WEBVTT 00:00:00.000 --> 00:00:02.980 You've got this great idea for an AI video. You 00:00:02.980 --> 00:00:06.540 type it in. You hit generate. And, well, it's 00:00:06.540 --> 00:00:09.119 a mess. Yeah, a total mess. Your main character's 00:00:09.119 --> 00:00:11.980 face just keeps morphing. It turns into someone 00:00:11.980 --> 00:00:14.560 else entirely in the background. Right. The visual 00:00:14.560 --> 00:00:17.760 style feels incredibly chaotic. What was supposed 00:00:17.760 --> 00:00:20.140 to be cinematic looks like a complete and utter 00:00:20.140 --> 00:00:22.480 accident. We have all been there. It's incredibly 00:00:22.480 --> 00:00:24.320 frustrating. You kind of feel like you're playing 00:00:24.320 --> 00:00:26.539 a slot machine. You just pull the lever and hope. 00:00:27.890 --> 00:00:31.269 Welcome to the deep dive. Today we are breaking 00:00:31.269 --> 00:00:34.329 down a step -by -step guide. We're exploring 00:00:34.329 --> 00:00:38.090 exactly how to build AI video from scratch. We're 00:00:38.090 --> 00:00:40.810 looking at a specific three tool sequence. We're 00:00:40.810 --> 00:00:44.539 talking Cloud AI, ChatGPT, and Google Flow. That's 00:00:44.539 --> 00:00:47.079 the stack. This sequence guarantees consistent, 00:00:47.420 --> 00:00:49.799 highly cinematic results. It's really about moving 00:00:49.799 --> 00:00:52.939 from total randomness to actual directorial control. 00:00:53.219 --> 00:00:55.479 Right. It really is a mindset shift. We need 00:00:55.479 --> 00:00:57.820 to understand why those initial generations turn 00:00:57.820 --> 00:01:01.000 into chaotic messes. It's honestly not about 00:01:01.000 --> 00:01:04.560 writing a better, more complex prompt. It is 00:01:04.560 --> 00:01:06.700 entirely about the sequence of events. The sequence. 00:01:06.859 --> 00:01:09.140 Let's unpack that. The fundamental rule is this. 00:01:09.599 --> 00:01:11.840 You must build your visuals as static images 00:01:11.840 --> 00:01:14.969 first. You have to do before animating anything 00:01:14.969 --> 00:01:18.189 at all. Generating static images is much faster. 00:01:18.409 --> 00:01:20.689 It's also significantly cheaper than generating 00:01:20.689 --> 00:01:23.010 video. You know, I still wrestle with prompt 00:01:23.010 --> 00:01:25.629 drift myself. Oh, yeah. Watching my character 00:01:25.629 --> 00:01:28.469 morph into a totally different person halfway 00:01:28.469 --> 00:01:31.329 through a scene, it is a really helpless feeling. 00:01:31.530 --> 00:01:33.469 Absolutely. Everyone experiences that prompt 00:01:33.469 --> 00:01:35.590 drift initially. It's like stacking Lego blocks 00:01:35.590 --> 00:01:38.629 of data. You need a solid, unmoving base before 00:01:38.629 --> 00:01:40.569 you add the moving pieces. Exactly. If you build 00:01:40.569 --> 00:01:42.670 on a shaky foundation, the whole scene just collapses. 00:01:42.829 --> 00:01:46.469 Yeah. But why is that? Why is fixing a mistake 00:01:46.469 --> 00:01:49.409 in a video so much harder than in a static image? 00:01:50.030 --> 00:01:52.930 Because video multiplies variables across time. 00:01:53.189 --> 00:01:55.769 Static images isolate those variables. Think 00:01:55.769 --> 00:01:58.390 about what a video actually is. Right. A standard 00:01:58.390 --> 00:02:01.349 film runs at 24 frames per second. If you generate 00:02:01.349 --> 00:02:04.530 a 10 -second clip, the AI isn't just making one 00:02:04.530 --> 00:02:08.610 picture. It's making 240 pictures. Exactly. 240 00:02:08.610 --> 00:02:11.930 distinct images. And here is the truly difficult 00:02:11.930 --> 00:02:14.870 part for the AI. It has to remember what happened 00:02:14.870 --> 00:02:18.129 in frame one. It has to apply that exact memory 00:02:18.129 --> 00:02:21.150 to frame two. Then it has to guess what naturally 00:02:21.150 --> 00:02:24.349 happens in frame three. That is an immense computational 00:02:24.349 --> 00:02:27.250 burden. The system has to maintain the physics 00:02:27.250 --> 00:02:30.150 of a moving scene. It is massive. Eventually 00:02:30.150 --> 00:02:32.530 the machine drops the ball. The math just gets 00:02:32.530 --> 00:02:35.110 too heavy. It forgets the color of the character's 00:02:35.110 --> 00:02:37.310 jacket. It forgets the exact lighting angle on 00:02:37.310 --> 00:02:39.969 the coffee cup. Wow. But an image generator only 00:02:39.969 --> 00:02:43.110 has to solve one single frame. It concentrates 00:02:43.110 --> 00:02:45.409 all its processing power on that exact moment. 00:02:45.490 --> 00:02:47.669 You get a highly detailed, perfectly accurate 00:02:47.669 --> 00:02:50.650 picture. So video multiplies the chaos across 00:02:50.650 --> 00:02:54.129 time. Images isolate the problem entirely. Exactly. 00:02:54.169 --> 00:02:56.490 It pins the butterfly to the board, so to speak. 00:02:56.750 --> 00:02:59.289 You lock in the exact visual style. You fix the 00:02:59.289 --> 00:03:01.509 character's faces. Right. You perfect the lighting 00:03:01.509 --> 00:03:03.770 before a single frame actually moves. Which means 00:03:03.770 --> 00:03:06.330 we need a rock solid textual blueprint to start. 00:03:06.409 --> 00:03:08.389 Yeah. And that brings us to the first tool in 00:03:08.389 --> 00:03:11.150 this sequence. We're talking about Claude AI. 00:03:11.400 --> 00:03:14.240 Right. Before you even look at an image generator, 00:03:14.379 --> 00:03:16.979 before you open a video tool, you need a very 00:03:16.979 --> 00:03:20.400 clear, structured plan. Claude AI handles this 00:03:20.400 --> 00:03:23.080 entire blueprint stage. The guide we are looking 00:03:23.080 --> 00:03:25.360 at uses a very specific example. It starts with 00:03:25.360 --> 00:03:28.520 a remarkably simple story. A father and daughter 00:03:28.520 --> 00:03:32.020 escape a sudden volcano eruption. They're driving 00:03:32.020 --> 00:03:34.879 around desperately searching for gas to survive. 00:03:35.080 --> 00:03:37.219 It's a great example. It has inherent tension, 00:03:37.680 --> 00:03:40.319 but visually it's very contained. But let me 00:03:40.319 --> 00:03:42.939 push back on that a bit. Doesn't giving the AI 00:03:42.939 --> 00:03:46.580 a highly complex story make for a richer video? 00:03:47.080 --> 00:03:49.919 Like, if I want a sprawling sci -fi epic, shouldn't 00:03:49.919 --> 00:03:52.479 my premise be sprawling, too? It sounds counterintuitive, 00:03:52.520 --> 00:03:55.000 but no. A simple premise is actually much better 00:03:55.000 --> 00:03:58.219 here. Think about how the AI processes text tokens. 00:03:58.259 --> 00:04:01.280 OK. If you give the AI a complex novel with 10 00:04:01.280 --> 00:04:04.060 subplots, it won't know what visual elements 00:04:04.060 --> 00:04:06.219 to prioritize. Its attention gets scattered. 00:04:06.400 --> 00:04:08.560 It loses focus on the actual scene. Precisely. 00:04:08.680 --> 00:04:11.099 Two characters and one clear goal is perfect. 00:04:11.340 --> 00:04:14.199 It allows the AI to extract very clear visual 00:04:14.199 --> 00:04:16.660 directions. It knows exactly who to light and 00:04:16.660 --> 00:04:19.439 what they're doing. Keep the core idea incredibly 00:04:19.439 --> 00:04:23.459 simple so the AI can extract clear visual directions. 00:04:24.079 --> 00:04:27.240 Beat. That's the secret. So you take that simple 00:04:27.240 --> 00:04:30.720 story to Claude, then you create an AI video 00:04:30.720 --> 00:04:32.839 prompt skill. Let's define that. For someone 00:04:32.839 --> 00:04:35.540 new to this, what is an AI video prompt skill? 00:04:35.920 --> 00:04:39.389 Simply put, A saved instruction file for repeated 00:04:39.389 --> 00:04:41.670 AI tasks. So you aren't just typing into a blank 00:04:41.670 --> 00:04:43.689 chat box every time? Exactly. You are setting 00:04:43.689 --> 00:04:46.569 up a permanent framework. You tell Claude how 00:04:46.569 --> 00:04:49.230 to behave as a master video director. Nice. You 00:04:49.230 --> 00:04:51.269 load this skill automatically whenever you want 00:04:51.269 --> 00:04:53.389 to make a video. It saves you from constantly 00:04:53.389 --> 00:04:57.050 explaining the rules to the AI. What are we actually 00:04:57.050 --> 00:04:59.310 instructing Claude to do with this skill? What 00:04:59.310 --> 00:05:01.870 are the outputs we need? This skill demands three 00:05:01.870 --> 00:05:03.910 very distinct things from your simple story. 00:05:04.370 --> 00:05:06.870 It forces Claude to break the story down systematically. 00:05:07.670 --> 00:05:10.310 First, it generates a design sheet prompt. A 00:05:10.310 --> 00:05:12.149 design sheet? Kind of like concept art for a 00:05:12.149 --> 00:05:14.680 movie. Right. Exactly like concept art. This 00:05:14.680 --> 00:05:17.500 prompt outlines the characters, the core environment, 00:05:17.680 --> 00:05:20.399 and the color palette. It establishes the visual 00:05:20.399 --> 00:05:23.920 roles. Second, Claude generates a storyboard 00:05:23.920 --> 00:05:27.079 prompt. This breaks the story down into specific 00:05:27.079 --> 00:05:30.079 camera angles for each shot. Close -ups, wide 00:05:30.079 --> 00:05:32.540 shots, panning descriptions. Yes, it acts as 00:05:32.540 --> 00:05:35.779 the cinematographer. Third, it generates individual 00:05:35.779 --> 00:05:38.860 video prompts for every single scene. These are 00:05:38.860 --> 00:05:41.540 highly detailed instructions for the final animation 00:05:41.540 --> 00:05:44.600 phase. So Claude gives us this perfect structured 00:05:44.600 --> 00:05:47.560 text. We have the design sheet prompt, the storyboard 00:05:47.560 --> 00:05:50.279 prompt, and the video prompt. Yeah. But text 00:05:50.279 --> 00:05:52.300 isn't a movie. We need to actually visualize 00:05:52.300 --> 00:05:54.180 these blueprints. This is where we transition 00:05:54.180 --> 00:05:56.629 to the second stage. This is where ChatGPT comes 00:05:56.629 --> 00:05:59.089 in to lock in the visual style. We're moving 00:05:59.089 --> 00:06:02.089 from planning to actual image generation. You 00:06:02.089 --> 00:06:05.290 open ChatGPT. Specifically, you want to use the 00:06:05.290 --> 00:06:08.029 Deli 3 image generation feature inside it. Your 00:06:08.029 --> 00:06:10.430 first move is creating that master design sheet. 00:06:10.610 --> 00:06:12.629 We paste the design sheet prompt that Claude 00:06:12.629 --> 00:06:15.490 just wrote for us. You do. But the guide emphasizes 00:06:15.490 --> 00:06:18.290 a crucial addition here. You must manually add 00:06:18.290 --> 00:06:21.529 a specific phrase to the end. OK. You type. High 00:06:21.529 --> 00:06:24.569 detail, wide format, every element clearly visible. 00:06:24.990 --> 00:06:27.449 Why those exact words? It sounds like an arbitrary 00:06:27.449 --> 00:06:29.870 magic spell. Why not just say, make it look good? 00:06:30.110 --> 00:06:32.629 Because diffusion models can be lazy. If you 00:06:32.629 --> 00:06:36.269 just say, make it look good. The AI focuses on 00:06:36.269 --> 00:06:38.790 aesthetics. It might use heavy shadows to look 00:06:38.790 --> 00:06:41.629 cinematic. It might blur the background for an 00:06:41.629 --> 00:06:43.990 artistic depth of field. Which hides the actual 00:06:43.990 --> 00:06:46.850 details we need. Precisely. We are building a 00:06:46.850 --> 00:06:49.569 reference document, not a final piece of art. 00:06:50.050 --> 00:06:53.449 Those specific words force the AI to prioritize 00:06:53.449 --> 00:06:56.449 spatial clarity. Makes sense. We need flat, even 00:06:56.449 --> 00:06:58.730 lighting. We need to see the character's exact 00:06:58.730 --> 00:07:01.230 face. We need to clearly see the buckle on their 00:07:01.230 --> 00:07:03.790 backpack. This image is the anchor for everything 00:07:03.790 --> 00:07:06.000 else. So you get this design sheet back, you 00:07:06.000 --> 00:07:08.920 look it over, but what happens if chat GPT gets 00:07:08.920 --> 00:07:11.860 a small detail wrong? Say the daughter's backpack 00:07:11.860 --> 00:07:14.939 is the wrong color, or a photograph on the table 00:07:14.939 --> 00:07:18.139 is facing the wrong way. The instinct is to just 00:07:18.139 --> 00:07:20.160 rewrite the whole prompt and try again. That 00:07:20.160 --> 00:07:22.519 is the biggest mistake people make. Do not rewrite 00:07:22.519 --> 00:07:24.439 the whole prompt. Why not? Because rewriting 00:07:24.439 --> 00:07:27.360 changes the underlying seed noise, the AI will 00:07:27.360 --> 00:07:29.259 generate a completely different image from scratch. 00:07:29.500 --> 00:07:32.000 Oh, wow. The lighting will change. The faces 00:07:32.000 --> 00:07:34.120 will change. You lose all the good stuff just 00:07:34.120 --> 00:07:37.720 to fix one tiny detail. So how do you fix it 00:07:37.720 --> 00:07:41.519 without destroying the image? You add one specific 00:07:41.519 --> 00:07:44.480 localized correction sentence. If the photograph 00:07:44.480 --> 00:07:47.139 is wrong, you literally just type only the back 00:07:47.139 --> 00:07:49.279 of the photograph is visible. You just replied 00:07:49.279 --> 00:07:52.360 the image with that one sentence. Yes. The AI 00:07:52.360 --> 00:07:54.839 understands localized constraints much better. 00:07:55.060 --> 00:07:57.439 It keeps the original context window intact. 00:07:57.800 --> 00:08:00.759 It just surgically alters that one specific element 00:08:00.759 --> 00:08:03.500 you mentioned. Let me play devil's advocate here. 00:08:03.579 --> 00:08:06.279 Why not just skip this design sheet entirely? 00:08:06.420 --> 00:08:09.160 Okay. If we have the text prompts, why not generate 00:08:09.069 --> 00:08:11.459 the storyboard right away. Because without that 00:08:11.459 --> 00:08:14.560 visual anchor, the AI is just guessing. It will 00:08:14.560 --> 00:08:17.639 hallucinate a new visual style for every single 00:08:17.639 --> 00:08:19.899 panel. The dreaded prompt drift again. Exactly. 00:08:20.060 --> 00:08:21.740 In panel one, your character is wearing a clean 00:08:21.740 --> 00:08:24.720 blue jacket. In panel two, it's suddenly a dirty 00:08:24.720 --> 00:08:27.519 denim vest. Right. The lighting shifts from overcast 00:08:27.519 --> 00:08:30.399 to bright sunlight. It looks amateurish. The 00:08:30.399 --> 00:08:33.059 design sheet is the exact visual anchor keeping 00:08:33.059 --> 00:08:36.120 the AI from hallucinating. It absolutely is. 00:08:36.360 --> 00:08:38.259 So once that design sheet is perfectly locked 00:08:38.259 --> 00:08:40.220 in, once the character is looking Exactly right. 00:08:40.379 --> 00:08:43.460 Then, and only then, do you generate the storyboard. 00:08:43.600 --> 00:08:45.740 You take the storyboard prompt Claude wrote earlier, 00:08:46.159 --> 00:08:48.659 you paste it into ChatGPT. But here is the critical 00:08:48.659 --> 00:08:52.620 step. You must attach that perfect design sheet 00:08:52.620 --> 00:08:55.700 image to the prompt. You are literally feeding 00:08:55.700 --> 00:08:59.340 the image back into the AI. You tell ChatGPT 00:08:59.340 --> 00:09:02.159 to match those exact characters and colors. You're 00:09:02.159 --> 00:09:04.399 giving it visual reference, not just text. Yes. 00:09:04.480 --> 00:09:07.409 And you add one more specific phrase here. You 00:09:07.409 --> 00:09:10.830 type, high detail, each panel clearly separated 00:09:10.830 --> 00:09:13.169 and readable. Because a storyboard is a grid 00:09:13.169 --> 00:09:15.029 of multiple images. Right. It's usually a two 00:09:15.029 --> 00:09:18.830 by six grid, 12 panels total. If the AI blends 00:09:18.830 --> 00:09:20.950 the borders of those panels together, it becomes 00:09:20.950 --> 00:09:23.590 a mess. You need clean, distinct shots because 00:09:23.590 --> 00:09:25.529 you will be isolating them soon. We're going 00:09:25.529 --> 00:09:27.690 to take a quick pause right here. We don't go 00:09:27.690 --> 00:09:31.889 anywhere. And we are back. So far, we have planned 00:09:31.889 --> 00:09:34.830 our text in Claude. We have locked in our visual 00:09:34.830 --> 00:09:37.730 style with a design sheet in ChatGPT. And we 00:09:37.730 --> 00:09:41.230 just generated a perfectly consistent 12 -panel 00:09:41.230 --> 00:09:44.009 storyboard. Everything matches. You have a comic 00:09:44.009 --> 00:09:46.610 book version of your film. It is static, but 00:09:46.610 --> 00:09:49.169 it is visually perfect. Now comes the magic. 00:09:49.789 --> 00:09:52.970 We finally breathe life into it. We are moving 00:09:52.970 --> 00:09:56.879 to the third tool. Google Flow. This is the animation 00:09:56.879 --> 00:09:59.879 phase. Google Flow uses an underlying technology 00:09:59.879 --> 00:10:02.860 called VO3. Let's define VO3 for the listener. 00:10:03.200 --> 00:10:06.080 An AI engine that turns static images into moving 00:10:06.080 --> 00:10:08.159 clips. Simple enough, how do we actually use 00:10:08.159 --> 00:10:11.000 it? You take that full 12 -panel storyboard image. 00:10:11.259 --> 00:10:13.620 You upload that single image file directly into 00:10:13.620 --> 00:10:16.860 Google Flow. OK. Then you paste those 12 specific 00:10:16.860 --> 00:10:19.500 video prompts, the ones Claude wrote for us back 00:10:19.500 --> 00:10:21.980 in Stage 1. So we are giving Flow the static 00:10:21.980 --> 00:10:24.480 visual grid plus the text instructions of how 00:10:24.480 --> 00:10:26.759 things should move. Exactly. And you add a master 00:10:26.759 --> 00:10:28.940 instruction to the top of the prompt. You tell 00:10:28.940 --> 00:10:31.600 Flow, generate a scene using shots in the uploaded 00:10:31.600 --> 00:10:33.500 film storyboard sequence. The guide mentions 00:10:33.500 --> 00:10:36.000 adding one more strict constraint here. You have 00:10:36.000 --> 00:10:39.159 to explicitly tell the AI, no subtitles, and 00:10:39.159 --> 00:10:41.419 no music. Why do we need to specify that? We 00:10:41.419 --> 00:10:43.799 obviously want sound and text in our final video 00:10:43.799 --> 00:10:47.039 eventually, right? We do. But AI video generators 00:10:47.039 --> 00:10:49.940 have a really bad habit. If they try to generate 00:10:49.940 --> 00:10:52.580 audio or text alongside the video, it bakes it 00:10:52.580 --> 00:10:54.700 directly into the file. It's permanently attached 00:10:54.700 --> 00:10:57.480 to the visuals. Yes. AI -generated text often 00:10:57.480 --> 00:11:00.679 looks like alien gibberish. It hallucinates weird 00:11:00.679 --> 00:11:03.679 letters. AI -generated music might have tempo 00:11:03.679 --> 00:11:06.200 changes you hate. That makes sense. If that audio 00:11:06.200 --> 00:11:08.620 is baked into your raw video file, you cannot 00:11:08.620 --> 00:11:11.500 separate it later. You ruin your non -linear 00:11:11.500 --> 00:11:14.019 editing options. Blank audio and clean frames 00:11:14.019 --> 00:11:16.220 give you total control during the final edit. 00:11:16.419 --> 00:11:18.919 It beat. Exactly. Keep the raw footage as clean 00:11:18.919 --> 00:11:21.600 as possible so you hit generate and flow goes 00:11:21.600 --> 00:11:24.440 to work. It processes the grid. It does. And 00:11:24.440 --> 00:11:28.929 whoo! Imagine turning a flat 12 -panel grid into 00:11:28.929 --> 00:11:31.990 15 seconds of cinematic motion instantly. Yeah. 00:11:32.309 --> 00:11:34.610 It analyzes the spatial relationships in the 00:11:34.610 --> 00:11:37.250 static image. It calculates the temporal consistency 00:11:37.250 --> 00:11:39.809 required to make it move. It just blows my mind 00:11:39.809 --> 00:11:42.470 every time I see it work. It's pretty wild to 00:11:42.470 --> 00:11:46.110 think about. You are taking a flat static comic 00:11:46.110 --> 00:11:48.950 strip and getting a living, breathing movie out 00:11:48.950 --> 00:11:51.190 of it. Absolutely. It feels like something straight 00:11:51.190 --> 00:11:54.419 out of a sci -fi novel. But wait. We have 12 00:11:54.419 --> 00:11:58.019 distinct panels, and flow generates a 15 -second 00:11:58.019 --> 00:12:01.059 clip. That math is pretty tight. It is very tight, 00:12:01.240 --> 00:12:03.940 averaging just over one second per shot. Some 00:12:03.940 --> 00:12:06.120 of those shots must flash by incredibly quickly. 00:12:06.240 --> 00:12:08.419 They absolutely do. Some shots will feel rushed. 00:12:08.700 --> 00:12:10.519 Occasionally, the AI might even skip a panel 00:12:10.519 --> 00:12:13.019 entirely if the transition is too complex. Oh, 00:12:13.019 --> 00:12:15.759 really? Yeah, but that is OK. This initial generation 00:12:15.759 --> 00:12:18.419 is basically an animatic. It is a rough cut. 00:12:18.500 --> 00:12:21.639 It shows you the overall pacing and flow. So 00:12:21.639 --> 00:12:23.779 you watch this rough 15 second clip, what are 00:12:23.779 --> 00:12:26.419 we looking for? You are checking for temporal 00:12:26.419 --> 00:12:29.539 consistency. Do the warm orange tones of the 00:12:29.539 --> 00:12:32.179 volcano ash stay consistent into the final gas 00:12:32.179 --> 00:12:34.820 station scene? Does the camera pan smoothly? 00:12:35.840 --> 00:12:39.039 If it feels too fast, or if a transition is jarring, 00:12:39.600 --> 00:12:41.620 how do we fix it? We aren't in a traditional 00:12:41.620 --> 00:12:44.080 editing timeline here. Flow has a feature called 00:12:44.080 --> 00:12:46.840 the describe your edits box. It uses natural 00:12:46.840 --> 00:12:49.230 language processing to adjust the timeline. You 00:12:49.230 --> 00:12:52.110 literally just type smooth the transition between 00:12:52.110 --> 00:12:55.950 each shot, or you type maintain consistent pacing 00:12:55.950 --> 00:12:58.309 throughout the video. You just ask it nicely 00:12:58.309 --> 00:13:01.669 to fix the edit. Exactly. The engine recalculates 00:13:01.669 --> 00:13:04.590 the latent space between the frames. It generates 00:13:04.590 --> 00:13:07.289 new transitional frames to smooth out the motion. 00:13:07.389 --> 00:13:10.669 Wow. It adjusts the pacing without changing the 00:13:10.669 --> 00:13:13.490 core visual assets you establish. That is incredible 00:13:13.490 --> 00:13:16.559 control. You are directing the edit with text. 00:13:16.720 --> 00:13:18.580 And because you did the hard work in stage one 00:13:18.580 --> 00:13:21.039 and two. Because you built that perfect design 00:13:21.039 --> 00:13:23.679 sheet. Right. The final motion clip maintains 00:13:23.679 --> 00:13:25.980 its integrity. The characters look right. The 00:13:25.980 --> 00:13:28.759 lighting matches. It feels intentional. The Claude 00:13:28.759 --> 00:13:31.840 to chat GPT to flow pipeline is undeniably powerful. 00:13:32.600 --> 00:13:35.059 But let's look at the bigger picture. Not everyone 00:13:35.059 --> 00:13:37.179 uses those specific tools. Maybe they don't have 00:13:37.179 --> 00:13:39.600 access. Maybe the subscription costs for three 00:13:39.600 --> 00:13:41.960 different premium AI services are just too high. 00:13:42.080 --> 00:13:45.100 That is a very valid concern. Generating images 00:13:45.100 --> 00:13:48.419 and video at scale gets expensive quickly. The 00:13:48.419 --> 00:13:51.200 beauty of this specific framework is its modularity. 00:13:51.840 --> 00:13:55.259 The tools are completely swappable. Let's dig 00:13:55.259 --> 00:13:58.139 into the tool built. What are the alternatives 00:13:58.139 --> 00:14:01.399 for stage one, the prompt writing phase? If you 00:14:01.399 --> 00:14:04.529 don't want to use Claude, Gemini, is a fantastic 00:14:04.529 --> 00:14:08.070 alternative. Google's Gemini has a very capable 00:14:08.070 --> 00:14:11.730 free tier. It handles large context windows beautifully, 00:14:11.809 --> 00:14:13.850 which is great for building those skills. What 00:14:13.850 --> 00:14:16.110 about Grok? Grok is another great option for 00:14:16.110 --> 00:14:18.889 the text phase. It is incredibly fast. It also 00:14:18.889 --> 00:14:21.549 has fewer safety restrictions, which can be helpful 00:14:21.549 --> 00:14:23.850 if your story involves action or conflict that 00:14:23.850 --> 00:14:26.710 other AIs might mistakenly flag as inappropriate. 00:14:27.309 --> 00:14:30.049 What about stage two, generating the static images? 00:14:30.549 --> 00:14:33.509 We need tools that excel at building consistent 00:14:33.509 --> 00:14:36.769 design sheets. Ideogram is a phenomenal alternative 00:14:36.769 --> 00:14:40.129 to chat GPT and Dale E3. Why ideogram specifically? 00:14:40.330 --> 00:14:42.570 Two reasons. First, it is currently the best 00:14:42.570 --> 00:14:45.009 in the market at rendering visible text accurately. 00:14:45.210 --> 00:14:47.309 Oh, nice. If your scene requires a neon sign 00:14:47.309 --> 00:14:50.370 or a newspaper headline, ideogram nails it. Second, 00:14:50.429 --> 00:14:52.590 it has unique built -in tools for maintaining 00:14:52.590 --> 00:14:54.629 strict character consistency across different 00:14:54.629 --> 00:14:57.149 prompts. Any other image alternatives? Leonardo 00:14:57.149 --> 00:14:59.950 AI. It has an amazing free tier. It gives you 00:14:59.950 --> 00:15:02.200 an incredible amount of granular control. over 00:15:02.200 --> 00:15:05.519 the artistic style. If you want a highly stylized, 00:15:05.620 --> 00:15:08.320 painted look rather than photorealism, Leonardo 00:15:08.320 --> 00:15:11.080 is brilliant. Now for the heavy hitter. Stage 00:15:11.080 --> 00:15:15.039 three, the video generation. This is almost always 00:15:15.039 --> 00:15:18.100 the most computationally expensive step. What 00:15:18.100 --> 00:15:21.659 can we use instead of Google Flow and VO3? Kling 00:15:21.659 --> 00:15:24.559 3 .0 is a massive contender right now. I have 00:15:24.559 --> 00:15:26.740 seen a lot of clips from Kling online. This looks 00:15:26.740 --> 00:15:30.379 very cinematic. It is. Kling 3 .0 handles physics 00:15:30.379 --> 00:15:33.600 simulations remarkably well. Water splashing, 00:15:34.000 --> 00:15:36.559 smoke billowing. It's often cheaper than enterprise 00:15:36.559 --> 00:15:39.559 solutions, and the quality is stunning. What 00:15:39.559 --> 00:15:42.190 else is out there for video? Pix4C1 is another 00:15:42.190 --> 00:15:44.350 strong option. It is specifically optimized for 00:15:44.350 --> 00:15:46.250 taking storyboard reference panels and translating 00:15:46.250 --> 00:15:48.590 them into consistent motion. Good to know. And 00:15:48.590 --> 00:15:52.190 if you are on a strict $0 budget, Hylua AI is 00:15:52.190 --> 00:15:54.330 perfect for completely free experimentation. 00:15:54.750 --> 00:15:56.529 So you have all these clips, but we still need 00:15:56.529 --> 00:15:58.470 to put it all together. We need to add the music 00:15:58.470 --> 00:16:01.529 and subtitles we avoided earlier. For final editing, 00:16:02.009 --> 00:16:04.649 CapCut is the dominant choice for creators. Right. 00:16:04.779 --> 00:16:07.440 It's free, incredibly intuitive, and there is 00:16:07.440 --> 00:16:09.539 no watermark if you use the desktop version. 00:16:10.220 --> 00:16:12.419 If you want Hollywood level control, DaVinci 00:16:12.419 --> 00:16:15.279 Resolve offers professional color grading. It 00:16:15.279 --> 00:16:18.240 is a massive complex program, but the base version 00:16:18.240 --> 00:16:21.159 is entirely free. And obviously Adobe Premiere 00:16:21.159 --> 00:16:23.620 if you are already in that ecosystem. But let 00:16:23.620 --> 00:16:26.360 me ask you this. If we swap out the software, 00:16:26.919 --> 00:16:29.200 if we use Gemini instead of Claude and Kling 00:16:29.200 --> 00:16:32.340 instead of Flow, does the quality of this specific 00:16:32.340 --> 00:16:35.169 sequence degrade? Does the magic disappear? Not 00:16:35.169 --> 00:16:37.250 at all. The discipline of the sequence is what 00:16:37.250 --> 00:16:39.509 creates the quality. It is absolutely not about 00:16:39.509 --> 00:16:42.149 the specific brand of AI. It's the architecture 00:16:42.149 --> 00:16:45.129 of the workflow. Precisely. The AI models will 00:16:45.129 --> 00:16:47.149 change. A new tool will launch next week that 00:16:47.149 --> 00:16:50.309 makes VO3 look old. But the logic of the pipeline 00:16:50.309 --> 00:16:53.110 remains. Right. You must build the textual blueprint 00:16:53.110 --> 00:16:55.870 first. You must lock in the static image second. 00:16:56.289 --> 00:16:58.870 You only animate as the absolute final step. 00:16:59.110 --> 00:17:01.450 The sequence is universal, even if the exact 00:17:01.450 --> 00:17:04.089 software you decide to use changes. Two -sex 00:17:04.089 --> 00:17:06.490 silence. That is the big takeaway here. Building 00:17:06.490 --> 00:17:10.069 an AI video from scratch feels like magic. When 00:17:10.069 --> 00:17:12.869 you watch a finished high -quality AI film, it 00:17:12.869 --> 00:17:15.369 looks like a miracle of prompting, but it isn't 00:17:15.369 --> 00:17:17.849 magical prompting at all. No, it's just a structured 00:17:17.849 --> 00:17:20.990 process. It is about having a clear, methodical 00:17:20.990 --> 00:17:23.990 sequence. You start with a simple story. You 00:17:23.990 --> 00:17:27.869 let the text tool structure your prompts. You 00:17:27.869 --> 00:17:31.769 build your static visuals to anchor the AI. Only 00:17:31.769 --> 00:17:34.509 then do you hit the animate button. Each tool 00:17:34.509 --> 00:17:37.930 handles one specific job. You isolate the variables. 00:17:38.289 --> 00:17:40.950 You fix problems while they are still flat images. 00:17:41.170 --> 00:17:43.349 Because fixing a moving target is nearly impossible. 00:17:43.650 --> 00:17:45.829 Exactly. You become a director, carefully setting 00:17:45.829 --> 00:17:48.390 up the shot. You stop being a gambler just pulling 00:17:48.390 --> 00:17:50.450 the slot machine lever. It changes everything. 00:17:50.509 --> 00:17:52.710 Yeah. I want to leave everyone with a final thought 00:17:52.710 --> 00:17:54.910 to chew on today. We have been talking about 00:17:54.910 --> 00:17:58.400 this specific AI video workflow. But think about 00:17:58.400 --> 00:18:00.880 your own workflows outside of video creation. 00:18:00.980 --> 00:18:02.779 Oh, it applies to almost everything. It really 00:18:02.779 --> 00:18:05.079 does. Think about analytical business projects 00:18:05.079 --> 00:18:07.819 or writing code or even designing a presentation. 00:18:08.539 --> 00:18:10.880 This methodology building the design sheet first 00:18:10.880 --> 00:18:14.039 is universal. How often do we rush to the final 00:18:14.039 --> 00:18:16.940 moving product? We jump straight into the execution 00:18:16.940 --> 00:18:20.569 phase. blindly because it feels productive, we 00:18:20.569 --> 00:18:23.250 try to animate our ideas before we have actually 00:18:23.250 --> 00:18:25.130 designed them. We skip the blueprint because 00:18:25.130 --> 00:18:27.049 we want to see the house. Exactly. And then we 00:18:27.049 --> 00:18:29.849 spend twice as long fixing structural errors 00:18:29.849 --> 00:18:32.490 that never should have happened. Right. Next 00:18:32.490 --> 00:18:35.509 time you start a complex new project, stop and 00:18:35.509 --> 00:18:38.230 ask yourself, did I build my static design sheet 00:18:38.230 --> 00:18:41.309 first or am I just hoping the sequence works 00:18:41.309 --> 00:18:44.150 itself out? Thank you for joining The Deep Dive. 00:18:44.309 --> 00:18:45.329 We will see you next time.