WEBVTT 00:00:00.000 --> 00:00:02.779 Have you ever tried to make an AI film where 00:00:02.779 --> 00:00:06.139 your main character, the one you spent ages designing 00:00:06.269 --> 00:00:09.830 Just changes faces in every other scene. Oh, 00:00:09.849 --> 00:00:13.810 that issue of character consistency. It is probably 00:00:13.810 --> 00:00:16.410 the biggest and most frustrating hurdle in the 00:00:16.410 --> 00:00:18.329 field right now. It just kills the immersion. 00:00:18.370 --> 00:00:20.190 It completely takes you out of it. But we have 00:00:20.190 --> 00:00:22.329 the golden key today, a trick called the three 00:00:22.329 --> 00:00:24.769 by three grid that locks the look right from 00:00:24.769 --> 00:00:27.170 the first picture. OK, let's unpack this. We're 00:00:27.170 --> 00:00:29.469 diving deep into a guide that's designed to fix 00:00:29.469 --> 00:00:32.530 that problem for good. That's right. Our listener, 00:00:32.810 --> 00:00:35.149 you, shared these really comprehensive notes 00:00:35.149 --> 00:00:39.020 on mastering character consistency in AI filmmaking. 00:00:39.539 --> 00:00:41.619 Our goal here is to move past those, you know, 00:00:41.899 --> 00:00:43.799 blurry changing versus and produce something 00:00:43.799 --> 00:00:45.880 that looks truly professional. We'll analyze 00:00:45.880 --> 00:00:48.280 the tools. We're going to break down this complex 00:00:48.280 --> 00:00:51.479 prompt formula into five clear parts and walk 00:00:51.479 --> 00:00:53.240 through the workflow. We're aiming for films 00:00:53.240 --> 00:00:55.240 that look polished, not like a collection of 00:00:55.240 --> 00:00:58.500 random photos. Exactly. You know, character consistency 00:00:58.500 --> 00:01:01.079 isn't just a technical thing. It's the key to 00:01:01.079 --> 00:01:03.380 storytelling. If your detective girl's yellow 00:01:03.380 --> 00:01:06.439 coat suddenly turns orange or her face changes, 00:01:06.519 --> 00:01:08.560 the whole film just looks unprofessional. So 00:01:08.560 --> 00:01:10.500 the three by three grid is basically a blueprint 00:01:10.500 --> 00:01:13.420 for that. Instead of one photo, we're creating 00:01:13.420 --> 00:01:16.450 a kind of reference sheet with nine different 00:01:16.450 --> 00:01:18.689 camera angles all at once. And that's how we 00:01:18.689 --> 00:01:21.930 lock the look. So what tools are powering this? 00:01:22.430 --> 00:01:25.170 The source suggests three crucial pieces of software. 00:01:25.310 --> 00:01:29.150 Right. First up is dzine .ai. Think of it as 00:01:29.150 --> 00:01:31.650 the creative hub built for filmmakers. It's what 00:01:31.650 --> 00:01:34.569 handles the character control. OK. Second is 00:01:34.569 --> 00:01:36.790 topaz gigapixel. This tool does something called 00:01:36.790 --> 00:01:39.409 upscaling. which just means making your images 00:01:39.409 --> 00:01:42.609 bigger, like 2x or 4x, but without losing quality. 00:01:42.810 --> 00:01:45.370 Exactly. It adds back details, like strands of 00:01:45.370 --> 00:01:48.150 hair. And then finally, for movement, you turn 00:01:48.150 --> 00:01:50.829 those high quality photos into smooth video using 00:01:50.829 --> 00:01:53.870 tools like Lumadri Machine or Runaway Gen 2. 00:01:53.989 --> 00:01:56.349 So why is generating a nine -shot reference sheet 00:01:56.349 --> 00:01:58.890 better than just, you know, generating nine separate 00:01:58.890 --> 00:02:01.790 images of the same character? Because the AI 00:02:01.790 --> 00:02:05.930 uses its internal memory, its short -term state, 00:02:06.469 --> 00:02:09.750 to maintain uniformity across that entire single 00:02:09.750 --> 00:02:13.349 page, it keeps it all consistent. Okay, this 00:02:13.349 --> 00:02:15.969 is the real heart of the process. A short, messy 00:02:15.969 --> 00:02:19.050 prompt means the AI is just going to fail. We 00:02:19.050 --> 00:02:21.389 need what the source calls an unbeatable formula, 00:02:21.629 --> 00:02:24.129 split into five clear parts. And before you even 00:02:24.129 --> 00:02:25.930 write anything, you have to set up the workspace 00:02:25.930 --> 00:02:29.699 in dzine .ai. Specifically, choosing the 1 .1, 00:02:29.719 --> 00:02:32.340 the square aspect ratio. Right. Why is that? 00:02:32.539 --> 00:02:34.560 A square shape just keeps those nine frames neat. 00:02:34.759 --> 00:02:36.759 It prevents the AI from cutting off the character's 00:02:36.759 --> 00:02:39.840 head or feet. Ah, that makes sense. Okay, so 00:02:39.840 --> 00:02:42.819 part one of the prompt. It's the command. A cinematic 00:02:42.819 --> 00:02:45.819 3x3 grid presenting multiple camera angles. That 00:02:45.819 --> 00:02:47.840 tells the AI to divide the image into netting 00:02:47.840 --> 00:02:50.199 sections. Simple. Part two is what you call the 00:02:50.199 --> 00:02:53.000 detail bomb. Subject and location. Yeah, don't 00:02:53.000 --> 00:02:55.419 just say a girl. Describe her hair, her eye color, 00:02:55.539 --> 00:02:58.139 the material of her clothes, any scars. The more 00:02:58.139 --> 00:02:59.979 detail you give it, the better the AI remembers 00:02:59.979 --> 00:03:02.080 her. And part three is where the magic happens. 00:03:02.360 --> 00:03:05.139 The nine specific camera angles. This makes sure 00:03:05.139 --> 00:03:07.379 your reference sheet is really versatile. You 00:03:07.379 --> 00:03:10.219 absolutely need these. You need an extreme close 00:03:10.219 --> 00:03:13.039 -up of the eyes, a medium shot from the waist 00:03:13.039 --> 00:03:15.699 up and over the shoulder, a wide shot, high angle 00:03:15.699 --> 00:03:17.800 top down, and a low angle, which makes the character 00:03:17.800 --> 00:03:20.000 look powerful. Right. Right. And then you need 00:03:20.000 --> 00:03:23.020 the side views, too. The profile, a 3 quarter 00:03:23.020 --> 00:03:25.819 rear view, and a full back view. If you skip 00:03:25.819 --> 00:03:28.620 those, the AI just improvises when your character 00:03:28.620 --> 00:03:31.379 turns around. And it's usually not good. OK, 00:03:31.379 --> 00:03:33.520 here's where it gets really interesting. Part 00:03:33.520 --> 00:03:37.300 four is cinema style and lighting. Things like 00:03:37.300 --> 00:03:40.539 volumetric fog, cinematic lighting, moody blue 00:03:40.539 --> 00:03:43.400 and orange. This is what Hollywood Eyes is the 00:03:43.400 --> 00:03:45.659 result. And it locks in the color palette. It's 00:03:45.659 --> 00:03:48.610 the glue. And then the final piece, part five, 00:03:48.669 --> 00:03:51.229 the cleanup command, images stacked together 00:03:51.229 --> 00:03:54.849 with no grid lines and no borders, no text. That 00:03:54.849 --> 00:03:57.849 no text part is so vital it saves you so much 00:03:57.849 --> 00:04:00.169 time later. The source gives this great example 00:04:00.169 --> 00:04:02.710 of a cyberpunk warrior prompt that uses all five 00:04:02.710 --> 00:04:05.590 parts. It just works. It works. So what happens 00:04:05.590 --> 00:04:07.870 if I skip the lighting keywords? What's the immediate 00:04:07.870 --> 00:04:10.270 consequence for my character? The lack of those 00:04:10.270 --> 00:04:12.349 lighting details will cause skin and clothes 00:04:12.349 --> 00:04:15.189 colors to change between scenes. Totally breaking 00:04:15.189 --> 00:04:20.240 the flow. OK, so we're moving from theory to 00:04:20.240 --> 00:04:22.779 application. We've made our 3x3 reference sheet. 00:04:23.360 --> 00:04:25.720 The source says you do a quality check, making 00:04:25.720 --> 00:04:28.620 sure details like eye color are about 80 % to 00:04:28.620 --> 00:04:31.540 90 % similar across the nine boxes. And once 00:04:31.540 --> 00:04:33.519 you've got that, the true secret is the character 00:04:33.519 --> 00:04:37.060 reference feature inside dzine .ai. OK. You save 00:04:37.060 --> 00:04:39.779 that successful 3x3 grid image, then you upload 00:04:39.779 --> 00:04:42.660 it using the little person icon button. And once 00:04:42.660 --> 00:04:46.319 it's uploaded, the AI now remembers the character. 00:04:46.439 --> 00:04:48.850 Yes. It has that file to check against. So now 00:04:48.850 --> 00:04:51.329 you can write short new prompts like, the girl 00:04:51.329 --> 00:04:53.949 is walking in the rain, and the AI automatically 00:04:53.949 --> 00:04:56.250 applies the face from your reference grid to 00:04:56.250 --> 00:04:58.290 that new scene. This whole process is so much 00:04:58.290 --> 00:05:00.370 faster than traditional filmmaking. Oh, it's 00:05:00.370 --> 00:05:02.569 not even close. After you master the workflow, 00:05:02.850 --> 00:05:05.370 the source suggests a full one -minute AI film 00:05:05.370 --> 00:05:07.569 takes only about two to three hours. Two to three 00:05:07.569 --> 00:05:10.290 hours. It's incredible. And after creating those 00:05:10.290 --> 00:05:12.370 new scenes, we need to upgrade them. That's where 00:05:12.370 --> 00:05:14.550 Topaz Gigapixel comes in. Right, for the upscaling. 00:05:14.750 --> 00:05:17.769 You drag the photos in, you choose 4x enlargement, 00:05:18.370 --> 00:05:21.069 and critically you select face recovery. That 00:05:21.069 --> 00:05:23.209 face recovery setting is what makes the eyes 00:05:23.209 --> 00:05:26.230 and skin look really, really real. It turns a 00:05:26.230 --> 00:05:29.149 good image into a high quality professional film 00:05:29.149 --> 00:05:32.850 asset. So what if I generate my grid? but one 00:05:32.850 --> 00:05:34.790 hand looks totally strange in the bottom right 00:05:34.790 --> 00:05:37.509 box. Do we have to restart everything? No, you 00:05:37.509 --> 00:05:39.730 don't. You can use the local edit tool to just 00:05:39.730 --> 00:05:42.069 brush over that area and fix that single box. 00:05:42.629 --> 00:05:44.370 All right. Let's move on to the pitfalls, the 00:05:44.370 --> 00:05:47.730 traps. The source warns against the prompt is 00:05:47.730 --> 00:05:50.129 too short trap. Right. This is for the scenes 00:05:50.129 --> 00:05:53.129 after you've made the grid. Exactly. If you don't 00:05:53.129 --> 00:05:56.269 detail the dirty green uniform or the old map 00:05:56.269 --> 00:05:58.769 in the character's hands. The AI will just randomly 00:05:58.769 --> 00:06:00.769 change the look. You have to keep giving it that 00:06:00.769 --> 00:06:03.230 context. Yeah, I still wrestle with prompt drift 00:06:03.230 --> 00:06:06.189 myself, especially when I'm shifting environments. 00:06:06.709 --> 00:06:09.290 You have to treat that initial prompt like a 00:06:09.290 --> 00:06:11.490 detailed map for the AI to follow every single 00:06:11.490 --> 00:06:13.750 time. That's a great way to put it. And another 00:06:13.750 --> 00:06:16.209 huge mistake is changing the lighting. Lighting 00:06:16.209 --> 00:06:19.050 is the glue that creates continuity. So if you 00:06:19.050 --> 00:06:21.980 start with soft fog, And then switch to harsh 00:06:21.980 --> 00:06:25.540 midday sun. The colors will break. The film will 00:06:25.540 --> 00:06:28.579 feel disjointed. Completely. And the simple technical 00:06:28.579 --> 00:06:32.339 error, forgetting no text or no borders. We mentioned 00:06:32.339 --> 00:06:35.160 this, but it's so important. The AI is trained 00:06:35.160 --> 00:06:37.480 on labeled data, so it tries to sneak in random 00:06:37.480 --> 00:06:39.839 words. And if your initial grid has any text 00:06:39.839 --> 00:06:42.300 on it, the source says to just scrap it. Generate 00:06:42.300 --> 00:06:44.790 a new, clean one. So what does this all mean, 00:06:45.029 --> 00:06:47.050 really? Beyond just the creative side, there 00:06:47.050 --> 00:06:50.329 are real monetary applications here. Absolutely. 00:06:50.910 --> 00:06:53.009 When you achieve character consistency, this 00:06:53.009 --> 00:06:55.670 becomes a viable skill. You can create ads for 00:06:55.670 --> 00:06:58.310 businesses. You could start YouTube story channels. 00:06:58.430 --> 00:07:00.889 Or sell the character concept art to game developers. 00:07:00.970 --> 00:07:03.110 Right. They need this stuff. Does this workflow 00:07:03.110 --> 00:07:06.290 require, like, a specialized gaming or rendering 00:07:06.290 --> 00:07:09.339 computer to even get started? Not usually. Most 00:07:09.339 --> 00:07:11.660 of the complex work happens online, though a 00:07:11.660 --> 00:07:14.899 good GPU definitely speeds up that Topaz upscaling 00:07:14.899 --> 00:07:17.660 part. So the big takeaway here is control. I 00:07:17.660 --> 00:07:20.079 mean, AI filmmaking can look like magic, but 00:07:20.079 --> 00:07:23.329 consistency is all about structure. The 3x3 grid 00:07:23.329 --> 00:07:26.410 shifts the whole process from just random generation 00:07:26.410 --> 00:07:29.230 to predictable, repeatable character design. 00:07:29.490 --> 00:07:32.730 We found the specific steps. Use that 1 .1 ratio, 00:07:33.269 --> 00:07:35.990 apply the five -part prompt formula, and then 00:07:35.990 --> 00:07:38.269 use the character reference feature to teach 00:07:38.269 --> 00:07:40.769 the AI what your character looks like. And that's 00:07:40.769 --> 00:07:42.829 the whole game. What's fascinating to me is just 00:07:42.829 --> 00:07:45.689 how quickly this field is evolving. We're talking 00:07:45.689 --> 00:07:47.610 about automating one of the hardest parts of 00:07:47.610 --> 00:07:51.699 visual art. Whoa. Imagine scaling this core workflow 00:07:51.699 --> 00:07:54.860 to a billion queries. It's staggering to think 00:07:54.860 --> 00:07:57.000 about. Yeah. And the source material ends by 00:07:57.000 --> 00:07:58.620 asking, you know, what's next? Should it be adding 00:07:58.620 --> 00:08:01.279 voices and music? Which raises a really important 00:08:01.279 --> 00:08:04.040 question. If the visual consistency is locked 00:08:04.040 --> 00:08:07.120 down, how do we ensure the audio consistency? 00:08:07.319 --> 00:08:09.259 That's a whole other deep dive. Yeah. For now, 00:08:09.339 --> 00:08:11.300 we hope this look at character consistency helps 00:08:11.300 --> 00:08:13.680 you realize your imagination can come to life 00:08:13.680 --> 00:08:16.879 faster than ever before. Go out there, try it, 00:08:17.000 --> 00:08:19.129 make mistakes. and try again. Thanks for sharing 00:08:19.129 --> 00:08:20.810 your sources with us for this deep dive.