. For the quiet 00:09:21.419 --> 00:09:24.840 opening scene, you choose medium vigilance. Beginners 00:09:24.840 --> 00:09:27.100 almost always push the emotional intensity too 00:09:27.100 --> 00:09:29.919 high. They choose maximum rage or sheer panic 00:09:29.919 --> 00:09:33.360 right away. Right, but high intensity looks incredibly 00:09:33.360 --> 00:09:36.399 cartoonish and dramatic for a quiet setup. A 00:09:36.399 --> 00:09:38.899 medium emotion level feels so much more grounded 00:09:38.899 --> 00:09:42.039 and natural. Beat. Then we get to the actual 00:09:42.039 --> 00:09:44.200 text prompt. We use something called the multi 00:09:44.200 --> 00:09:46.500 -shot framework. You break the written prompt 00:09:46.500 --> 00:09:49.779 into clear, numbered shots. Shot one, shot two, 00:09:49.879 --> 00:09:52.100 shot three. You describe the specific camera 00:09:52.100 --> 00:09:54.840 action for each individual shot. And you put 00:09:54.840 --> 00:09:57.600 the spoken dialogue lines directly in quotation 00:09:57.600 --> 00:09:59.860 marks. It is brilliant. Instead of throwing a 00:09:59.860 --> 00:10:01.940 whole bucket of text at the wall, it's like stacking 00:10:01.940 --> 00:10:04.919 Lego blocks of data. That is exactly how it feels. 00:10:05.080 --> 00:10:07.779 It is highly structured and remarkably orderly. 00:10:07.860 --> 00:10:10.419 Shot one is the establishing wide shot of the 00:10:10.419 --> 00:10:14.100 bridge. Shot two focuses on Ryder crouching behind 00:10:14.100 --> 00:10:17.779 tactical cover. Shot three reveals Vance locked 00:10:17.779 --> 00:10:20.399 in his high sniper position. And then shot four 00:10:20.399 --> 00:10:23.100 delivers the actual dialogue. Ryder says, everyone 00:10:23.100 --> 00:10:26.200 in position. The AI follows this logical sequence 00:10:26.200 --> 00:10:28.019 step by step. It doesn't get overwhelmed and 00:10:28.019 --> 00:10:30.659 confused by a massive wall of text. If we leave 00:10:30.659 --> 00:10:33.039 the audio on end during generation, are we getting 00:10:33.360 --> 00:10:35.580 Actual usable dialogue right out of the box. 00:10:35.740 --> 00:10:38.480 Yes, the tool generates the vocal dialogue internally. 00:10:38.840 --> 00:10:41.899 It synthesizes the voice and adds ambient environmental 00:10:41.899 --> 00:10:45.340 sounds. It bakes those basic effects right into 00:10:45.340 --> 00:10:47.340 the final video file. So you don't have to record 00:10:47.340 --> 00:10:49.600 separate voice lines later. Exactly. Yeah, so 00:10:49.600 --> 00:10:51.480 it bakes in the voices and sound effects automatically. 00:10:51.779 --> 00:10:56.220 Huge time saver. Two sec silence. We generated 00:10:56.220 --> 00:10:59.419 our first successful scene, but one cool clip 00:10:59.419 --> 00:11:02.740 doesn't make a coherent movie. Step three is 00:11:02.740 --> 00:11:05.399 keeping every subsequent clip logically connected. 00:11:05.639 --> 00:11:08.120 This is exactly where most AI films completely 00:11:08.120 --> 00:11:10.899 fall apart. You string two generated clips together 00:11:10.899 --> 00:11:14.220 and the cut feels entirely wrong. The underlying 00:11:14.220 --> 00:11:17.100 mood suddenly resets to zero. People usually 00:11:17.100 --> 00:11:19.539 try the outdated traditional method first. They 00:11:19.539 --> 00:11:22.159 take the very last frame of clip one. They use 00:11:22.159 --> 00:11:24.700 that single image to start clip two. But that 00:11:24.700 --> 00:11:27.940 methodology fundamentally fails. A single still 00:11:27.940 --> 00:11:30.779 image cannot carry emotional tension. It only 00:11:30.779 --> 00:11:33.600 captures one static, frozen moment in time. It 00:11:33.600 --> 00:11:36.080 has absolutely no velocity. You need the next 00:11:36.080 --> 00:11:38.200 clip to implicitly remember what just physically 00:11:38.200 --> 00:11:40.480 happened. That is exactly why you use video reference 00:11:40.480 --> 00:11:43.019 instead. Let's define that critical concept quickly. 00:11:43.259 --> 00:11:46.139 What is video reference? Feeding the whole previous 00:11:46.139 --> 00:11:48.929 video to keep the exact same mood. So the AI 00:11:48.929 --> 00:11:51.629 mathematically reads the scene's pacing. It reads 00:11:51.629 --> 00:11:54.789 the ongoing camera motion. It reads the lingering 00:11:54.789 --> 00:11:57.269 tension from the previous scene. Let's dig into 00:11:57.269 --> 00:11:59.289 the mechanics of that. How does it actually read 00:11:59.289 --> 00:12:02.139 tension? Well, it reads the pixel movement and 00:12:02.139 --> 00:12:05.500 the audio waveforms. It analyzes the speed and 00:12:05.500 --> 00:12:08.039 direction of the motion vectors. So yes, it mathematically 00:12:08.039 --> 00:12:10.299 calculates and continues the kinetic energy. 00:12:10.519 --> 00:12:13.600 Clip two uses clip one as its direct video reference. 00:12:13.860 --> 00:12:17.460 This is our big explosion scene. You change writer's 00:12:17.460 --> 00:12:20.379 underlying emotion from vigilance to pure rage. 00:12:20.600 --> 00:12:23.399 You keep Vance firmly on vigilance, and you finally 00:12:23.399 --> 00:12:25.899 swap in the damaged bridge reference image. You 00:12:25.899 --> 00:12:29.159 write the next multi -shot prompt sequence. Ryder 00:12:29.159 --> 00:12:31.700 aggressively triggers the remote explosion. The 00:12:31.700 --> 00:12:34.379 bridge violently erupts in thick smoke and heavy 00:12:34.379 --> 00:12:36.240 debris. The scene carries forward incredibly 00:12:36.240 --> 00:12:39.360 smoothly. Then we get to clip three. This is 00:12:39.360 --> 00:12:41.399 easily the hardest action scene to generate. 00:12:41.789 --> 00:12:45.470 It is an absolute chaotic battle zone. Multiple 00:12:45.470 --> 00:12:49.009 characters are moving simultaneously. Heavy vehicles 00:12:49.009 --> 00:12:52.269 are reacting to the blast. Debris is falling 00:12:52.269 --> 00:12:55.210 everywhere across the frame. Because the visual 00:12:55.210 --> 00:12:57.950 data is so dense, the source material suggests 00:12:57.950 --> 00:13:01.350 a specific tactic. You bump the total generations 00:13:01.350 --> 00:13:05.490 up to three. I want to understand this. Why do 00:13:05.490 --> 00:13:07.600 we need more generations here? You fundamentally 00:13:07.600 --> 00:13:10.519 need better statistical odds. Dense action has 00:13:10.519 --> 00:13:13.600 so many overlapping moving parts. The AI doesn't 00:13:13.600 --> 00:13:15.820 truly understand physics. It just predicts pixels. 00:13:16.039 --> 00:13:18.299 Right. One take might miss the explosive timing 00:13:18.299 --> 00:13:20.559 completely. Another might accidentally block 00:13:20.559 --> 00:13:22.700 the main character's face. So for that messy 00:13:22.700 --> 00:13:24.899 action scene, we aren't expecting perfection, 00:13:25.139 --> 00:13:27.000 just asking for a few takes to choose the best 00:13:27.000 --> 00:13:29.460 one. Precisely. You watch all three generated 00:13:29.460 --> 00:13:31.879 variations and pick the clearest, most accurate 00:13:31.879 --> 00:13:34.159 take. Exactly. We're just playing the odds when 00:13:34.159 --> 00:13:35.860 the action gets heavy. Then you move on to clip 00:13:35.860 --> 00:13:38.679 four. This is the narrative ending. You need 00:13:38.679 --> 00:13:40.879 to deliberately slow the visual pacing down. 00:13:41.039 --> 00:13:43.399 You drop the generations back down to one. The 00:13:43.399 --> 00:13:45.840 immediate physical threat is almost entirely 00:13:45.840 --> 00:13:49.100 over. The emotional shift in this phase is very 00:13:49.100 --> 00:13:51.639 important. You change Vance's baseline emotion 00:13:51.639 --> 00:13:55.120 from vigilance to trust. If he visibly stays 00:13:55.120 --> 00:13:58.419 tense, the ending feels completely unresolved. 00:13:58.799 --> 00:14:01.799 The trust emotion physically tells the AI to 00:14:01.799 --> 00:14:04.559 relax the character's posture. Thick smoke slowly 00:14:04.559 --> 00:14:07.740 drifts across the ruined bridge. Ryder cautiously 00:14:07.740 --> 00:14:11.259 lowers his tactical weapon. Vance finally relaxes 00:14:11.259 --> 00:14:13.659 his grip on the high cliff. Ryder simply says, 00:14:13.700 --> 00:14:16.600 bridge is clear. The narrative sequence is completely 00:14:16.600 --> 00:14:18.799 resolved. Now you have four deeply connected 00:14:18.799 --> 00:14:22.019 clips. They logically follow a continuous physical 00:14:22.019 --> 00:14:24.659 path. They actually feel like a real directed 00:14:24.659 --> 00:14:27.259 sequence. Which naturally brings us to step four. 00:14:27.379 --> 00:14:30.120 You have to finally edit and export the final 00:14:30.120 --> 00:14:32.659 film. You assemble the final product in a program 00:14:32.659 --> 00:14:35.139 like CapCut. It is incredibly simple and fast 00:14:35.139 --> 00:14:37.419 for beginners to use. You import the generated 00:14:37.419 --> 00:14:39.879 clips in exact chronological order. Yeah. The 00:14:39.879 --> 00:14:43.470 quiet setup. The massive explosion. the chaotic 00:14:43.470 --> 00:14:46.730 action, the calm end. You watch the full timeline 00:14:46.730 --> 00:14:49.330 sequentially to feel the narrative rhythm, but 00:14:49.330 --> 00:14:51.789 you keep the actual edits incredibly simple. 00:14:51.870 --> 00:14:53.750 Wait, let me push back on this. So we are barely 00:14:53.750 --> 00:14:55.860 editing at all. That goes against everything 00:14:55.860 --> 00:14:58.059 I know about traditional video production. The 00:14:58.059 --> 00:15:00.299 edit bay is usually where the movie is actually 00:15:00.299 --> 00:15:02.139 made. I completely understand that reaction, 00:15:02.379 --> 00:15:05.279 but the video reference step already did the 00:15:05.279 --> 00:15:07.879 incredibly hard work. Yeah. That is the real 00:15:07.879 --> 00:15:10.299 paradigm shift here. Okay. The lighting and the 00:15:10.299 --> 00:15:13.159 atmospheric mood already match perfectly. The 00:15:13.159 --> 00:15:15.860 dialogue audio is already seamlessly baked in. 00:15:15.919 --> 00:15:17.659 Because we controlled everything in the reference 00:15:17.659 --> 00:15:20.379 and generation phases, the edit bay is just for 00:15:20.379 --> 00:15:24.549 polishing. Yes, editing simply becomes basic 00:15:24.549 --> 00:15:27.470 superficial cleanup. You are no longer desperately 00:15:27.470 --> 00:15:30.990 trying to save a broken, disjointed film. The 00:15:30.990 --> 00:15:33.710 real heavy lifting happens long before we ever 00:15:33.710 --> 00:15:36.509 open CapCut. You simply trim the slightly weak 00:15:36.509 --> 00:15:39.669 clip openings. You cut any awkwardly long pauses. 00:15:39.690 --> 00:15:42.110 You carefully remove any visually broken frame. 00:15:42.350 --> 00:15:44.629 You absolutely do not add heavy, distracting 00:15:44.629 --> 00:15:46.730 visual transitions. Then you hit export. You 00:15:46.730 --> 00:15:49.870 use a 16 .9 ratio for a classic cinematic look, 00:15:49.990 --> 00:15:53.409 or you use 9 .16 for modern vertical platforms. 00:15:53.570 --> 00:15:54.830 You gotta always watch it outside the editor 00:15:54.830 --> 00:15:57.230 first. on your phone, you will catch small continuity 00:15:57.230 --> 00:15:59.389 problems or pacing issues much easier that way. 00:15:59.509 --> 00:16:02.519 Two sec silence. So what does this all essentially 00:16:02.519 --> 00:16:04.799 mean? Let's recap the big idea here. The main 00:16:04.799 --> 00:16:07.600 takeaway is entirely about intention and control. 00:16:07.879 --> 00:16:11.399 Do not ever make the AI guess your movie. You 00:16:11.399 --> 00:16:13.600 show it exactly what your specific world looks 00:16:13.600 --> 00:16:16.360 like first. Only then do you tell it what should 00:16:16.360 --> 00:16:19.259 actually happen. The magic formula is quite clear 00:16:19.259 --> 00:16:22.019 and replicable. Solid references, structured 00:16:22.019 --> 00:16:25.320 video generation, mathematical continuity, simple 00:16:25.320 --> 00:16:28.000 edit. You meticulously build your characters 00:16:28.000 --> 00:16:31.200 and locations. You write a smart multi -shot 00:16:31.200 --> 00:16:33.879 prompt sequence. You mathematically connect clips 00:16:33.879 --> 00:16:36.539 with video references. You easily clean it up 00:16:36.539 --> 00:16:39.299 in CapCut. It completely changes how you approach 00:16:39.299 --> 00:16:41.980 these generative AI tools. It turns a folder 00:16:41.980 --> 00:16:45.419 of random chaotic clips into a real watchable 00:16:45.419 --> 00:16:48.240 film. It is a profound shift in mindset. You 00:16:48.240 --> 00:16:50.539 become a deliberate director, not just a passive 00:16:50.539 --> 00:16:53.200 prompter. Before we sign off, I want to leave 00:16:53.200 --> 00:16:55.120 you with a final thought to mull over. I always 00:16:55.120 --> 00:16:57.179 love these conceptual questions. Let's hear it. 00:16:57.299 --> 00:16:59.179 We just spent this entire time talking about 00:16:59.179 --> 00:17:01.840 how perfectly an AI can maintain a fictional 00:17:01.840 --> 00:17:04.339 world. It mathematically does it just from a 00:17:04.339 --> 00:17:07.099 few carefully constructed reference shots. But 00:17:07.099 --> 00:17:09.660 what happens when we inevitably start feeding 00:17:09.660 --> 00:17:13.480 it visual references of our actual lives? Pictures 00:17:13.480 --> 00:17:16.140 of our real homes, audio of our past memories. 00:17:16.279 --> 00:17:18.859 How long until the films we casually generate 00:17:18.859 --> 00:17:21.220 are completely indistinguishable from the lives 00:17:21.220 --> 00:17:23.740 we've actually lived? That is a fascinating and 00:17:23.740 --> 00:17:26.180 perhaps slightly terrifying question to consider. 00:17:26.579 --> 00:17:29.319 The line between generated memory and real memory 00:17:29.319 --> 00:17:32.279 is getting very thin. Thank you for diving deep 00:17:32.279 --> 00:17:34.779 with us today. Keep learning, keep experimenting 00:17:34.779 --> 00:17:36.660 with these tools, and we'll see you on the next 00:17:36.660 --> 00:17:38.259 deep dive. Out to your own music.