WEBVTT 00:00:00.000 --> 00:00:02.120 I want you to just visualize the math for a second. 00:00:02.779 --> 00:00:05.360 You have a job, a standard architectural rendering 00:00:05.360 --> 00:00:10.300 job, and you are staring at 75 hours of waiting, 00:00:10.460 --> 00:00:14.820 75 hours of just fans spinning, your office getting 00:00:14.820 --> 00:00:18.780 hot. Yeah. Then you look at this new way and 00:00:18.780 --> 00:00:21.780 it's five minutes. Five minutes. You are taking 00:00:21.780 --> 00:00:25.660 a workflow that needs 850 individual expensive 00:00:25.660 --> 00:00:28.500 images and you are replacing it with just two. 00:00:28.839 --> 00:00:30.800 It sounds, I mean, honestly, it sounds like snake 00:00:30.800 --> 00:00:32.799 oil. It does. It sounds like one of those late 00:00:32.799 --> 00:00:35.159 night ads promising you can learn Spanish in 00:00:35.159 --> 00:00:37.619 your sleep. But we're not talking about a scam 00:00:37.619 --> 00:00:40.280 today. We are looking at a fundamental shift 00:00:40.280 --> 00:00:43.679 in the physics of how architecture gets visualized. 00:00:43.700 --> 00:00:46.240 Right. We're talking about the $500 mistake you 00:00:46.240 --> 00:00:48.520 might be making every single time you hit that 00:00:48.520 --> 00:00:51.159 render button. It is a really startling claim. 00:00:51.340 --> 00:00:54.700 But the numbers, they seem to back it up. Welcome 00:00:54.700 --> 00:00:57.320 to the Deep Dive. I'm your host, and today we 00:00:57.320 --> 00:01:00.420 are unpacking a fascinating and frankly a kind 00:01:00.420 --> 00:01:03.140 of aggressive piece by Max Anne. It's titled 00:01:03.140 --> 00:01:06.200 The $500 Mistake Why You Should Stop Sending 00:01:06.200 --> 00:01:09.430 Your Archviz to a Render Farm. We're going to 00:01:09.430 --> 00:01:12.390 get into the nuts and bolts of Kling AI 2 .6, 00:01:12.510 --> 00:01:15.030 this whole concept of the first and last frame 00:01:15.030 --> 00:01:18.049 workflow, and how some specific prompts, what 00:01:18.049 --> 00:01:20.810 Anne calls power prompts, are completely rewriting 00:01:20.810 --> 00:01:23.010 the rules. And look, usually when we talk design 00:01:23.010 --> 00:01:25.290 tech, it's about the art, right? The aesthetics. 00:01:25.549 --> 00:01:27.650 But today we really need to talk about the financial 00:01:27.650 --> 00:01:29.950 reality. Because if you're still doing things 00:01:29.950 --> 00:01:32.629 the old way, the way we've done them for 20 years, 00:01:32.849 --> 00:01:34.730 Anne suggests you might be leaving about $100 00:01:34.730 --> 00:01:37.109 ,000 a year on the table. That number just... 00:01:37.280 --> 00:01:39.439 It jumped out at me immediately. That's a salary. 00:01:39.560 --> 00:01:41.780 It's a whole employee. It's a whole person. But 00:01:41.780 --> 00:01:43.680 before we get to the cash, let's start with the 00:01:43.680 --> 00:01:46.700 pain. Ann calls it the progress bar nightmare. 00:01:47.079 --> 00:01:50.019 Oh, I felt that in my soul. If you work in architectural 00:01:50.019 --> 00:01:52.560 visualization or really any high -end 3D work, 00:01:52.700 --> 00:01:55.099 you know this pain viscerally. It's the time 00:01:55.099 --> 00:01:58.180 debt. Let's unpack that term, time debt. Because 00:01:58.180 --> 00:02:00.760 for someone who just sees the final pretty building, 00:02:00.959 --> 00:02:03.340 the process is invisible, what are we actually 00:02:03.340 --> 00:02:05.299 paying with here? It's the part of the job the 00:02:05.299 --> 00:02:08.860 client never, ever sees. Traditionally, if you 00:02:08.860 --> 00:02:11.919 want just a standard 10 -second animation, nothing 00:02:11.919 --> 00:02:15.020 fancy, just a simple walkthrough, at 30 frames 00:02:15.020 --> 00:02:19.400 per second, you need 300 individual images. 300? 00:02:19.919 --> 00:02:22.240 And in high -end archviz, we're not talking about, 00:02:22.280 --> 00:02:24.599 like... Video game graphics. We're talking about 00:02:24.599 --> 00:02:27.039 simulating light bouncing off velvet refraction 00:02:27.039 --> 00:02:29.680 through glass. The details. All the tiny details. 00:02:29.960 --> 00:02:33.199 One single frame can take 15 minutes to render 00:02:33.199 --> 00:02:36.400 on a powerful machine. So if I'm doing the math 00:02:36.400 --> 00:02:39.300 on that, 300 frames times 15 minutes, that's 00:02:39.300 --> 00:02:41.599 not a lunch break. No. You're looking at anywhere 00:02:41.599 --> 00:02:45.639 from 23 to 75 hours. That's three days. Yeah. 00:02:45.919 --> 00:02:48.400 Just babysitting a computer. That describes it 00:02:48.400 --> 00:02:50.280 perfectly. He says you stop being a designer 00:02:50.280 --> 00:02:53.759 and you start being a highly paid computer technician 00:02:53.759 --> 00:02:56.439 waiting for a loading screen. And here's the 00:02:56.439 --> 00:02:58.960 kicker. The thing that drives professionals crazy. 00:02:59.460 --> 00:03:02.580 What happens if the client calls you on hour 00:03:02.580 --> 00:03:05.680 60 and says, hey, can we move that chair three 00:03:05.680 --> 00:03:07.960 inches to the left? Oh, no, you have to start 00:03:07.960 --> 00:03:11.159 over. You start the 75 hour clock all over again. 00:03:11.280 --> 00:03:13.240 And that's where the money burns. Because it's 00:03:13.240 --> 00:03:15.699 not just your time. You're paying for the computing 00:03:15.699 --> 00:03:17.900 power. This is the part I think people outside 00:03:17.900 --> 00:03:19.960 the industry just don't get. You don't just have 00:03:19.960 --> 00:03:21.599 the computer power for this. You usually have 00:03:21.599 --> 00:03:24.439 to rent it. Right. From a render farm, these 00:03:24.439 --> 00:03:27.219 are massive data centers, huge racks of servers 00:03:27.219 --> 00:03:29.740 that you rent to do all the number crunching 00:03:29.740 --> 00:03:32.939 for you. And they charge per core hour. Can you 00:03:32.939 --> 00:03:35.180 define that for the listener who hasn't seen 00:03:35.180 --> 00:03:37.620 an invoice for one of these? Sure. A core hour 00:03:37.620 --> 00:03:41.120 is basically the cost to use one processor core 00:03:41.120 --> 00:03:43.810 for one hour. It's usually between, say, $0 .02 00:03:43.810 --> 00:03:46.909 and $0 .10, which sounds cheap, right? Just pennies. 00:03:47.150 --> 00:03:51.349 But for a complex scene with 850 frames, you're 00:03:51.349 --> 00:03:54.129 looking at a bill of over $500. For one version. 00:03:54.330 --> 00:03:56.289 For one version. If you do three revisions for 00:03:56.289 --> 00:04:01.500 the client, that's $1 ,500. It is just wild that 00:04:01.500 --> 00:04:03.879 this has been accepted as the cost of doing business. 00:04:04.120 --> 00:04:06.539 It raises a really big question for me. I mean, 00:04:06.580 --> 00:04:08.620 architects, designers, they're obsessed with 00:04:08.620 --> 00:04:11.479 efficiency. Why have smart professionals accepted 00:04:11.479 --> 00:04:14.680 this financial bleed as just normal for so long? 00:04:14.900 --> 00:04:17.019 It really just comes down to necessity. It was 00:04:17.019 --> 00:04:19.459 the only way to get that quality until now. Normal 00:04:19.459 --> 00:04:21.819 didn't mean good. It just meant it was possible. 00:04:22.379 --> 00:04:24.680 If you wanted photorealism, you had to pay the 00:04:24.680 --> 00:04:27.870 tax. But that definition of possible has just 00:04:27.870 --> 00:04:30.230 shifted. Let's talk about the tool that changes 00:04:30.230 --> 00:04:34.009 the physics of this. Cling 2 .6. Released late 00:04:34.009 --> 00:04:37.069 2025. This is the big disruptive element here. 00:04:37.250 --> 00:04:38.970 Now, I have to play devil's advocate a little. 00:04:39.110 --> 00:04:41.670 We've seen AI video before. We've all seen the 00:04:41.670 --> 00:04:44.050 memes. Oh, yeah. The walls are breathing. The 00:04:44.050 --> 00:04:46.410 furniture turns into a dog. The camera feels 00:04:46.410 --> 00:04:48.310 like it's floating in soup. It's hallucinogenic. 00:04:49.319 --> 00:04:51.379 Why is this any different? So you're describing 00:04:51.379 --> 00:04:53.259 the single image problem. Yeah. In the older 00:04:53.259 --> 00:04:55.399 AI models, you give it a picture of a living 00:04:55.399 --> 00:04:58.720 room and say, animate this. The AI has to guess 00:04:58.720 --> 00:05:00.220 what the rest of the room looks like. It has 00:05:00.220 --> 00:05:02.439 to guess what's behind the sofa. And it's a terrible 00:05:02.439 --> 00:05:04.879 guesser. A terrible guesser, especially when 00:05:04.879 --> 00:05:07.800 it comes to strict geometry. It improvises. Yeah. 00:05:07.839 --> 00:05:10.459 And architects, they hate improvisation. They 00:05:10.459 --> 00:05:13.019 hate it. They want stability. If a column moves, 00:05:13.300 --> 00:05:15.870 the whole building falls down. So Kling 2 .6 00:05:15.870 --> 00:05:18.509 introduces this first and last frame control. 00:05:18.730 --> 00:05:20.810 This is the genius part. You don't just give 00:05:20.810 --> 00:05:23.230 it the start. You go back to your 3D software 00:05:23.230 --> 00:05:26.129 blender, 3's Max, whatever, and you render frame 00:05:26.129 --> 00:05:28.910 one. And then you render frame 300. The destination. 00:05:29.170 --> 00:05:31.509 Exactly. You give the AI the beginning and the 00:05:31.509 --> 00:05:34.649 end. Both are perfect, geometrically accurate 00:05:34.649 --> 00:05:37.610 renders from your software. And then you tell 00:05:37.610 --> 00:05:40.110 the AI, find the path. So it's not creating the 00:05:40.110 --> 00:05:42.149 room from scratch. No. It's just connecting the 00:05:42.149 --> 00:05:45.379 dots. Precisely. It's interpolation. It calculates 00:05:45.379 --> 00:05:47.720 the movement between two locked points. It literally 00:05:47.720 --> 00:05:49.819 can't make the wall warp because the wall is 00:05:49.819 --> 00:05:51.579 locked in place at the finish line. It has no 00:05:51.579 --> 00:05:53.959 choice but to stay straight. That is astounding. 00:05:54.420 --> 00:05:58.639 It's elegantly simple. It really is. And Kling 00:05:58.639 --> 00:06:02.300 2 .6 also added native audio, so it'll generate 00:06:02.300 --> 00:06:05.660 sound effects like ambient noise, footsteps in 00:06:05.660 --> 00:06:08.180 that same pass. But the visual stability is the 00:06:08.180 --> 00:06:12.279 headline. You are replacing... 850 frames of 00:06:12.279 --> 00:06:15.540 heavy rendering with just two. Two frames. That 00:06:15.540 --> 00:06:19.600 is a 99 % reduction in render time. Whoa, that's 00:06:19.600 --> 00:06:21.439 almost hard to wrap your head around. You're 00:06:21.439 --> 00:06:23.259 basically building a bridge and you only have 00:06:23.259 --> 00:06:25.439 to build the two pillars on the banks and the 00:06:25.439 --> 00:06:29.220 AI just manifests the bridge in between. That 00:06:29.220 --> 00:06:31.420 is a great analogy. And because you built the 00:06:31.420 --> 00:06:33.939 pillars, you know the bridge lands exactly where 00:06:33.939 --> 00:06:36.310 it's supposed to. So just to be crystal clear, 00:06:36.449 --> 00:06:38.889 the AI isn't hallucinating the building. It's 00:06:38.889 --> 00:06:41.629 strictly filling in the gap. Exactly. It's constrained 00:06:41.629 --> 00:06:44.970 logic. It forces stability by locking that destination. 00:06:45.329 --> 00:06:46.970 OK, so that's the theory. But I want to know 00:06:46.970 --> 00:06:48.970 how a professional actually does this, because 00:06:48.970 --> 00:06:51.410 I assume you can't just type make it cool and 00:06:51.410 --> 00:06:53.589 get a client ready video. There has to be some 00:06:53.589 --> 00:06:56.129 craft involved. Or absolutely. Amy is very clear 00:06:56.129 --> 00:06:58.810 about this. You can't just vibe your way through 00:06:58.810 --> 00:07:00.730 it. You need a workflow. So walk us through it. 00:07:00.730 --> 00:07:04.050 Phase one. Phase one is the 3D prep. You're still 00:07:04.050 --> 00:07:06.230 a designer. You set up your camera path and your 00:07:06.230 --> 00:07:09.449 software. But, and this is a massive constraint, 00:07:09.649 --> 00:07:12.449 the movement has to be intentional. Define intentional 00:07:12.449 --> 00:07:16.269 for me. No roller coasters, no crazy acrobatics, 00:07:16.310 --> 00:07:20.370 slow pushes, smooth pans. Why? Why can't I do 00:07:20.370 --> 00:07:24.620 a 360 spin? Because of object permanence. If 00:07:24.620 --> 00:07:27.639 the camera spins too fast, the AI loses track 00:07:27.639 --> 00:07:29.399 of what the objects are supposed to look like. 00:07:29.439 --> 00:07:31.579 It just gets confused. So you keep it cinematic. 00:07:31.579 --> 00:07:33.980 You export frame one and frame 300. Standard 00:07:33.980 --> 00:07:37.480 HD resolution is fine. Okay. And phase two? Phase 00:07:37.480 --> 00:07:40.079 two is the setup inside Kling. You select image 00:07:40.079 --> 00:07:43.259 to video and you enable that first and last frame 00:07:43.259 --> 00:07:46.079 toggle. And it suggests starting with five second 00:07:46.079 --> 00:07:48.199 clips just to get the hang of it. You know, walk 00:07:48.199 --> 00:07:50.660 before you run. Makes sense. Now phase three. 00:07:51.279 --> 00:07:53.100 This is the part that really interested me. The 00:07:53.100 --> 00:07:55.740 language we use. The power prompt. Yes. This 00:07:55.740 --> 00:07:57.480 is where you make or break the entire shot. And 00:07:57.480 --> 00:08:00.220 breaks it down into a formula. What is it? It's 00:08:00.220 --> 00:08:03.019 a four -part structure. You need movement, scene 00:08:03.019 --> 00:08:05.819 description, lighting, and technical rules. Okay. 00:08:05.860 --> 00:08:07.899 Give me an example of how a novice would screw 00:08:07.899 --> 00:08:11.120 this up versus how a pro does it. Okay. A bad 00:08:11.120 --> 00:08:14.819 prompt, a novice prompt, is just camera moves 00:08:14.819 --> 00:08:18.839 through room. Vague. Very vague. Very. The AI 00:08:18.839 --> 00:08:20.579 will do whatever it wants. Here's the expert 00:08:20.579 --> 00:08:23.079 version, and it uses slow -forward tracking shot 00:08:23.079 --> 00:08:25.459 through a modern minimalist living room, natural 00:08:25.459 --> 00:08:28.180 daylight streaming through windows, stable geometry, 00:08:28.480 --> 00:08:31.639 no morphing, cinematic photography. Stable geometry, 00:08:31.920 --> 00:08:34.519 no morphing. So you're explicitly telling it 00:08:34.519 --> 00:08:37.320 what not to do. That's the technical rules part 00:08:37.320 --> 00:08:39.419 of the formula. It's the safety rail. It's so 00:08:39.419 --> 00:08:41.320 interesting that we have to speak to the machine 00:08:41.320 --> 00:08:44.080 in its own language of constraints. We have to 00:08:44.080 --> 00:08:46.580 tell it, hey, don't hallucinate. We do. We have 00:08:46.580 --> 00:08:48.559 to remind it of the laws of physics. So looking 00:08:48.559 --> 00:08:51.159 at that whole formula, what is the one specific 00:08:51.159 --> 00:08:53.580 variable in there that prevents all the weirdness? 00:08:53.860 --> 00:08:55.500 It's definitely that technical rules section, 00:08:55.659 --> 00:08:58.080 explicitly commanding stable geometry and no 00:08:58.080 --> 00:09:00.720 morphing. Okay, we're back. We have the tool. 00:09:00.779 --> 00:09:03.340 We have the prompt. Now let's talk about selling 00:09:03.340 --> 00:09:06.259 this. Because technology is cool, but applications 00:09:06.259 --> 00:09:09.919 are what get invoices paid. And outlines a few 00:09:09.919 --> 00:09:12.379 of these magic applications. And these are honestly 00:09:12.379 --> 00:09:15.120 brilliant because they turn what could be a tech 00:09:15.120 --> 00:09:18.340 limitation into a feature. The first one he calls 00:09:18.340 --> 00:09:21.179 magic staging. This is huge for renovation concepts. 00:09:21.399 --> 00:09:24.899 You render an empty room as frame one. And then 00:09:24.899 --> 00:09:26.860 the fully furnished room as the last frame. And 00:09:26.860 --> 00:09:29.220 the AI just fills in the middle. Yes, but with 00:09:29.220 --> 00:09:32.090 nuance. The prompt is something like. Furniture 00:09:32.090 --> 00:09:34.669 and decorations should appear piece by piece, 00:09:34.850 --> 00:09:37.129 popping up one by one. So it actually creates 00:09:37.129 --> 00:09:39.649 a time -lapse effect. Exactly. A clean, smooth 00:09:39.649 --> 00:09:42.190 transition of a room furnishing itself. That 00:09:42.190 --> 00:09:44.409 used to take days of manual keyframing. Wow. 00:09:44.549 --> 00:09:47.309 Now, three minutes. That is a killer marketing 00:09:47.309 --> 00:09:49.309 asset. I can see that on Instagram immediately. 00:09:49.929 --> 00:09:52.230 It is. Another one is atmospheric lighting shifts. 00:09:52.509 --> 00:09:54.409 Oh, I like this. Day to night. Day to night. 00:09:54.710 --> 00:09:57.169 Frame one is bright daylight. The last frame 00:09:57.169 --> 00:10:00.110 is a moody evening with warm lamps. The prompt 00:10:00.110 --> 00:10:02.830 describes golden afternoon sunbeams crawling 00:10:02.830 --> 00:10:05.269 across the floor. Crawling across the floor. 00:10:05.429 --> 00:10:08.230 That's poetic. It is. And the AI understands 00:10:08.230 --> 00:10:11.190 that flow. Yeah. It animates the shadows lengthening, 00:10:11.190 --> 00:10:13.149 the lights flickering on. It's very emotional. 00:10:13.409 --> 00:10:16.149 And emotion sells architecture. Always. Now, 00:10:16.210 --> 00:10:17.850 I have to ask about the elephant in the room. 00:10:17.929 --> 00:10:20.750 People. 3D people usually look like zombies. 00:10:20.809 --> 00:10:22.730 They slide across the floor. Their eyes are dead. 00:10:23.610 --> 00:10:27.039 Does Kling fix the zombie problem? It helps. 00:10:27.200 --> 00:10:30.940 And calls this application humanity. How does 00:10:30.940 --> 00:10:33.940 it work? You don't use motion capture data, which 00:10:33.940 --> 00:10:35.840 is expensive and really hard to clean up. You 00:10:35.840 --> 00:10:38.840 just set frame A, person is at the door. Frame 00:10:38.840 --> 00:10:41.139 B, person is sitting on the sofa. And the AI 00:10:41.139 --> 00:10:43.679 figures out how they walked there. It fills in 00:10:43.679 --> 00:10:46.179 the natural movement. Walking, sitting, shifting 00:10:46.179 --> 00:10:48.940 their weight. No rigging required at all. He 00:10:48.940 --> 00:10:50.759 gives an example of a woman walking to a sofa 00:10:50.759 --> 00:10:53.580 with a cup of tea. The AI handles the subtle 00:10:53.580 --> 00:10:55.879 physics of holding the cup, the fabric of her 00:10:55.879 --> 00:11:00.120 shirt moving. But be honest, is it perfect? Or 00:11:00.120 --> 00:11:02.759 does she suddenly grow a third arm halfway there? 00:11:03.019 --> 00:11:06.039 It's not perfect. There's a catch. You have to 00:11:06.039 --> 00:11:08.519 keep the people secondary. If you make them the 00:11:08.519 --> 00:11:11.240 main focus, you zoom right in on their face, 00:11:11.419 --> 00:11:14.259 the flaws show up. The uncanny valley is still 00:11:14.259 --> 00:11:17.409 there. Okay. But it's background life. It's incredible. 00:11:17.590 --> 00:11:19.190 And then there's this falling furniture concept. 00:11:19.289 --> 00:11:21.269 Yeah, physics and stop motion. This is purely 00:11:21.269 --> 00:11:24.470 for viral reels. You prompt for a stop motion 00:11:24.470 --> 00:11:27.429 animation style. Furniture slides and skids into 00:11:27.429 --> 00:11:30.549 place. It's playful. It's not physically perfect, 00:11:30.629 --> 00:11:32.909 but it absolutely grabs attention on Instagram. 00:11:33.360 --> 00:11:35.299 It seems like the whole paradigm is shifting 00:11:35.299 --> 00:11:39.379 from perfect simulation to mood and flow. That's 00:11:39.379 --> 00:11:41.379 a great distinction. Yeah. So does this work 00:11:41.379 --> 00:11:44.399 for really complex interactions or is it just 00:11:44.399 --> 00:11:46.500 for mood setting? It's mostly mood and flow. 00:11:46.559 --> 00:11:48.460 You have to keep people secondary to the architecture, 00:11:48.559 --> 00:11:50.779 at least for now. Okay. So it's time for a reality 00:11:50.779 --> 00:11:52.759 check. We're painting a very rosy picture here. 00:11:52.820 --> 00:11:56.220 Five minutes, $500 saved. But is the render farm 00:11:56.220 --> 00:12:00.240 actually dead? Not entirely. And An is honest 00:12:00.240 --> 00:12:03.120 about this. He puts the tech at about 80 % ready. 00:12:03.259 --> 00:12:06.500 What's in that missing 20 %? Complex camera moves 00:12:06.500 --> 00:12:10.080 like spirals or really intricate loops or extreme 00:12:10.080 --> 00:12:13.659 close -ups on textures. If you need to show the 00:12:13.659 --> 00:12:16.200 specific weave of a fabric for a manufacturer 00:12:16.200 --> 00:12:20.259 or if you need absolute... dimensional accuracy 00:12:20.259 --> 00:12:22.759 for a legal submission. You can't have the wall 00:12:22.759 --> 00:12:25.539 wobble even a single millimeter. Right. For regulatory 00:12:25.539 --> 00:12:27.820 work or when you're in court proving a sight 00:12:27.820 --> 00:12:30.100 line, you stick to the traditional render farm. 00:12:30.460 --> 00:12:32.740 You need the physics engine, not the AI, I guess. 00:12:32.940 --> 00:12:35.519 But for everything else, concepts, marketing, 00:12:35.659 --> 00:12:38.419 social media, client buy -in. It's a no -brainer. 00:12:38.840 --> 00:12:41.139 And this is where we get back to the money, the 00:12:41.139 --> 00:12:42.759 money math. Let's break it down per project. 00:12:43.080 --> 00:12:45.860 Okay. A traditional 10 -second walkthrough costs 00:12:45.860 --> 00:12:49.850 about $2 ,500. That's the render farm fees, plus 00:12:49.850 --> 00:12:51.990 about 20 hours of your labor at your billing 00:12:51.990 --> 00:12:54.129 rate. Expensive. The AI -powered version, about 00:12:54.129 --> 00:12:58.230 $250, $10 in AI credits, and maybe two and a 00:12:58.230 --> 00:13:00.570 half hours of labor. So you're saving $2 ,250 00:13:00.570 --> 00:13:03.690 per project. Exactly. If you do four projects 00:13:03.690 --> 00:13:06.710 a month, you are effectively finding $100 ,000 00:13:06.710 --> 00:13:09.509 a year in overhead. That is life -changing money 00:13:09.509 --> 00:13:11.940 for a small studio. That's the difference between 00:13:11.940 --> 00:13:14.259 profitability and bankruptcy. It really is. It's 00:13:14.259 --> 00:13:17.039 not just saving money. It's reclaiming your profit 00:13:17.039 --> 00:13:20.940 margin. I have to admit, though, listening to 00:13:20.940 --> 00:13:24.940 all this, there is something a little scary about 00:13:24.940 --> 00:13:27.860 it. I still wrestle with prompt drift myself 00:13:27.860 --> 00:13:31.480 or just the fear of the black box. When you render 00:13:31.480 --> 00:13:34.529 manually. You control every photon. You know 00:13:34.529 --> 00:13:37.570 exactly why a shadow falls where it does. When 00:13:37.570 --> 00:13:39.850 you write a prompt, you're trusting an algorithm. 00:13:40.110 --> 00:13:42.450 It feels like giving up the steering wheel. It's 00:13:42.450 --> 00:13:44.909 a loss of control, sure. But it's a gain in leverage. 00:13:44.950 --> 00:13:46.649 You're saying, I don't need to control the photon. 00:13:46.710 --> 00:13:49.149 I just need the result. So based on those numbers, 00:13:49.269 --> 00:13:52.450 is the render farm completely obsolete? Not entirely. 00:13:52.610 --> 00:13:54.850 It's obsolete for concepts. Yeah. But it's still 00:13:54.850 --> 00:13:57.149 essential for that legal precision. It's a hybrid 00:13:57.149 --> 00:13:59.490 world now. Got it. Okay. Now for the listeners 00:13:59.490 --> 00:14:01.389 who are thinking, okay, I'm in, but I want to 00:14:01.389 --> 00:14:03.769 do this at a high level. And outline some advanced 00:14:03.769 --> 00:14:06.309 techniques to really polish this stuff. He does. 00:14:06.450 --> 00:14:09.409 The first one is for length. If you try to generate, 00:14:09.429 --> 00:14:12.669 say, a 20 -second clip in one go, the AI will 00:14:12.669 --> 00:14:15.669 drift. It gets amnesia. It loses the plot. Right. 00:14:15.750 --> 00:14:19.190 So you use multi -segment animations. You generate 00:14:19.190 --> 00:14:21.940 a five -second clip. clip A, and you take the 00:14:21.940 --> 00:14:25.759 very last frame of clip A and you make that the 00:14:25.759 --> 00:14:27.580 start frame of clip B. So you're stitching them 00:14:27.580 --> 00:14:30.600 together like a relay race. Exactly. It prevents 00:14:30.600 --> 00:14:32.879 the drifting because you keep re -anchoring the 00:14:32.879 --> 00:14:35.139 reality every five seconds. That's really smart. 00:14:35.480 --> 00:14:37.700 Then there's the hybrid workflow. This is for 00:14:37.700 --> 00:14:40.559 when the AI gets you 90 % there, but there's 00:14:40.559 --> 00:14:42.279 a little glitch in the corner. Maybe a plant 00:14:42.279 --> 00:14:45.039 looks weird. You can export the frames, fix them 00:14:45.039 --> 00:14:48.039 in Photoshop, manually clean it up, and then 00:14:48.039 --> 00:14:50.070 feed them back in. So you're actually collaborating 00:14:50.070 --> 00:14:52.690 with the AI. You fix its mistakes. Precisely. 00:14:52.730 --> 00:14:56.529 And finally, style consistency. If you're doing 00:14:56.529 --> 00:14:58.730 a whole house, you don't want the kitchen to 00:14:58.730 --> 00:15:01.409 look like a moody film noir and the bedroom to 00:15:01.409 --> 00:15:03.269 look like a Pixar cartoon. That would be a little 00:15:03.269 --> 00:15:05.669 jarring. So you use style anchors. These are 00:15:05.669 --> 00:15:07.710 phrases you repeat in every single tromp for 00:15:07.710 --> 00:15:10.190 that project. Like what? Architectural digest 00:15:10.190 --> 00:15:14.169 photography style. Or Scandinavian minimalist 00:15:14.169 --> 00:15:17.590 aesthetic. You train the AI on the vibe. You 00:15:17.590 --> 00:15:20.519 do. And the last technical tip is upscaling. 00:15:20.620 --> 00:15:24.480 Cling outputs at 1080p. Most clients want 4K. 00:15:24.860 --> 00:15:27.200 So you run it through a tool like Topaz Video 00:15:27.200 --> 00:15:30.539 AI to upscale and sharpen it. It bridges that 00:15:30.539 --> 00:15:33.259 last gap to professional delivery. So if I'm 00:15:33.259 --> 00:15:35.240 understanding this right, the key to not letting 00:15:35.240 --> 00:15:38.120 the style drift is those style anchors repeating 00:15:38.120 --> 00:15:40.659 the exact same aesthetic keywords every single 00:15:40.659 --> 00:15:43.840 time. Yes. Consistency in your language equals 00:15:43.840 --> 00:15:46.980 consistency in the visuals. This has been really 00:15:46.980 --> 00:15:48.679 eye -opening. We are definitely looking at a 00:15:48.679 --> 00:15:50.919 hybrid future. We are. You know, we look back 00:15:50.919 --> 00:15:53.659 at early CGI in movies, I think the 90s, and 00:15:53.659 --> 00:15:56.659 it looks charming but primitive. The Scorpion 00:15:56.659 --> 00:15:58.960 King comes to mind. Exactly. We're in the Scorpion 00:15:58.960 --> 00:16:01.700 King era of AI video right now. It's impressive, 00:16:01.759 --> 00:16:03.580 but it's going to get so much better. But the 00:16:03.580 --> 00:16:06.259 point Ahn makes is so crucial. You can't wait 00:16:06.259 --> 00:16:08.600 for it to be perfect. No, because by the time 00:16:08.600 --> 00:16:11.039 it's perfect, everyone will use it. The competitive 00:16:11.039 --> 00:16:12.879 advantage belongs to the people who master it 00:16:12.879 --> 00:16:15.500 now, while it's still a little bit messy. The 00:16:15.500 --> 00:16:17.559 people who figure out how to reclaim that hundred 00:16:17.559 --> 00:16:20.779 grand a year. Exactly. Stop asking, can this 00:16:20.779 --> 00:16:23.600 replace rendering? And start asking, how can 00:16:23.600 --> 00:16:26.279 I use this to do things I couldn't do before? 00:16:26.480 --> 00:16:28.820 That's the real takeaway. Don't let the tool 00:16:28.820 --> 00:16:31.519 replace you. Let it multiply you. Couldn't have 00:16:31.519 --> 00:16:33.539 said it better myself. And here's a thought to 00:16:33.539 --> 00:16:36.080 leave you with. If the visualization becomes 00:16:36.080 --> 00:16:39.419 this fast and this cheap, does the design process 00:16:39.419 --> 00:16:42.519 itself change? If you can see the finished building 00:16:42.519 --> 00:16:45.179 in five minutes instead of three days, do you 00:16:45.179 --> 00:16:48.379 take more risks? Do you iterate more? Or do we 00:16:48.379 --> 00:16:50.220 just end up churning out more generic buildings 00:16:50.220 --> 00:16:53.559 but faster? That is the really exciting and kind 00:16:53.559 --> 00:16:55.519 of terrifying question, isn't it? Something to 00:16:55.519 --> 00:16:57.399 mull over. Thanks for diving in with us. Thanks 00:16:57.399 --> 00:16:58.700 for having me. We'll see you on the next one.