WEBVTT 00:00:00.270 --> 00:00:04.089 The headlines, they promise Hollywood level AI 00:00:04.089 --> 00:00:07.230 movies. You know, just from a few words, you 00:00:07.230 --> 00:00:09.109 see these incredible clips online and you think, 00:00:09.230 --> 00:00:12.630 wow, the era of limitless creation is here. But 00:00:12.630 --> 00:00:16.390 the reality is, making a full story, where the 00:00:16.390 --> 00:00:19.370 main character doesn't change their face or their 00:00:19.370 --> 00:00:21.469 voice every eight seconds, that's actually the 00:00:21.469 --> 00:00:23.010 number one technical challenge right now. Yeah, 00:00:23.030 --> 00:00:25.649 it really is. So today, we're diving into the 00:00:25.649 --> 00:00:28.339 systematic process that actually, well, solve 00:00:28.339 --> 00:00:30.640 that consistency problem. Welcome to the deep 00:00:30.640 --> 00:00:33.460 dive. We've got a, I think, really necessary, 00:00:33.679 --> 00:00:36.560 repeatable four -step method. It basically turns 00:00:36.560 --> 00:00:40.119 those chaotic one -off AI video clips into something 00:00:40.119 --> 00:00:42.759 continuous, a coherent story. Right. And this 00:00:42.759 --> 00:00:45.359 is pretty critical for anyone hoping to graduate 00:00:45.359 --> 00:00:48.640 from just generating cool, isolated shots to 00:00:48.640 --> 00:00:50.799 actually crafting a real narrative. That's the 00:00:50.799 --> 00:00:52.659 mission for today. We'll show you exactly how 00:00:52.659 --> 00:00:54.619 to set up your characters, let's call it DNA, 00:00:55.079 --> 00:00:58.380 using just a single static image. than how to 00:00:58.380 --> 00:01:00.619 lock that visual identity into every scene afterwards. 00:01:01.020 --> 00:01:03.719 And crucially, that final step. fixing the audio 00:01:03.719 --> 00:01:06.099 inconsistencies, because those just instantly 00:01:06.099 --> 00:01:08.200 ruin the immersion, don't they? They absolutely 00:01:08.200 --> 00:01:10.799 do. So this is your operational blueprint, really, 00:01:10.920 --> 00:01:14.480 your shortcut to getting reliable AI storytelling. 00:01:14.760 --> 00:01:16.079 We're going to cut through the hype and give 00:01:16.079 --> 00:01:18.900 you the actual workflow reality. OK, let's unpack 00:01:18.900 --> 00:01:21.040 this core issue then. Yeah. We kind of assume 00:01:21.040 --> 00:01:24.280 AI remembers things, right? If you use an LLM 00:01:24.280 --> 00:01:27.439 for a story, maybe a specialized chat GPT, the 00:01:27.439 --> 00:01:30.340 characters stick. The model holds onto that context. 00:01:30.879 --> 00:01:33.280 But why is it that when we shift to AI video, 00:01:33.480 --> 00:01:37.439 that memory just seems to vanish completely? 00:01:37.620 --> 00:01:40.040 Yeah, that's the blank slate problem. And it's 00:01:40.040 --> 00:01:42.040 fascinating, really, because the underlying tech 00:01:42.040 --> 00:01:44.120 is just fundamentally different LLMs. They work 00:01:44.120 --> 00:01:46.579 with tokens, text, they have a context window, 00:01:46.959 --> 00:01:49.510 a kind of working memory. OK. But when you generate 00:01:49.510 --> 00:01:51.829 video, usually with a diffusion model, you're 00:01:51.829 --> 00:01:53.709 often starting from pure noise. You're generating 00:01:53.709 --> 00:01:56.269 pixels, not words. Starting from noise. Right, 00:01:56.290 --> 00:01:58.370 exactly. So every single time you ask the tool, 00:01:58.510 --> 00:02:01.069 make me a clip, it genuinely starts fresh. Even 00:02:01.069 --> 00:02:03.930 if you feed it the exact same words, the random 00:02:03.930 --> 00:02:06.109 noise seed is different. The output's slightly 00:02:06.109 --> 00:02:08.479 different. And poof, continuity is gone. gone 00:02:08.479 --> 00:02:11.180 immediately. The model just doesn't have that 00:02:11.180 --> 00:02:13.860 built -in persistent memory for a character's 00:02:13.860 --> 00:02:16.280 visual look across generations. So it's not that 00:02:16.280 --> 00:02:18.919 the system is actively trying to forget our character. 00:02:19.639 --> 00:02:23.419 It just, well, it lacks the architecture to persistently 00:02:23.419 --> 00:02:25.939 remember visual details when starting a new task. 00:02:26.199 --> 00:02:28.120 Precisely. You know, you'll get that perfect 00:02:28.120 --> 00:02:31.060 eight second clip. Your brave knight shining 00:02:31.060 --> 00:02:33.300 armor looks great. Then you generate the next 00:02:33.300 --> 00:02:35.599 scene, maybe the knight walking into a castle. 00:02:36.020 --> 00:02:38.699 And suddenly the armor shifts from silver to, 00:02:38.699 --> 00:02:40.979 I don't know, bronze. His face looks five years 00:02:40.979 --> 00:02:44.870 older maybe. And the voice. Totally different 00:02:44.870 --> 00:02:47.449 it completely breaks the suspension of disbelief 00:02:47.449 --> 00:02:49.569 as a viewer you immediately feel like okay This 00:02:49.569 --> 00:02:51.770 is kind of amateur it does and you know I still 00:02:51.770 --> 00:02:53.909 wrestle with prompt drift myself sometimes if 00:02:53.909 --> 00:02:55.969 I don't use a really strict external reference 00:02:55.969 --> 00:02:58.409 It's a very very common frustration even for 00:02:58.409 --> 00:03:01.409 people doing this a lot So if the tools themselves 00:03:01.409 --> 00:03:04.669 lack that core internal memory, what's the fundamental 00:03:04.669 --> 00:03:07.289 idea? We need to use instead. How do we externalize 00:03:07.289 --> 00:03:09.889 that character identity? Well, we have to externalize 00:03:09.889 --> 00:03:13.039 the memory And we do that using a single consistent 00:03:13.039 --> 00:03:16.979 visual reference image. OK, step one. It sounds 00:03:16.979 --> 00:03:19.020 a bit counterintuitive for making video, doesn't 00:03:19.020 --> 00:03:22.280 it? We start by creating a single still picture. 00:03:23.090 --> 00:03:26.750 Why is this character image the DNA for the whole 00:03:26.750 --> 00:03:29.569 project? That picture becomes the immutable reference 00:03:29.569 --> 00:03:32.330 point. It's what you feed back to the AI again 00:03:32.330 --> 00:03:35.069 and again to basically force consistency. Okay. 00:03:35.229 --> 00:03:38.449 The key here is extreme, almost surgical detail 00:03:38.449 --> 00:03:41.169 in that first prompt. You need to define every 00:03:41.169 --> 00:03:43.150 little pixel of that character you want. So you 00:03:43.150 --> 00:03:44.969 can't just say a robot. You've got to blueprint 00:03:44.969 --> 00:03:47.189 the character meticulously. Absolutely. Think 00:03:47.189 --> 00:03:49.110 like an engineer designing it. Like the example 00:03:49.110 --> 00:03:52.449 prompt. A friendly, futuristic robot librarian. 00:03:52.939 --> 00:03:55.879 smooth, white metallic body, glowing blue lines, 00:03:56.300 --> 00:03:58.520 simple dome head, large digital visor showing 00:03:58.520 --> 00:04:01.500 friendly animated eyes, wearing a smart gray 00:04:01.500 --> 00:04:04.479 vest. That level of detail. That's the specificity 00:04:04.479 --> 00:04:05.919 you need. And here's an important little technique 00:04:05.919 --> 00:04:08.360 at the start. Turn off any style consistency 00:04:08.360 --> 00:04:10.259 settings when you're creating this very first 00:04:10.259 --> 00:04:12.900 image. Ah, okay. Why's that? You want the AI 00:04:12.900 --> 00:04:15.840 to give you maximum creativity initially. Let 00:04:15.840 --> 00:04:18.740 it explore a bit. Then you review the options 00:04:18.740 --> 00:04:21.519 and pick the single best, usually the full frontal 00:04:21.519 --> 00:04:25.540 view image that nails your vision. OK, that makes 00:04:25.540 --> 00:04:28.120 sense. But let's say I love the white robot librarian 00:04:28.120 --> 00:04:30.800 I got. But maybe halfway through making my video, 00:04:30.939 --> 00:04:34.399 I think, hmm, those blue lines, maybe they should 00:04:34.399 --> 00:04:37.560 be a warm orange instead. If I just change the 00:04:37.560 --> 00:04:40.259 prompt text, won't the whole robot image drift? 00:04:40.860 --> 00:04:42.819 Well, that depends on the settings you use next. 00:04:42.939 --> 00:04:45.279 If you just edit the text prompt alone, yes, 00:04:45.519 --> 00:04:47.379 you absolutely risk drift. Things will change 00:04:47.379 --> 00:04:50.000 subtly. all over. Right. This is where the tools 00:04:50.000 --> 00:04:52.120 offer features usually called something like 00:04:52.120 --> 00:04:54.259 precise reference or maybe structure reference. 00:04:54.259 --> 00:04:55.939 You have to use one of those. OK, walk us through 00:04:55.939 --> 00:04:58.100 the difference there. Precise versus structure 00:04:58.100 --> 00:05:00.500 reference. Sure. So if you choose precise reference, 00:05:00.579 --> 00:05:02.959 you're telling the AI to lock pretty much everything. 00:05:03.040 --> 00:05:06.540 The texture, the color, the fine details. Then 00:05:06.540 --> 00:05:09.420 if you prompt for orange lines, the AI really 00:05:09.420 --> 00:05:12.500 tries hard to change only the lines while keeping 00:05:12.500 --> 00:05:14.759 everything else identical. Now, if you choose 00:05:14.759 --> 00:05:17.019 structure reference, you're locking more the 00:05:17.019 --> 00:05:20.360 skeleton. or the pose, the overall shape. So 00:05:20.360 --> 00:05:22.439 with structure lock, you could maybe change the 00:05:22.439 --> 00:05:25.040 material from metal to wood, but the robot would 00:05:25.040 --> 00:05:28.300 keep its exact shape and stance. I see. For character 00:05:28.300 --> 00:05:30.459 consistency, like keeping the robot looking like 00:05:30.459 --> 00:05:32.759 the same robot, we generally stick with the precise 00:05:32.759 --> 00:05:36.019 option. So that precise lock lets you make small, 00:05:36.399 --> 00:05:39.120 controlled changes, like the line color, without 00:05:39.120 --> 00:05:41.480 the whole character morphing into something else. 00:05:41.540 --> 00:05:44.699 That's the nuance, exactly. And that brings us 00:05:44.699 --> 00:05:48.040 neatly to step two. the starting frames. We take 00:05:48.040 --> 00:05:50.439 that perfect reference image we made. The DNA 00:05:50.439 --> 00:05:53.300 image. The DNA image, right. We upload it and 00:05:53.300 --> 00:05:55.180 critically, we make sure that precise reference 00:05:55.180 --> 00:05:57.980 feature stays on. Now we're setting the stage 00:05:57.980 --> 00:06:00.720 for each scene. So we're creating the static 00:06:00.720 --> 00:06:03.100 starting point for each video clip. We place 00:06:03.100 --> 00:06:05.100 our locked character into a new background, a 00:06:05.100 --> 00:06:07.139 new environment. Let's use the contrast examples. 00:06:07.420 --> 00:06:09.939 Scene one. The robot is pointing to a book for 00:06:09.939 --> 00:06:13.839 a young student. And then scene two. Same robot, 00:06:14.199 --> 00:06:16.040 but now it's leaning forward, maybe listening 00:06:16.040 --> 00:06:19.240 carefully to an elderly man sitting in a comfy 00:06:19.240 --> 00:06:22.740 armchair. And the key is the robot's face, its 00:06:22.740 --> 00:06:26.500 body. Those color lines, they look visually identical 00:06:26.500 --> 00:06:29.160 across both of those static frames. It's only 00:06:29.160 --> 00:06:31.259 the background and the other people that change. 00:06:31.560 --> 00:06:34.360 The reference image is really doing all the heavy 00:06:34.360 --> 00:06:36.879 lifting for consistency there. So when moving 00:06:36.879 --> 00:06:39.259 from step one, creating that reference, to step 00:06:39.259 --> 00:06:42.100 two, making these starting frames, what's the 00:06:42.100 --> 00:06:43.899 immediate problem if someone forgets to turn 00:06:43.899 --> 00:06:46.279 that precise reference or locking feature on? 00:06:46.459 --> 00:06:48.839 Well, the character will immediately start to 00:06:48.839 --> 00:06:51.139 drift in the new scene. It might pick up lighting 00:06:51.139 --> 00:06:52.879 cues or color tones from the new background, 00:06:53.040 --> 00:06:54.699 and bam, you're right back to square one with 00:06:54.699 --> 00:06:57.240 inconsistency. Okay. Step three is where things 00:06:57.240 --> 00:06:59.660 actually start moving. We take that static frame 00:06:59.660 --> 00:07:01.420 from step two, the one with the locked character 00:07:01.420 --> 00:07:03.779 in the scene, and we use an image -to -video 00:07:03.779 --> 00:07:05.899 tool to bring it to life. Yeah, this is where 00:07:05.899 --> 00:07:08.740 the magic happens. But it's fragile magic, you 00:07:08.740 --> 00:07:11.459 know? We're telling the AI what should move, 00:07:11.540 --> 00:07:14.019 and just as importantly, how it should move. 00:07:14.019 --> 00:07:16.550 Right. This requires really precise... instructions 00:07:16.550 --> 00:07:19.310 about the motion, the choreography. It's not 00:07:19.310 --> 00:07:21.470 just describing the background anymore. So we 00:07:21.470 --> 00:07:24.350 need another super detailed prompt, but this 00:07:24.350 --> 00:07:27.389 time focusing on the action. Just saying the 00:07:27.389 --> 00:07:29.629 robot points its finger up at the book. Yeah. 00:07:29.949 --> 00:07:31.990 That's not gonna cut it anymore, is it? Absolutely 00:07:31.990 --> 00:07:34.389 not. We have to be directive, tell it the pace, 00:07:34.550 --> 00:07:36.769 the scope of the movement. So for that first 00:07:36.769 --> 00:07:39.610 clip, the robot and the student, the full prompt 00:07:39.610 --> 00:07:42.310 needs to be something like that. Okay. The robot 00:07:42.310 --> 00:07:45.110 librarian points its finger up at the book. Its 00:07:45.110 --> 00:07:47.350 blue visor blinks slowly. The young girl looks 00:07:47.350 --> 00:07:50.490 up. Then maybe add camera direction. The camera 00:07:50.490 --> 00:07:53.629 slowly pushes in towards the girl's face. Both 00:07:53.629 --> 00:07:56.029 characters are relatively still. The movement 00:07:56.029 --> 00:07:58.730 is very gentle. Make the scene eight seconds 00:07:58.730 --> 00:08:02.170 long. Whoa. Okay, imagine having a system that 00:08:02.170 --> 00:08:05.350 can reliably take those super precise, gentle 00:08:05.350 --> 00:08:08.069 motion instructions, turn them into a cohesive 00:08:08.069 --> 00:08:10.879 eight second clip. and then do that consistently 00:08:10.879 --> 00:08:13.899 across, say, 100 scenes. That's the real power 00:08:13.899 --> 00:08:15.839 we're trying to unlock here. It really is. It's 00:08:15.839 --> 00:08:18.180 all about control. And an advanced tip here, 00:08:18.399 --> 00:08:20.420 technically, sometimes you need to use what's 00:08:20.420 --> 00:08:23.000 called prompt weighting syntax, things like using 00:08:23.000 --> 00:08:24.879 parentheses or special keywords. To tell the 00:08:24.879 --> 00:08:27.079 model what's most important. Exactly. To tell 00:08:27.079 --> 00:08:30.220 it, OK, focus maybe 90 % of your effort on the 00:08:30.220 --> 00:08:33.419 slow camera push and only 10 % on the robot's 00:08:33.419 --> 00:08:35.750 little secondary action, like the blinking. That 00:08:35.750 --> 00:08:37.769 helps systematize the prompt engineering, doesn't 00:08:37.769 --> 00:08:40.149 it? Makes scaling it up feel more achievable, 00:08:40.450 --> 00:08:42.929 which leads to the idea of maybe using a prompt 00:08:42.929 --> 00:08:46.129 helper, like creating a custom AI assistant, 00:08:46.490 --> 00:08:48.669 maybe a specialized Gemini or something, whose 00:08:48.669 --> 00:08:51.129 only job is to translate a simple idea like, 00:08:51.169 --> 00:08:54.549 robot shows the girl a book, into that really 00:08:54.549 --> 00:08:57.370 technical, weighted, super detailed video prompt. 00:08:57.549 --> 00:08:59.549 That's absolutely the way forward for efficiency. 00:08:59.750 --> 00:09:01.850 You systematize the inputs to try and stabilize 00:09:01.850 --> 00:09:04.820 the outputs. So if the goal here is both consistency 00:09:04.820 --> 00:09:07.419 and quality in these video clips. What's the 00:09:07.419 --> 00:09:10.000 one cardinal rule a creator should never break 00:09:10.000 --> 00:09:12.440 when actually generating the video in step three? 00:09:12.600 --> 00:09:15.600 Oh, simple. Never generate only one video output. 00:09:16.059 --> 00:09:18.820 You absolutely must create multiples I'd say 00:09:18.820 --> 00:09:20.519 at least three to five versions of each clips. 00:09:20.559 --> 00:09:23.220 You can pick the best one You need quality control 00:09:23.220 --> 00:09:25.799 looking for glitches weird flickers distortions 00:09:25.799 --> 00:09:28.740 right always generate options Okay, so we've 00:09:28.740 --> 00:09:30.360 worked through visual consistency with the first 00:09:30.360 --> 00:09:33.539 three steps reference image starting frame motion 00:09:33.539 --> 00:09:37.179 generation But the second huge technical trap 00:09:37.179 --> 00:09:41.059 is audio mm -hmm big one our robot now looks 00:09:41.059 --> 00:09:44.000 identical in every scene, which is great, but 00:09:44.000 --> 00:09:46.539 when it speaks, it might sound like a completely 00:09:46.539 --> 00:09:49.279 different actor in every single clip. How do 00:09:49.279 --> 00:09:51.799 we fix this immersion breaking problem? Right. 00:09:51.899 --> 00:09:54.159 We have to isolate the audio and treat it completely 00:09:54.159 --> 00:09:56.299 separately. That's step four. You pretty much 00:09:56.299 --> 00:09:59.519 have to use an external voice cloning or a text 00:09:59.519 --> 00:10:01.480 -to -speech service, something like 11 Labs, 00:10:01.519 --> 00:10:04.299 for example. OK. The goal is to create a single, 00:10:04.620 --> 00:10:06.639 consistent voice for your character. So we pick 00:10:06.639 --> 00:10:09.100 one definitive voice, let's call it Rachel, for 00:10:09.100 --> 00:10:11.399 our robot. And that specific voice profile will 00:10:11.399 --> 00:10:13.659 be used for our robot across every single scene 00:10:13.659 --> 00:10:15.539 it speaks in. It doesn't matter what prompt we 00:10:15.539 --> 00:10:17.679 use for the video itself. Correct. So the workflow 00:10:17.679 --> 00:10:19.500 usually starts with scripting the dialogue first. 00:10:19.600 --> 00:10:21.940 Then you take your final video clip, upload it 00:10:21.940 --> 00:10:23.860 to the voice tool, choose your specific voice 00:10:23.860 --> 00:10:26.559 profile or Rachel, and generate that new clean 00:10:26.559 --> 00:10:30.039 consistent audio track just for the dialog. And 00:10:30.039 --> 00:10:34.320 this process. it has to be repeated meticulously 00:10:34.320 --> 00:10:37.159 for every single scene where the robot speaks, 00:10:37.360 --> 00:10:39.500 right? Using the exact same voice model, the 00:10:39.500 --> 00:10:42.559 same settings every time. Exactly. It's repetitive, 00:10:42.559 --> 00:10:44.740 yeah, but that's the only way to guarantee audio 00:10:44.740 --> 00:10:47.080 consistency for that character. Now, you mentioned 00:10:47.080 --> 00:10:49.580 this is the difficult part, the final edit. So 00:10:49.580 --> 00:10:52.299 we have our video clips with consistent visuals, 00:10:52.919 --> 00:10:56.559 and we have our new consistent Rachel audio tracks 00:10:56.559 --> 00:10:59.700 for the robot's lines. But the original video 00:10:59.700 --> 00:11:01.570 clips, They have audio too, right? Well, they 00:11:01.570 --> 00:11:04.590 do, and it's usually unusable because it's inconsistent 00:11:04.590 --> 00:11:07.509 or just noise. So in your video editor, the first 00:11:07.509 --> 00:11:09.629 thing you do is mute the original audio track 00:11:09.629 --> 00:11:11.389 from the video clip completely. Just silence 00:11:11.389 --> 00:11:14.559 it. Okay, easy enough, but wait. If our voice 00:11:14.559 --> 00:11:17.919 tool, like 11 Labs, processes the dialogue, doesn't 00:11:17.919 --> 00:11:19.840 it usually process all the dialogue in the clip? 00:11:19.940 --> 00:11:21.940 Yeah. What about the voice of the young girl 00:11:21.940 --> 00:11:24.139 or the elderly man? Wouldn't they also sound 00:11:24.139 --> 00:11:26.919 like Rachel? Ah, yes. That's the flaw you typically 00:11:26.919 --> 00:11:29.419 have to fix manually. Often, the voice tool will 00:11:29.419 --> 00:11:31.299 change all the voices in the segment to your 00:11:31.299 --> 00:11:33.679 chosen one, the robot voice in this case. So 00:11:33.679 --> 00:11:35.559 everything sounds like the robot. Potentially, 00:11:35.799 --> 00:11:38.320 yes. So you have to do some careful audio surgery 00:11:38.320 --> 00:11:41.019 in your editor. You need to identify precisely 00:11:41.019 --> 00:11:43.899 where the robot stops speaking and say, the girl 00:11:43.899 --> 00:11:46.940 starts. You keep the robot's original track muted. 00:11:47.460 --> 00:11:50.279 You place the new clean Rachel audio track for 00:11:50.279 --> 00:11:53.960 the robot's lines. But then you have to... painstakingly 00:11:53.960 --> 00:11:55.700 cut out the parts of that new Rachel track where 00:11:55.700 --> 00:11:57.559 the girl was supposed to be speaking. And let 00:11:57.559 --> 00:11:59.980 her original audio come through. Or replace it, 00:12:00.000 --> 00:12:01.940 too. Exactly. You either let her original generated 00:12:01.940 --> 00:12:04.720 voice come through from the muted track by unmuting 00:12:04.720 --> 00:12:07.759 just those sections, or ideally, you generate 00:12:07.759 --> 00:12:09.600 another consistent voice for the girl using the 00:12:09.600 --> 00:12:12.759 same process and layer that in. Wow, OK. That 00:12:12.759 --> 00:12:15.299 is intricate. That's like frame accurate audio 00:12:15.299 --> 00:12:18.090 editing. Step four isn't just clicking generate 00:12:18.090 --> 00:12:21.450 audio. It's manually solving a multi -voice synchronization 00:12:21.450 --> 00:12:24.070 and replacement puzzle in a final video editor. 00:12:24.269 --> 00:12:26.350 It really is. It's the necessary final polish. 00:12:26.830 --> 00:12:29.070 And then adding some subtle background sound, 00:12:29.330 --> 00:12:32.250 like a constant library hum or room tone throughout 00:12:32.250 --> 00:12:35.990 the entire scene that helps mask any tiny imperfections 00:12:35.990 --> 00:12:38.429 from the cuts and really completes the illusion 00:12:38.429 --> 00:12:42.350 of quality. So if a creator just skips this whole 00:12:42.350 --> 00:12:47.970 audio workflow, Ignore step four. What's the 00:12:47.970 --> 00:12:50.250 immediate flaw that every single viewer is going 00:12:50.250 --> 00:12:52.230 to notice, even if the visuals are absolutely 00:12:52.230 --> 00:12:54.789 perfect? Oh, it's instant. Inconsistent character 00:12:54.789 --> 00:12:57.210 voices just make the whole thing feel amateurish 00:12:57.210 --> 00:12:59.809 or kind of cheap, cobbled together. It completely 00:12:59.809 --> 00:13:02.370 shatters the believability. OK, let's think about 00:13:02.370 --> 00:13:05.220 scaling this up. What if we have multiple recurring 00:13:05.220 --> 00:13:08.639 characters, like our robot librarian and say 00:13:08.639 --> 00:13:11.000 a little floating drone sidekick that follows 00:13:11.000 --> 00:13:13.779 it around? How does this four -step method handle 00:13:13.779 --> 00:13:16.440 that added complexity? Well, the process is fundamentally 00:13:16.440 --> 00:13:18.840 the same, but you basically duplicate the complexity 00:13:18.840 --> 00:13:21.440 and maybe even square it. You generate a separate 00:13:21.440 --> 00:13:23.639 reference image for each character. Robot A gets 00:13:23.639 --> 00:13:26.820 an image. Drone B gets its own image. Then in 00:13:26.820 --> 00:13:28.740 step two, setting up the scene frames, you'd 00:13:28.740 --> 00:13:31.259 ideally upload both reference images. But now 00:13:31.259 --> 00:13:34.220 you run into the challenge of regional prompting. 00:13:34.340 --> 00:13:36.840 Regional prompting. You mean telling the AI which 00:13:36.840 --> 00:13:38.840 reference image applies to which part of the 00:13:38.840 --> 00:13:41.120 scene, like this blob is the robot, that blob 00:13:41.120 --> 00:13:44.019 is the drone? Exactly that. Because without more 00:13:44.019 --> 00:13:46.480 advanced tools, tools, maybe like masking, if 00:13:46.480 --> 00:13:48.620 you just upload two reference images in a scene 00:13:48.620 --> 00:13:51.940 prompt, they often conflict or kind of blend 00:13:51.940 --> 00:13:54.620 together into mush. You need to use advanced 00:13:54.620 --> 00:13:56.820 prompt syntax or specific features if the tool 00:13:56.820 --> 00:14:01.139 supports it to say, okay, apply robot A's reference 00:14:01.139 --> 00:14:04.639 to the character in the center. Apply Drone B's 00:14:04.639 --> 00:14:06.500 reference only to the object in the top right 00:14:06.500 --> 00:14:08.679 of the frame. That sounds like a significant 00:14:08.679 --> 00:14:11.399 technical hurdle that wasn't obvious in this 00:14:11.399 --> 00:14:13.639 simple one -character workflow. It definitely 00:14:13.639 --> 00:14:15.679 is. It adds a layer of complexity. And then, 00:14:15.679 --> 00:14:18.279 of course, in step four, the audio, you need 00:14:18.279 --> 00:14:21.399 to select two distinct, consistent voices. Maybe 00:14:21.399 --> 00:14:24.019 Nova for the robot and a higher -pitched, whirring 00:14:24.019 --> 00:14:26.809 Sparky for the drone. And then in your final 00:14:26.809 --> 00:14:28.870 edit, you're managing potentially three separate 00:14:28.870 --> 00:14:31.409 audio tracks that need careful cutting and syncing. 00:14:31.549 --> 00:14:34.029 The robot track, the drone track, and any human 00:14:34.029 --> 00:14:36.570 character tracks. Okay, that paints a clearer 00:14:36.570 --> 00:14:39.850 picture of the real work involved. Now we hear 00:14:39.850 --> 00:14:42.230 constant news about next -generation tools. You 00:14:42.230 --> 00:14:44.570 know, models like Sora, they talk about built 00:14:44.570 --> 00:14:47.909 -in features like cameo or continuity. Does this 00:14:47.909 --> 00:14:50.690 whole four -step multi -tool method we've outlined, 00:14:51.190 --> 00:14:53.570 does it become obsolete once those tools launch 00:14:53.570 --> 00:14:56.200 widely? You know, I really don't think so. Not 00:14:56.200 --> 00:14:58.629 completely anyway. That cameo feature, from what 00:14:58.629 --> 00:15:01.090 we understand, seems focused on tracking real 00:15:01.090 --> 00:15:03.909 people across clips to keep them visually consistent. 00:15:04.350 --> 00:15:06.570 That doesn't directly help you create and maintain 00:15:06.570 --> 00:15:09.350 a fictional character, like our specific robot 00:15:09.350 --> 00:15:11.850 librarian design. OK, that makes sense. And what 00:15:11.850 --> 00:15:14.149 about the promised recut or continuity features? 00:15:14.350 --> 00:15:16.549 Those sound like they'll help smooth out transitions 00:15:16.549 --> 00:15:19.549 between generated scenes, maybe ensure physics 00:15:19.549 --> 00:15:21.750 remain stable, things like that. It's more about 00:15:21.750 --> 00:15:24.429 the quality control of the video generation itself. 00:15:24.610 --> 00:15:27.019 It still doesn't seem to solve the fundamental 00:15:27.019 --> 00:15:29.080 problem of creating the consistent character 00:15:29.080 --> 00:15:32.679 look in the first place, nor does it solve the 00:15:32.679 --> 00:15:35.500 absolutely critical issue of multi -character 00:15:35.500 --> 00:15:38.500 audio consistency. So I think this four -step 00:15:38.500 --> 00:15:41.399 method or something very like it remains the 00:15:41.399 --> 00:15:43.340 essential backbone of the production strategy 00:15:43.340 --> 00:15:45.840 for the foreseeable future. So the realistic 00:15:45.840 --> 00:15:49.399 toolkit picture emerges. To make even that short, 00:15:49.399 --> 00:15:53.870 consistent video of the robot librarian. We potentially 00:15:53.870 --> 00:15:56.570 ended up using, what, up to six different specialized 00:15:56.570 --> 00:15:58.129 tools? Yeah, let's count them. You've got your 00:15:58.129 --> 00:16:00.269 image creation tool for step one, your scene 00:16:00.269 --> 00:16:02.690 framing tool or feature for step two, using the 00:16:02.690 --> 00:16:04.909 reference, maybe an optional AI prompt helper, 00:16:05.710 --> 00:16:07.970 then the core image to video tool for step three, 00:16:08.370 --> 00:16:10.549 the external voice tool for step four's audio 00:16:10.549 --> 00:16:13.549 generation. And finally, a complex video editor 00:16:13.549 --> 00:16:16.049 for step four's audio fixing and final assembly. 00:16:16.289 --> 00:16:18.049 It's really a pipeline, isn't it? Not a single 00:16:18.049 --> 00:16:20.110 magic app. It's absolutely a workflow, a pipeline. 00:16:20.429 --> 00:16:22.690 So looking at this complex multi -tool approach, 00:16:23.389 --> 00:16:25.169 what do you think is the biggest trap related 00:16:25.169 --> 00:16:27.450 to just strategic planning that creators need 00:16:27.450 --> 00:16:30.289 to consciously avoid? Hmm. I'd say rushing those 00:16:30.289 --> 00:16:32.850 crucial early steps. If you rush the character 00:16:32.850 --> 00:16:35.090 design in step one, if you settle for a mediocre 00:16:35.090 --> 00:16:38.350 reference image, All the complexity and effort 00:16:38.350 --> 00:16:41.090 you put into the next five tools, it just ends 00:16:41.090 --> 00:16:44.750 up amplifying that initial flaw. Bad character 00:16:44.750 --> 00:16:47.330 design inevitably leads to a bad final video, 00:16:47.629 --> 00:16:50.409 no matter how fancy the tools are later. That's 00:16:50.409 --> 00:16:52.889 spot on. So the core takeaway for you listening 00:16:52.889 --> 00:16:56.159 is pretty clear. Consistent AI video isn't achieved 00:16:56.159 --> 00:16:59.200 by some single magic button right now. It's about 00:16:59.200 --> 00:17:01.919 intelligently combining several specialized tools 00:17:01.919 --> 00:17:04.700 into this reliable four -step process. Right. 00:17:04.960 --> 00:17:07.200 You start with that solid reference image, you 00:17:07.200 --> 00:17:09.220 move to careful starting frames for each scene, 00:17:09.839 --> 00:17:11.920 you ensure quality movement generation with detail 00:17:11.920 --> 00:17:14.220 prompts, and then critically you tackle the audio 00:17:14.220 --> 00:17:16.799 correction manually and meticulously. The consistency 00:17:16.799 --> 00:17:19.680 of your characters visually and audibly. It's 00:17:19.680 --> 00:17:22.259 now entirely in your hands, really, through disciplined 00:17:22.259 --> 00:17:24.500 execution of this kind of process. And this applies 00:17:24.500 --> 00:17:26.500 everywhere, doesn't it? Consistent marketing 00:17:26.500 --> 00:17:29.259 mascots, characters in explainer videos, virtual 00:17:29.259 --> 00:17:31.619 presenters for education. Any project needing 00:17:31.619 --> 00:17:34.619 consistency. Planning that reference image properly 00:17:34.619 --> 00:17:37.319 up front, that's the key to unlocking success 00:17:37.319 --> 00:17:39.299 for all those projects. But what's the best way 00:17:39.299 --> 00:17:42.559 to actually learn this? Honestly, just start 00:17:42.559 --> 00:17:45.309 small. Grab the tools you have access to, pick 00:17:45.309 --> 00:17:48.130 a really simple character idea, and just dedicate 00:17:48.130 --> 00:17:50.269 some time to practicing each of these four steps. 00:17:50.369 --> 00:17:52.769 Get a feel for it today. OK, so we've kind of 00:17:52.769 --> 00:17:54.849 solved the technical challenge of consistency, 00:17:55.049 --> 00:17:57.470 which turns out to require combining maybe half 00:17:57.470 --> 00:18:01.589 a dozen different tools intelligently. But if 00:18:01.589 --> 00:18:04.170 AI creativity right now means we have to constantly 00:18:04.170 --> 00:18:06.390 patchwork these specialized services together, 00:18:07.079 --> 00:18:09.740 Will this fractured workflow ever really converge 00:18:09.740 --> 00:18:13.380 into a single cohesive interface? Or is managing 00:18:13.380 --> 00:18:16.279 this complex multi -tool pipeline actually the 00:18:16.279 --> 00:18:18.980 future of deep AI creativity? That's something 00:18:18.980 --> 00:18:20.900 for you to mull over. Thanks for diving deep 00:18:20.900 --> 00:18:22.299 with us today. We'll catch you next time.