WEBVTT 00:00:00.000 --> 00:00:02.060 Welcome to the deep dive. We are really glad 00:00:02.060 --> 00:00:03.720 you're joining us today Yeah, thanks for having 00:00:03.720 --> 00:00:05.660 me excited to get into this one because look 00:00:05.660 --> 00:00:08.140 I know the mindset you are bringing to this conversation 00:00:08.140 --> 00:00:10.820 You're a learner. You want to actually grasp 00:00:10.820 --> 00:00:14.199 how these complex systems work Without drowning 00:00:14.199 --> 00:00:16.420 in the overwhelm, right? Nobody wants to read 00:00:16.420 --> 00:00:18.719 a thousand page technical manual just to make 00:00:18.719 --> 00:00:22.769 a video exactly and That is the mission today. 00:00:23.309 --> 00:00:26.469 We are diving into an incredibly insightful source 00:00:26.469 --> 00:00:29.350 text called The Director's Roadmap to Professional 00:00:29.350 --> 00:00:32.189 AI Filmmaking. It's a great piece. It really 00:00:32.189 --> 00:00:35.850 is. We are going to shortcut that massive frustrating 00:00:35.850 --> 00:00:39.469 learning curve of AI video generation. The goal 00:00:39.469 --> 00:00:41.530 is to take you from being a beginner who just 00:00:41.530 --> 00:00:44.609 types a prompt and prays for luck to a true digital 00:00:44.609 --> 00:00:46.789 director running a professional pipeline. And 00:00:46.789 --> 00:00:48.429 getting to that professional level, it really 00:00:48.429 --> 00:00:52.960 requires dismantling a huge A trap almost everyone 00:00:52.960 --> 00:00:55.259 falls into right away. The empathy trap. Yeah, 00:00:55.340 --> 00:00:57.460 exactly. People just intuitively assume that 00:00:57.460 --> 00:01:00.320 AI understands human feelings, like a human collaborator 00:01:00.320 --> 00:01:03.179 would. Right. So you type in this deeply passionate 00:01:03.179 --> 00:01:05.799 emotional paragraph about the exact vibe you 00:01:05.799 --> 00:01:09.150 want, and you hit Enter. And the AI spits out 00:01:09.150 --> 00:01:12.930 a character with a terrifying twisted face or 00:01:12.930 --> 00:01:15.510 some background that defies the laws of physics 00:01:15.510 --> 00:01:17.510 entirely. Because it doesn't care about your 00:01:17.510 --> 00:01:19.730 passion. It's an algorithm. It follows strict 00:01:19.730 --> 00:01:22.709 mathematical logic. But when users don't grasp 00:01:22.709 --> 00:01:25.709 that logic, they get stuck. They operate them 00:01:25.709 --> 00:01:27.989 to this false assumption that if they just add 00:01:27.989 --> 00:01:30.230 more adjective, the computer will finally get 00:01:30.230 --> 00:01:32.409 it. But it doesn't. You might get one lucky clip, 00:01:32.409 --> 00:01:34.930 but the next scene is a total mess. The narrative 00:01:34.930 --> 00:01:37.250 continuity completely collapses. Yeah, you just 00:01:37.250 --> 00:01:39.209 pull in the slot machine lever over and over. 00:01:39.310 --> 00:01:41.829 OK, let's unpack this. Because the roadmap lays 00:01:41.829 --> 00:01:45.189 out the solution across five very clear levels. 00:01:45.650 --> 00:01:47.930 And level one is what the source calls the idea 00:01:47.930 --> 00:01:50.810 prompt. Which is where most people start, and 00:01:50.810 --> 00:01:52.650 unfortunately, where most people stay. Yeah. 00:01:52.950 --> 00:01:54.879 Operating at level one is basically treating 00:01:54.879 --> 00:01:58.140 the AI like a chaotic freelance artist. Sometimes 00:01:58.140 --> 00:02:00.719 it's stunning, usually it's useless for a real 00:02:00.719 --> 00:02:03.340 movie. And the biggest beginner mistake here 00:02:03.340 --> 00:02:06.180 is the length myth. Oh, the giant paragraphs. 00:02:06.659 --> 00:02:09.120 Right. People think detail equals precision. 00:02:10.340 --> 00:02:12.479 But modern tools, they don't need a novel. They 00:02:12.479 --> 00:02:15.259 need clarity. The source gives a perfect example 00:02:15.259 --> 00:02:17.800 of this using Runway. The prompt is just so simple. 00:02:17.979 --> 00:02:20.439 It's literally a fat cat wearing a space suit. 00:02:20.669 --> 00:02:22.909 is sitting and fishing on a planet with a pink 00:02:22.909 --> 00:02:25.810 sky surrounded by floating rocks and shining 00:02:25.810 --> 00:02:29.169 stars in a 3D Pixar movie style. What's fascinating 00:02:29.169 --> 00:02:31.449 here is that the video quality doesn't come from 00:02:31.449 --> 00:02:34.449 those specific words. The visual fidelity comes 00:02:34.449 --> 00:02:37.409 from the AI model itself. The words are just 00:02:37.409 --> 00:02:39.310 guardrails. Just keeping it from hallucinating. 00:02:39.469 --> 00:02:41.610 Exactly. Let's break down why that prompt works. 00:02:42.409 --> 00:02:45.469 First, you have the subject, fat cat in a spacesuit, 00:02:45.810 --> 00:02:48.169 that is a massively recognizable archetype in 00:02:48.169 --> 00:02:50.580 the training data. So the AI grabs that easily, 00:02:50.620 --> 00:02:52.960 but what about the movement? That's the genius 00:02:52.960 --> 00:02:55.460 of the action they chose. Sitting and fishing. 00:02:55.599 --> 00:02:58.060 It's calm. It's highly contained. If you ask 00:02:58.060 --> 00:03:00.759 for backflips and laser fire, the AI's predictive 00:03:00.759 --> 00:03:03.060 processing glitches out. You get blurred pixels. 00:03:03.419 --> 00:03:05.659 By keeping the movement calm, the render stays 00:03:05.659 --> 00:03:08.680 pristine. And then adding 3D Pixar style and 00:03:08.680 --> 00:03:10.900 pink sky that maps the lighting immediately without 00:03:10.900 --> 00:03:13.300 needing a technical essay on light sources. Right. 00:03:13.360 --> 00:03:16.039 It's incredibly efficient. The roadmap also brings 00:03:16.039 --> 00:03:19.830 up Luma AI for level one. The prompt there was 00:03:19.830 --> 00:03:22.650 a close -up shot of a robot hand carefully holding 00:03:22.650 --> 00:03:25.729 a rose made of glass. The sun reflects through 00:03:25.729 --> 00:03:27.870 the petals and creates beautiful light on the 00:03:27.870 --> 00:03:30.780 dusty ground. Now that is a prompt designed to 00:03:30.780 --> 00:03:33.060 exploit what the AI does best. It completely 00:03:33.060 --> 00:03:35.240 ignores the boring buzzwords people usually use. 00:03:35.360 --> 00:03:38.419 No UltraShark. No 8K resolution. It focuses on 00:03:38.419 --> 00:03:41.039 materials. Yes. The two hardest things in AI 00:03:41.039 --> 00:03:44.240 filmmaking are materials and lighting. Glass. 00:03:44.479 --> 00:03:48.039 Metal. Dust. It forces the AI to dedicate its 00:03:48.039 --> 00:03:51.219 power to simulating real -world light refraction. 00:03:51.500 --> 00:03:54.419 But we do have to call out. the major limitation 00:03:54.419 --> 00:03:56.659 of level one. The lack of control. Total lack 00:03:56.659 --> 00:03:59.180 of control. If you hit generate on that fat cat 00:03:59.180 --> 00:04:01.979 prompt 10 times, you get 10 totally different 00:04:01.979 --> 00:04:04.620 cats. Which is fine for a single TikTok video, 00:04:04.840 --> 00:04:07.360 but useless for a consistent long -form movie 00:04:07.360 --> 00:04:09.139 where we need to follow one character. Which 00:04:09.139 --> 00:04:11.340 brings us to level two, structured prompting. 00:04:11.460 --> 00:04:13.400 This is where you stop talking to the AI like 00:04:13.400 --> 00:04:15.000 it's a search engine. You start speaking the 00:04:15.000 --> 00:04:17.040 computer's native language. Series filmmakers 00:04:17.040 --> 00:04:19.980 use a template, the cinematic formula. Subject, 00:04:20.100 --> 00:04:23.370 action, scene. Camera, style. Exactly in that 00:04:23.370 --> 00:04:26.329 order. The source uses Killing AI to demonstrate 00:04:26.329 --> 00:04:29.389 this. Subject, an old detective wearing a long 00:04:29.389 --> 00:04:32.670 coat. Action standing and smoking. Environment, 00:04:32.970 --> 00:04:35.990 1950s London in the rain. Camera medium shot 00:04:35.990 --> 00:04:38.569 moving to face. Style, black and white noir. 00:04:38.839 --> 00:04:41.240 And the beauty of this formula is troubleshooting. 00:04:41.420 --> 00:04:43.660 Yes. If the clip fails, you don't throw away 00:04:43.660 --> 00:04:45.860 the whole idea. You just isolate the variable. 00:04:46.259 --> 00:04:49.139 Oh, the lighting is weird. Fix the style variable. 00:04:49.259 --> 00:04:51.319 Keep the camera and subject exactly the same. 00:04:51.839 --> 00:04:54.079 And to take that structural control even further, 00:04:54.279 --> 00:04:56.720 the author recommends using JSON formatting. 00:04:56.839 --> 00:04:59.899 Using code, basically. Well, using Chad GPT to 00:04:59.899 --> 00:05:02.779 format your ideas into JSON. It gives your project 00:05:02.779 --> 00:05:05.439 a consistent DNA. Right, the example was the 00:05:05.439 --> 00:05:07.720 old warrior with broken armor kneeling in sunflowers. 00:05:07.779 --> 00:05:10.540 By formatting it as JSON, the AI parses the exact 00:05:10.540 --> 00:05:13.920 same structural data every single time. And standardizing 00:05:13.920 --> 00:05:16.319 that data unlocks the multi -shot technique. 00:05:16.680 --> 00:05:19.060 This is huge. I love this part. You don't have 00:05:19.060 --> 00:05:21.000 to stitch together five -second clips anymore. 00:05:21.439 --> 00:05:24.600 No, you can generate a whole sequence. Like the 00:05:24.600 --> 00:05:28.029 example, a girl... opens a door, turns on a light 00:05:28.029 --> 00:05:30.550 to see a gift, and then a close -up of happy 00:05:30.550 --> 00:05:33.329 tears. All in one go. The model holds the whole 00:05:33.329 --> 00:05:36.110 sequence in its context window. It bakes the 00:05:36.110 --> 00:05:38.550 temporal consistency in from the start. Here's 00:05:38.550 --> 00:05:40.569 where it gets really interesting, because text 00:05:40.569 --> 00:05:44.129 prompts still hit a ceiling. Even with JSON, 00:05:44.569 --> 00:05:47.649 forcing the AI to keep a character's face perfectly 00:05:47.649 --> 00:05:50.149 identical across different scenes using only 00:05:50.149 --> 00:05:53.040 text. It's a nightmare. It is. Which is why Level 00:05:53.040 --> 00:05:55.759 3 is reference control. We stop relying on words. 00:05:55.959 --> 00:05:58.819 We force the AI's hand, using images and videos 00:05:58.819 --> 00:06:01.420 as maps. You basically become a casting director. 00:06:01.879 --> 00:06:04.360 Let's look at image video using Pika. You don't 00:06:04.360 --> 00:06:06.060 roll the dice on a text prompt for the face. 00:06:06.420 --> 00:06:08.579 You use mid -journey to generate your exact actor. 00:06:08.839 --> 00:06:10.819 You lock the look. Lock the look. Bring that 00:06:10.819 --> 00:06:13.300 static image into Pika and tell the AI keep the 00:06:13.300 --> 00:06:15.100 clothes, keep the hair, just make them smile 00:06:15.100 --> 00:06:17.279 and nod. But what if you need complex movement, 00:06:17.379 --> 00:06:19.350 like... Trying to type out the exact physics 00:06:19.350 --> 00:06:21.490 of casting a fishing rod in English is impossible. 00:06:21.870 --> 00:06:24.490 The AI will totally mess up the physics of the 00:06:24.490 --> 00:06:26.790 human arm. That's why video -to -video in Runway 00:06:26.790 --> 00:06:29.529 is brilliant. You literally act it out. Just 00:06:29.529 --> 00:06:31.110 record yourself on your phone in your living 00:06:31.110 --> 00:06:33.970 room. Mind the fishing motion, upload it, and 00:06:33.970 --> 00:06:36.709 Runway maps your natural human kinematics right 00:06:36.709 --> 00:06:40.069 onto the space cat. No complex English required. 00:06:40.449 --> 00:06:42.910 And you can steal million -dollar Hollywood camera 00:06:42.910 --> 00:06:45.850 moves the same way, right? Camera sync. The source 00:06:45.850 --> 00:06:48.829 talks about sedans for this. They applied a professional 00:06:48.829 --> 00:06:51.350 tracking shot to a white -haired anime girl. 00:06:51.629 --> 00:06:54.949 And the big update here is Sedence 2 .0, which 00:06:54.949 --> 00:06:58.649 dropped late February 2026. It handles multiple 00:06:58.649 --> 00:07:01.410 references at once seamlessly. So it locks the 00:07:01.410 --> 00:07:03.269 face from one image and grabs the camera move 00:07:03.269 --> 00:07:05.730 from a video all at the same time. Exactly. But 00:07:05.730 --> 00:07:07.990 man, doing that manually for a hundred shots 00:07:07.990 --> 00:07:09.810 would cause instant brain -out. Which is why 00:07:09.810 --> 00:07:13.769 Level 4 is all about custom GPTs. custom assistants. 00:07:14.209 --> 00:07:16.370 You turn Chat GPT into your script assistant, 00:07:16.730 --> 00:07:18.689 you give it your basic idea, and it hands you 00:07:18.689 --> 00:07:21.430 back three prompt choices optimized for caling 00:07:21.430 --> 00:07:24.910 AI. Wideshot, close -up, tracking shot all with 00:07:24.910 --> 00:07:27.430 professional lighting cues built in. You step 00:07:27.430 --> 00:07:31.250 out of the weeds, you become the boss, just choosing 00:07:31.250 --> 00:07:33.389 the best option. Which gives you the energy for 00:07:33.389 --> 00:07:36.370 level five. The full pipeline. If we connect 00:07:36.370 --> 00:07:38.670 this to the bigger picture, a professional movie 00:07:38.670 --> 00:07:41.670 isn't just one cool clip. It's a massive start 00:07:41.670 --> 00:07:43.910 -to -finish process. Starting with storyboarding. 00:07:44.470 --> 00:07:46.790 Before you spend a single credit, use Mid Journey 00:07:46.790 --> 00:07:50.389 to make a 3x3 grid. A 9 -frame comic book page. 00:07:50.569 --> 00:07:52.930 Like the spaceship crashing on a strange planet. 00:07:53.350 --> 00:07:55.670 It ensures your story flows and saves a ton of 00:07:55.670 --> 00:07:59.209 money. Then you need voices. Silent AI characters 00:07:59.209 --> 00:08:02.860 have that uncanny stiffness. The source uses 00:08:02.860 --> 00:08:06.279 11 labs, but with a highly specific trick. The 00:08:06.279 --> 00:08:08.120 square brackets. I thought this was brilliant. 00:08:08.279 --> 00:08:10.579 It changes everything. You don't just type the 00:08:10.579 --> 00:08:13.180 dialogue. You prompt it like bracket, exhausted, 00:08:13.540 --> 00:08:15.420 whispering, end bracket. I checked everything. 00:08:15.540 --> 00:08:18.019 It gives the character soul. Natural breathing. 00:08:18.199 --> 00:08:20.160 And finally, you have to sync that voice to the 00:08:20.160 --> 00:08:23.160 face using Creatify Aurora. But the source has 00:08:23.160 --> 00:08:25.139 a crucial warning here. Only do it when they 00:08:25.139 --> 00:08:28.180 are standing still. Right. Lipsync models hate 00:08:28.180 --> 00:08:30.649 fast movement. If your character is running or 00:08:30.649 --> 00:08:33.210 jumping, the lip -sync will turn into a glitchy, 00:08:33.230 --> 00:08:36.049 distorted mess. Keep the dialogue to the slow, 00:08:36.389 --> 00:08:38.769 stationary close -ups. Always. So evaluating 00:08:38.769 --> 00:08:41.789 this massive 2026 arsenal we just went through. 00:08:42.049 --> 00:08:44.309 This raises an important question. Where do you 00:08:44.309 --> 00:08:46.610 actually spend your subscription money? Let's 00:08:46.610 --> 00:08:49.090 do a rapid -fire review of the tools based on 00:08:49.090 --> 00:08:53.370 the roadmap, Kling AI. Amazing physics and long 00:08:53.370 --> 00:08:56.909 videos, but the faces can get weird. It's best 00:08:56.909 --> 00:08:59.700 for action sequences. runway best motion control 00:08:59.700 --> 00:09:02.500 by far but very expensive it's really a vfx tool 00:09:02.500 --> 00:09:05.360 lumadream machine unbelievable lighting and glass 00:09:05.360 --> 00:09:09.269 textures but glitchy movement Use it for artistic 00:09:09.269 --> 00:09:12.269 slow shots, mpica, fun physics, highly reliable 00:09:12.269 --> 00:09:15.230 but less cinematic, great for viral clips, animation, 00:09:15.549 --> 00:09:18.250 and sedents, incredible facial consistency, steep 00:09:18.250 --> 00:09:20.549 learning curve, but it's the king of anime and 00:09:20.549 --> 00:09:22.330 character films right now. But tools are just 00:09:22.330 --> 00:09:25.029 tools. The roadmap ends with three golden rules. 00:09:25.509 --> 00:09:27.990 Rule one, don't let the AI do everything. They 00:09:27.990 --> 00:09:30.409 have no idea what pacing is. Use it for raw materials, 00:09:30.470 --> 00:09:32.210 but you have to use your human brain to edit 00:09:32.210 --> 00:09:35.190 it together. Rule two, avoid the AI smell. That 00:09:35.190 --> 00:09:38.500 plastic, perfectly smooth look, It screams fake. 00:09:38.840 --> 00:09:41.860 So you add dust, film grain, handheld camera 00:09:41.860 --> 00:09:43.779 shake to your prompts. You have to dirty it up 00:09:43.779 --> 00:09:46.159 to make it real. And rule three check copyrights. 00:09:46.519 --> 00:09:49.539 The 2026 laws are no joke. Yeah. Make sure you 00:09:49.539 --> 00:09:51.679 actually have commercial rights to your generations. 00:09:52.139 --> 00:09:54.759 So what does this all mean? For you listening 00:09:54.759 --> 00:09:57.320 to this right now, the journey from level one 00:09:57.320 --> 00:10:00.500 to level five is an identity shift. You are moving 00:10:00.500 --> 00:10:03.679 from someone playing with a neat toy to an executive 00:10:03.679 --> 00:10:06.340 director running a creative pipeline. Master 00:10:06.340 --> 00:10:09.220 one level at a time. The tools will update every 00:10:09.220 --> 00:10:11.960 week, but the core structural mindset stays the 00:10:11.960 --> 00:10:15.799 same. Build your own unique visual DNA. The final 00:10:15.799 --> 00:10:17.940 analogy in the source nails it. The computer 00:10:17.940 --> 00:10:20.480 provides the fire, but you must direct the heat. 00:10:20.779 --> 00:10:22.940 Without you, it just burns in random directions. 00:10:23.279 --> 00:10:25.379 You know, thinking about that deliberate degradation, 00:10:25.919 --> 00:10:27.700 adding the dust and the mistakes to make it look 00:10:27.700 --> 00:10:30.000 real, it leaves me with a pretty massive thought. 00:10:30.159 --> 00:10:32.679 What's that? If the technical barriers of Hollywood 00:10:32.679 --> 00:10:36.750 are gone, if literally anyone can generate perfect, 00:10:37.070 --> 00:10:40.110 flawless, 4K cinematic lighting from their bedroom, 00:10:40.690 --> 00:10:42.830 will perfection just become completely boring? 00:10:43.289 --> 00:10:45.669 In a world where AI effortlessly creates a flawless 00:10:45.669 --> 00:10:49.330 image, will human mistakes, our messy, unpredictable, 00:10:49.350 --> 00:10:51.649 lived -in imperfections, will that actually become 00:10:51.649 --> 00:10:54.769 the new premium currency of storytelling? That 00:10:54.769 --> 00:10:57.490 is something to chew on. Perfect is cheap. Messy 00:10:57.490 --> 00:10:59.710 is human. Thank you so much for joining us on 00:10:59.710 --> 00:11:02.210 this deep dive. Keep directing that heat, and 00:11:02.210 --> 00:11:03.149 we will see you next time.