WEBVTT 00:00:00.000 --> 00:00:03.200 Imagine directing an entire cinematic movie where 00:00:03.200 --> 00:00:07.099 your only camera is a keyboard. Beat. No set. 00:00:07.320 --> 00:00:11.279 No crew. Just you. Right. Just you. And a machine 00:00:11.279 --> 00:00:14.460 learning to dream on command. Welcome to this 00:00:14.460 --> 00:00:16.379 deep dive. I am so glad you were here with us 00:00:16.379 --> 00:00:17.859 today. Yeah, thanks for having me. I'm excited 00:00:17.859 --> 00:00:19.899 for this one. We are going to take our time with 00:00:19.899 --> 00:00:22.500 this topic. We're slowing things down just a 00:00:22.500 --> 00:00:26.480 bit to break apart a highly structured, repeatable 00:00:26.480 --> 00:00:30.089 workflow for AI video creation. Which is something 00:00:30.089 --> 00:00:32.570 a lot of people desperately need right now. Absolutely. 00:00:32.780 --> 00:00:35.600 We're exploring how to combine two powerful systems, 00:00:36.020 --> 00:00:38.979 Google Flow and Claude AI. The goal here is to 00:00:38.979 --> 00:00:42.100 bypass that chaotic learning curve of AI video 00:00:42.100 --> 00:00:44.780 and actually build consistent, controllable scenes. 00:00:44.960 --> 00:00:46.619 I mean, it's a huge shift. We aren't just looking 00:00:46.619 --> 00:00:48.820 at software tools today. We are looking at a 00:00:48.820 --> 00:00:51.479 complete architectural process. First, we build 00:00:51.479 --> 00:00:54.079 the writer's room. Then we construct the visual 00:00:54.079 --> 00:00:56.439 blueprint. And finally, we direct the animation 00:00:56.439 --> 00:00:58.619 piece by piece. To really understand this workflow, 00:00:58.880 --> 00:01:01.179 we first have to understand why beginners usually 00:01:01.179 --> 00:01:04.260 fail. because it comes down to choosing the wrong 00:01:04.260 --> 00:01:07.019 type of tool from the absolute start. Yeah, that 00:01:07.019 --> 00:01:09.959 is the great AI video divide. There's basically 00:01:09.959 --> 00:01:13.400 this steep learning curve that stems from fragmented 00:01:13.400 --> 00:01:15.719 tools. Right, you have individual models. Exactly, 00:01:16.019 --> 00:01:19.040 models like Kling or Runway. They give you incredibly 00:01:19.040 --> 00:01:22.400 deep control, but they require juggling multiple 00:01:22.400 --> 00:01:26.640 platforms and... multiple subscriptions. It gets 00:01:26.640 --> 00:01:29.599 overwhelming. It does. But then, there are all 00:01:29.599 --> 00:01:32.239 -in -one workspaces. Google Flow solves that 00:01:32.239 --> 00:01:34.879 juggling act by combining three specific models 00:01:34.879 --> 00:01:37.019 into one place. So what are those three? Well, 00:01:37.120 --> 00:01:39.540 you've got Nano Banana Pro for your images, you've 00:01:39.540 --> 00:01:42.780 got VO3 .1 for your video generation, and then 00:01:42.780 --> 00:01:45.640 Gemini Omni. Which brings in multimodal editing. 00:01:45.760 --> 00:01:47.879 Yes. Which, just to define that simply, means 00:01:47.879 --> 00:01:51.379 combining text, images, and video to guide the 00:01:51.379 --> 00:01:53.920 AI. Right, exactly. But the really interesting 00:01:53.920 --> 00:01:56.319 part is Claude AI's role in all this. Yeah, because 00:01:56.319 --> 00:01:57.900 Claude doesn't actually generate any art, right? 00:01:57.900 --> 00:02:00.700 No, not a single pixel. It sits alongside Google 00:02:00.700 --> 00:02:03.280 Flow, purely as a scene planner and a prompt 00:02:03.280 --> 00:02:05.519 writer. I kind of like to compare this setup 00:02:05.519 --> 00:02:08.159 to a traditional movie set. Oh, yeah. Yeah, so 00:02:08.159 --> 00:02:09.819 Google Flow is your director of photography, 00:02:10.300 --> 00:02:12.780 and Claude AI is your head writer. That's a perfect 00:02:12.780 --> 00:02:15.419 way to look at it. But let me ask you this. If 00:02:15.419 --> 00:02:19.099 Google Flow is a true all -in -one tool, why 00:02:19.099 --> 00:02:22.060 rely on an external text model like Claude at 00:02:22.060 --> 00:02:25.840 all? Well, Google Flow is optimized for manipulating 00:02:25.840 --> 00:02:28.860 visual data. Claude is just vastly superior at 00:02:28.860 --> 00:02:31.639 maintaining narrative logic over long context 00:02:31.639 --> 00:02:34.560 windows. So Claude handles the logic, freeing 00:02:34.560 --> 00:02:37.439 Google Flow to focus purely on the visuals. Precisely. 00:02:37.560 --> 00:02:39.699 You let the writer write, And the camera film. 00:02:39.939 --> 00:02:42.159 Makes sense. So now that we have our writer and 00:02:42.159 --> 00:02:44.500 our DP, we have to actually teach the writer 00:02:44.500 --> 00:02:46.520 how to talk to the DP. Yeah, and this is where 00:02:46.520 --> 00:02:49.159 people get stuck. I mean, I still wrestle with 00:02:49.159 --> 00:02:51.180 prompt drift myself where my text instructions 00:02:51.180 --> 00:02:53.539 just get messy over time. Oh, absolutely. Writing 00:02:53.539 --> 00:02:55.560 prompts from scratch every time just leads to 00:02:55.560 --> 00:02:58.879 inconsistent garbage. The AI loses the thread. 00:02:59.300 --> 00:03:02.000 Right. So the solution is creating a custom skill 00:03:02.000 --> 00:03:05.210 inside Claude AI. You literally call it the AI 00:03:05.210 --> 00:03:07.169 video prompt writer. How does that actually work 00:03:07.169 --> 00:03:09.610 in the interface? You just go to customize, then 00:03:09.610 --> 00:03:12.129 skills, then create with Claude. And you paste 00:03:12.129 --> 00:03:14.629 in exact instructions to always generate three 00:03:14.629 --> 00:03:17.509 specific types of prompts. Three types. OK, what's 00:03:17.509 --> 00:03:19.210 the first one? Number one is the design sheet 00:03:19.210 --> 00:03:22.129 that covers your characters, your props, the 00:03:22.129 --> 00:03:24.949 overall style. Number two is the storyboard. 00:03:25.030 --> 00:03:27.650 That's your panel by panel camera angles. And 00:03:27.650 --> 00:03:29.669 number three is the scene prompt, which are the 00:03:29.669 --> 00:03:32.919 direct instructions for VO 3 .1. That's so structured. 00:03:33.340 --> 00:03:35.219 The source actually uses this great example. 00:03:35.699 --> 00:03:38.960 A woman and her dog escaping Manhattan during 00:03:38.960 --> 00:03:41.539 a zombie outbreak. Yeah, a classic setup. You 00:03:41.539 --> 00:03:43.659 literally just typed that one simple sentence 00:03:43.659 --> 00:03:46.219 and Claude generates all the structured technical 00:03:46.219 --> 00:03:48.500 prompts for you. It really does. It removes all 00:03:48.500 --> 00:03:51.080 the friction. But wait, why do we need three 00:03:51.080 --> 00:03:54.280 separate highly specific prompts instead of just 00:03:54.280 --> 00:03:56.400 one master prompt describing the whole video? 00:03:56.509 --> 00:03:59.629 Because if you feed a model a massive block of 00:03:59.629 --> 00:04:03.030 text, it simply drops details. It hallucinates. 00:04:03.569 --> 00:04:06.090 It can't balance all those variables at once. 00:04:06.289 --> 00:04:08.009 Breaking it down prevents the AI from getting 00:04:08.009 --> 00:04:10.150 overwhelmed and mixing up complex instructions. 00:04:10.449 --> 00:04:12.770 Exactly. You have to segment the cognitive load. 00:04:13.229 --> 00:04:15.909 OK. So with Claude generating these text blueprints, 00:04:16.230 --> 00:04:18.430 we must translate those into a visual foundation 00:04:18.430 --> 00:04:21.310 before generating any video. Right. And this 00:04:21.310 --> 00:04:24.509 is where we move into Google Flow, specifically 00:04:24.509 --> 00:04:27.810 using Nano Banana Pro for images. So we are building 00:04:27.810 --> 00:04:30.910 the design sheet next. Yes. You must definitively 00:04:30.910 --> 00:04:33.310 establish the world. What does the main character 00:04:33.310 --> 00:04:37.149 look like? The dog. The zombie. What are the 00:04:37.149 --> 00:04:39.430 clothing and props like? You're defining the 00:04:39.430 --> 00:04:42.329 color palette too. Everything. And here's a crucial 00:04:42.329 --> 00:04:46.540 tip. Start with low resolution. Oh, to save generation 00:04:46.540 --> 00:04:48.680 credits. Yeah, it saves credits and it generates 00:04:48.680 --> 00:04:51.860 way faster. You use that low -res draft to check 00:04:51.860 --> 00:04:54.720 for mistakes. If an element is missing, you don't 00:04:54.720 --> 00:04:56.920 fix it in Google Flow. You go back to Claude. 00:04:57.079 --> 00:04:59.100 Right. You go back to Claude for text revision. 00:04:59.579 --> 00:05:01.860 And once the image is perfect, then you render 00:05:01.860 --> 00:05:05.519 it in full resolution. That is smart. So then 00:05:05.519 --> 00:05:08.209 we move to the storyboard. We use Claude to generate 00:05:08.209 --> 00:05:11.129 a 12 -panel storyboard prompt. Like the convenience 00:05:11.129 --> 00:05:13.550 store, the zombie attack, the escape. Right. 00:05:13.709 --> 00:05:16.610 And we generate this in Google Flow using the 00:05:16.610 --> 00:05:18.910 design sheet as a visual reference to lock in 00:05:18.910 --> 00:05:20.850 that consistency. You have to attach it. It's 00:05:20.850 --> 00:05:23.230 mandatory. So what actually happens under the 00:05:23.230 --> 00:05:25.810 hood if you get impatient and skip the design 00:05:25.810 --> 00:05:28.170 sheet step? The model just invents a completely 00:05:28.170 --> 00:05:30.970 new reality for every panel. The woman and the 00:05:30.970 --> 00:05:33.149 dog will look different in every single shot. 00:05:33.509 --> 00:05:35.930 Without it, the AI literally forgets what your 00:05:35.930 --> 00:05:37.569 characters look like between every shot. Yeah, 00:05:37.569 --> 00:05:40.389 it has zero object permanence without that visual 00:05:40.389 --> 00:05:43.529 anchor. Sponsor. Okay. So we have our storyboard. 00:05:43.709 --> 00:05:46.329 We have our character DNA. Now we finally step 00:05:46.329 --> 00:05:48.670 onto the stage to make things move. The fun part. 00:05:48.970 --> 00:05:52.230 Switching over to VO 3 .1 for video generation 00:05:52.230 --> 00:05:55.050 in Google Flow. But before writing the scene 00:05:55.050 --> 00:05:58.310 prompt, the guide says you must upload two references. 00:05:58.430 --> 00:06:01.750 Yes. First, the design sheet for your world and 00:06:01.750 --> 00:06:04.750 character consistency, and second, the specific 00:06:04.750 --> 00:06:07.569 storyboard panel image. Which locks in your framing 00:06:07.569 --> 00:06:10.730 and composition. Exactly. Then, and only then, 00:06:10.790 --> 00:06:13.490 do you feed VO3 .1 the scene prompt from Claude. 00:06:13.709 --> 00:06:15.790 So for the first four panels, it's walking through 00:06:15.790 --> 00:06:18.230 the dark store, looking nervous, and then a zombie 00:06:18.230 --> 00:06:21.709 jumps out. Right. Whoa. I just have to pause 00:06:21.709 --> 00:06:24.930 and think about this. Imagine scaling to a billion 00:06:24.930 --> 00:06:27.740 queries across the globe. But here we're just 00:06:27.740 --> 00:06:30.240 intimately tweaking one single perfect frame 00:06:30.240 --> 00:06:34.439 of a zombie attack. It's wild. It really is a 00:06:34.439 --> 00:06:37.339 staggering amount of compute power, just focused 00:06:37.339 --> 00:06:40.399 on the shadow of a zombie in an aisle. But it 00:06:40.399 --> 00:06:42.060 doesn't always come out perfectly on the first 00:06:42.060 --> 00:06:44.759 try. No, definitely not. So we have to use this 00:06:44.759 --> 00:06:47.600 review and improve methodology. You do not throw 00:06:47.600 --> 00:06:50.480 away a whole clip if one cut looks wrong. Never. 00:06:50.720 --> 00:06:54.160 A weak first output is totally normal. You generate 00:06:54.160 --> 00:06:56.620 a second version with specific instructions. 00:06:56.720 --> 00:06:59.259 Like telling it to do a direct cut to a close 00:06:59.259 --> 00:07:01.720 -up. Exactly. And then you combine the strongest 00:07:01.720 --> 00:07:04.459 parts of both outputs in post. But why can't 00:07:04.459 --> 00:07:08.779 VO3 .1 just recognize a bad cut and fix it automatically 00:07:08.779 --> 00:07:11.160 in a single generation? Because it doesn't understand 00:07:11.160 --> 00:07:14.680 human anatomy or cinematic timing. It just predicts 00:07:14.680 --> 00:07:18.560 pixel patterns based on data. The AI lacks human 00:07:18.560 --> 00:07:21.540 taste. It needs us to stitch the best parts together. 00:07:21.600 --> 00:07:23.000 Right, you have to be the editor. So generating 00:07:23.000 --> 00:07:26.180 one good scene is great. But a 12 panel story 00:07:26.180 --> 00:07:28.279 will fall apart if you try to render it all at 00:07:28.279 --> 00:07:30.699 once. We have to control the AI's pacing. Yeah, 00:07:30.740 --> 00:07:32.860 you cannot do all 12 panels at once. This introduces 00:07:32.860 --> 00:07:35.360 the rule of chunking. Which means animating only 00:07:35.360 --> 00:07:39.180 four panels at a time. Right. Trying all 12 overwhelms 00:07:39.180 --> 00:07:43.060 VO 3 .1. It causes random hallucinatory transitions. 00:07:43.279 --> 00:07:45.720 It's like staffing Lego blocks of data. You do 00:07:45.720 --> 00:07:47.519 it piece by piece so the whole thing doesn't 00:07:47.519 --> 00:07:50.089 topple. I love that analogy. So for each chunk, 00:07:50.509 --> 00:07:53.230 you need three references uploaded. Yes. Number 00:07:53.230 --> 00:07:56.250 one, the cropped storyboard row. Number two, 00:07:56.490 --> 00:07:59.910 the design sheet. And number three, the final 00:07:59.910 --> 00:08:02.490 frame of the previous clip. That last one seems 00:08:02.490 --> 00:08:04.649 like a real secret weapon. Oh, it absolutely 00:08:04.649 --> 00:08:08.269 is. What about fixing weak cuts? Like, if a character 00:08:08.269 --> 00:08:10.970 shifts positions suddenly mid -scene. If that 00:08:10.970 --> 00:08:13.230 happens, instruct Vio to hold on a close -up 00:08:13.230 --> 00:08:15.949 before cutting to the next action. It hides the 00:08:15.949 --> 00:08:18.560 error. And again, you combine the best elements. 00:08:18.879 --> 00:08:20.259 I want to go back to that third reference for 00:08:20.259 --> 00:08:22.480 a second. Why is feeding the final frame of the 00:08:22.480 --> 00:08:25.120 previous clip back into the machine so critically 00:08:25.120 --> 00:08:27.040 important? Because it creates a rigid anchor 00:08:27.040 --> 00:08:30.060 in time. It prevents the model from subtly changing 00:08:30.060 --> 00:08:32.500 the lighting or the camera distance. It forces 00:08:32.500 --> 00:08:34.980 the new scene to mathematically lock into the 00:08:34.980 --> 00:08:37.379 last clip's ending. Perfectly said. Down to the 00:08:37.379 --> 00:08:40.799 exact pixel. Even with chunking, the entire project 00:08:40.799 --> 00:08:44.220 can unravel if you forget the overarching philosophy 00:08:44.220 --> 00:08:48.139 of AI video. Visual consistency over everything 00:08:48.139 --> 00:08:50.500 else. Without a doubt. The most common reason 00:08:50.500 --> 00:08:53.080 videos fall apart isn't bad text prompts. It's 00:08:53.080 --> 00:08:56.120 losing visual consistency. The characters change. 00:08:56.519 --> 00:08:58.399 The environments shift. Yeah, it just looks amateur. 00:08:58.820 --> 00:09:01.120 So rule number one is that the design sheet is 00:09:01.120 --> 00:09:03.840 the visual anchor. Keep it attached as a reference 00:09:03.840 --> 00:09:06.639 for every single generation. Let's talk about 00:09:06.639 --> 00:09:08.899 some mistakes that just burn time and credits. 00:09:09.320 --> 00:09:13.159 Number one. generating from text only. That forces 00:09:13.159 --> 00:09:16.779 the AI to guess entirely on its own. Number two, 00:09:17.039 --> 00:09:19.240 animating more than four panels at once. We covered 00:09:19.240 --> 00:09:22.340 that. It destroys the pacing. Number three, skipping 00:09:22.340 --> 00:09:24.799 the revision process or just expecting a perfect 00:09:24.799 --> 00:09:26.779 first output. You have to think like an editor. 00:09:26.840 --> 00:09:30.460 It's an iterative process. And number four, moving 00:09:30.460 --> 00:09:32.740 forward with an incomplete design sheet. Right, 00:09:32.740 --> 00:09:35.039 because any error there propagates into every 00:09:35.039 --> 00:09:36.820 single scene that follows. I do want to push 00:09:36.820 --> 00:09:39.600 back gently on that first mistake, though. Since 00:09:39.600 --> 00:09:43.059 text is how we naturally talk to AI, why is generating 00:09:43.059 --> 00:09:45.679 from text only considered such a massive mistake? 00:09:46.240 --> 00:09:48.379 Because human language is just too imprecise. 00:09:48.500 --> 00:09:51.399 If you say dark store, that means a billion different 00:09:51.399 --> 00:09:53.720 pixel variations to the model. Words are too 00:09:53.720 --> 00:09:56.419 ambiguous for video. Visual references provide 00:09:56.419 --> 00:09:58.940 the only undeniable truth. Exactly. You have 00:09:58.940 --> 00:10:02.220 to show it, not just tell it. So to recap this 00:10:02.220 --> 00:10:05.279 whole structured journey, AI video isn't luck. 00:10:05.639 --> 00:10:08.639 It's a process. It really is. You build the design 00:10:08.639 --> 00:10:11.899 sheet, you map the storyboard, you use Claude 00:10:11.899 --> 00:10:15.360 for precise prompts, you generate in small four 00:10:15.360 --> 00:10:19.039 -panel chunks with VO, and always, always keep 00:10:19.039 --> 00:10:21.340 your visual references attached. If you're going 00:10:21.340 --> 00:10:23.139 to try this for the first time, my advice is 00:10:23.139 --> 00:10:25.539 to keep it incredibly simple. One character, 00:10:25.620 --> 00:10:28.559 one location, one action scene. Master the workflow 00:10:28.559 --> 00:10:31.100 before building a complex epic. Walk before you 00:10:31.100 --> 00:10:33.259 run, for sure. It's just wild to think about 00:10:33.259 --> 00:10:36.299 where this is heading. Pete. If these dual AI 00:10:36.299 --> 00:10:39.379 tools can synthesize a cohesive, terrifying zombie 00:10:39.379 --> 00:10:41.779 escape from just a few structured constraints, 00:10:42.639 --> 00:10:45.179 Beat, what happens when this workflow gets fully 00:10:45.179 --> 00:10:47.000 automated? Yeah, that's the big question. If 00:10:47.000 --> 00:10:49.700 the AI eventually learns to manage its own visual 00:10:49.700 --> 00:10:52.759 consistency between shots, what exactly becomes 00:10:52.759 --> 00:10:54.740 the human director's role in the filmmaking of 00:10:54.740 --> 00:10:57.200 the future? Yeah, Beat, it's something to think 00:10:57.200 --> 00:10:58.659 about. Thank you so much for joining us on this 00:10:58.659 --> 00:11:00.879 Deep Dives. Take care. UT over music.