WEBVTT 00:00:00.000 --> 00:00:02.240 You've got five different browser tabs open right 00:00:02.240 --> 00:00:05.139 now. You're exporting footage from one AI tool 00:00:05.139 --> 00:00:07.320 to another. You just want to stitch a single 00:00:07.320 --> 00:00:09.919 video together. Yeah, it's a massive glitchy 00:00:09.919 --> 00:00:12.300 headache for modern creators. The final export 00:00:12.300 --> 00:00:15.179 usually feels completely broken and disjointed. 00:00:15.300 --> 00:00:17.600 Oh, completely broken. Today, we're exploring 00:00:17.600 --> 00:00:20.219 a totally new structural framework. Welcome to 00:00:20.219 --> 00:00:22.019 the deep dive, everyone. We're looking at the 00:00:22.019 --> 00:00:25.059 LTX desktop video application. And we're diving 00:00:25.059 --> 00:00:28.579 deep into the LTX 2 .3 model. Right. This is 00:00:28.579 --> 00:00:30.339 going to be a really fun one. We'll cover how 00:00:30.339 --> 00:00:33.460 it merges generation and editing natively. We'll 00:00:33.460 --> 00:00:35.380 see how developers smashed arbitrary hardware 00:00:35.380 --> 00:00:38.399 limits completely. And we'll explore some wild 00:00:38.399 --> 00:00:41.960 timeline native AI generation features. Finally, 00:00:41.979 --> 00:00:44.060 we'll ask what this actually means for human 00:00:44.060 --> 00:00:46.810 editors. This completely changes how you think 00:00:46.810 --> 00:00:48.710 about creating videos. You never actually have 00:00:48.710 --> 00:00:50.469 to leave the video editor. You aren't wasting 00:00:50.469 --> 00:00:52.950 hours managing random exported files anymore. 00:00:53.229 --> 00:00:55.810 So before the new workflow, we need some context. 00:00:56.070 --> 00:00:58.350 We have to understand the engine powering this 00:00:58.350 --> 00:01:00.990 system. Traditional editors usually just bolt 00:01:00.990 --> 00:01:04.290 AI onto old architecture. But LTX is built entirely 00:01:04.290 --> 00:01:07.129 around a multimodal model. Which is a massive 00:01:07.129 --> 00:01:09.769 fundamental shift in software design. Let's define 00:01:09.769 --> 00:01:12.459 what we mean by a multimodal model. It's an AI 00:01:12.459 --> 00:01:15.599 generating video, audio, and text simultaneously. 00:01:16.120 --> 00:01:18.239 That's a crucial definition for everything we 00:01:18.239 --> 00:01:21.719 discussed today. So LTX 2 .3 brings four major 00:01:21.719 --> 00:01:24.909 systemic upgrades. First, they rebuilt the visual 00:01:24.909 --> 00:01:27.590 autoencoder from scratch. Yeah, and that autoencoder 00:01:27.590 --> 00:01:30.489 controls the generated texture sharpness. Older 00:01:30.489 --> 00:01:32.709 models compressed things way too aggressively 00:01:32.709 --> 00:01:35.670 during generation. You lost crucial edge details 00:01:35.670 --> 00:01:38.209 during the decompression phase. The result was 00:01:38.209 --> 00:01:40.670 a notoriously muddy or blurry texture. Right. 00:01:40.750 --> 00:01:43.930 But this rebuilt version handles raw pixel data 00:01:43.930 --> 00:01:47.049 beautifully. You get incredibly clean, sharp 00:01:47.049 --> 00:01:49.810 object edges everywhere now. They also completely 00:01:49.810 --> 00:01:53.079 retrained the model's motion data. I'll admit, 00:01:53.180 --> 00:01:56.140 I still wrestle with prompt drift myself. It's 00:01:56.140 --> 00:01:57.840 so frustrating when models just freeze halfway 00:01:57.840 --> 00:02:00.019 through. That was genuinely the biggest early 00:02:00.019 --> 00:02:02.079 user complaint. The AI would just forget the 00:02:02.079 --> 00:02:03.760 physics of a scene. Yeah, it ruins the shot. 00:02:03.959 --> 00:02:06.760 But LTX 2 .3 keeps the generated movement completely 00:02:06.760 --> 00:02:09.180 natural. It understands how objects maintain 00:02:09.180 --> 00:02:11.840 momentum over time perfectly. The third upgrade 00:02:11.840 --> 00:02:15.080 focuses entirely on native vertical video. It 00:02:15.080 --> 00:02:18.530 supports 1080 by 1920. perfectly out of the box 00:02:18.530 --> 00:02:20.949 which is an absolute game changer for short form 00:02:20.949 --> 00:02:23.469 video creators and the fourth major upgrade brings 00:02:23.469 --> 00:02:27.189 a hi -fi gn vocoder Right. That new vocoder creates 00:02:27.189 --> 00:02:31.009 incredibly clean audio sync. It reconstructs 00:02:31.009 --> 00:02:33.330 audio waveforms to match micro movements perfectly. 00:02:33.629 --> 00:02:36.349 It successfully removes awkward silences and 00:02:36.349 --> 00:02:39.430 digital noise artifacts. Why does training natively 00:02:39.430 --> 00:02:41.770 on vertical video actually matter? Why not just 00:02:41.770 --> 00:02:44.509 crop standard horizontal landscape footage? Well, 00:02:44.590 --> 00:02:46.770 cropping horizontal video almost always ruin 00:02:46.770 --> 00:02:49.289 your composition. Native training helps the AI 00:02:49.289 --> 00:02:52.009 frame vertical subjects properly. The generated 00:02:52.009 --> 00:02:54.449 subjects fit the tall aspect ratio naturally. 00:02:54.759 --> 00:02:57.000 So native training prevents awkward cropping 00:02:57.000 --> 00:02:59.479 and frames perfectly. Exactly. Now, how do we 00:02:59.479 --> 00:03:02.219 get this running locally? LTX Desktop is a fully 00:03:02.219 --> 00:03:04.620 local open source editor. There are absolutely 00:03:04.620 --> 00:03:07.680 no subscriptions or per generation costs. You 00:03:07.680 --> 00:03:09.620 get complete and total privacy for your projects. 00:03:10.110 --> 00:03:13.009 That privacy aspect is absolutely huge for studio 00:03:13.009 --> 00:03:15.530 workflows. You literally download the installer 00:03:15.530 --> 00:03:17.969 straight to your machine. The installation process 00:03:17.969 --> 00:03:22.129 is smooth, but the file is massive. You're going 00:03:22.129 --> 00:03:24.969 to need about 70 to 150 gigabytes. Yeah, it has 00:03:24.969 --> 00:03:27.530 to download all the required models. And Windows 00:03:27.530 --> 00:03:29.889 users must remember to run as administrator. 00:03:30.250 --> 00:03:32.710 That simple step prevents the software freezing 00:03:32.710 --> 00:03:35.349 during setup. During that setup, you face a very 00:03:35.349 --> 00:03:38.750 interesting choice. You can use the LTX API for 00:03:38.750 --> 00:03:41.729 text encoding. Or you can download a math of 00:03:41.729 --> 00:03:44.409 local encoder instead. The text encoder basically 00:03:44.409 --> 00:03:47.229 translates your written text prompts. The API 00:03:47.229 --> 00:03:49.990 option is completely free for anyone. It saves 00:03:49.990 --> 00:03:52.669 you about 25 gigabytes of storage space. But 00:03:52.669 --> 00:03:55.050 the local encoder guarantees a fully offline 00:03:55.050 --> 00:03:57.560 workflow. Wait, let me push back on that specific 00:03:57.560 --> 00:03:59.860 choice. So it's totally private and secure on 00:03:59.860 --> 00:04:02.259 my machine, unless I really want to save local 00:04:02.259 --> 00:04:04.340 hard drive space. In that case, my text prompts 00:04:04.340 --> 00:04:06.599 leave my computer. Yeah, that's the exact technical 00:04:06.599 --> 00:04:08.919 trade -off you make here. The API sends just 00:04:08.919 --> 00:04:12.039 your text prompts to external servers. The actual 00:04:12.039 --> 00:04:14.199 video generation still happens entirely on your 00:04:14.199 --> 00:04:17.920 computer. But for total isolation, you must download 00:04:17.920 --> 00:04:20.759 that local encoder. Is there any difference in 00:04:20.759 --> 00:04:23.579 video quality between them? No, the final generated 00:04:23.579 --> 00:04:25.939 video quality remains exactly the same. It's 00:04:25.939 --> 00:04:28.319 purely a difference in local data routing and 00:04:28.319 --> 00:04:31.779 storage. So the API choice only impacts local 00:04:31.779 --> 00:04:36.060 storage routing. But there is a massive hardware 00:04:36.060 --> 00:04:38.579 roadblock sitting ahead. Right. The official 00:04:38.579 --> 00:04:41.959 requirement is 32 gigabytes of VRAM. Let's clarify 00:04:41.959 --> 00:04:45.060 that term really quickly for the listener. VRAM 00:04:45.060 --> 00:04:47.379 is computer memory strictly dedicated to processing 00:04:47.379 --> 00:04:49.759 graphics. That intense requirement basically 00:04:49.759 --> 00:04:52.480 demands an outrageously expensive card. You'd 00:04:52.480 --> 00:04:55.339 need something like an RTX 1590 to run it. Most 00:04:55.339 --> 00:04:57.379 normal people simply do not have that enterprise 00:04:57.379 --> 00:04:59.040 hardware. This is where the story gets really 00:04:59.040 --> 00:05:01.740 fascinating. The open source community completely 00:05:01.740 --> 00:05:04.439 revolted against this hardware limit. They literally 00:05:04.439 --> 00:05:07.300 used AI coding tools to remove it. They built 00:05:07.300 --> 00:05:10.209 alternate forks like 1GP almost instantly. They 00:05:10.209 --> 00:05:12.589 actually got it running on 12 gigabyte consumer 00:05:12.589 --> 00:05:15.730 cards. People are using standard 30 series gaming 00:05:15.730 --> 00:05:18.209 graphics cards now. And that happened within 00:05:18.209 --> 00:05:22.470 a single week of release. Whoa. Imagine scaling 00:05:22.470 --> 00:05:26.470 an enterprise level software wall down to a consumer 00:05:26.470 --> 00:05:30.610 GPU in just seven days. Two sec silence. It completely 00:05:30.610 --> 00:05:32.990 changes how we view software development timelines. 00:05:33.529 --> 00:05:35.810 It absolutely democratizes access to professional 00:05:35.810 --> 00:05:39.529 generative video tools. Mac optimization is also 00:05:39.529 --> 00:05:41.850 actively in progress right now. Apple Silicon 00:05:41.850 --> 00:05:44.170 users currently have to use the API connection, 00:05:44.430 --> 00:05:47.069 but native local support is coming very soon. 00:05:47.269 --> 00:05:50.970 Does bypassing the VRAM gate slow down render 00:05:50.970 --> 00:05:53.949 times? Yes. Generating these clips will definitely 00:05:53.949 --> 00:05:56.490 take much longer, but it completely democratizes 00:05:56.490 --> 00:05:58.889 the software for everyday users. You trade rendering 00:05:58.889 --> 00:06:01.089 speed for actual software accessibility. We're 00:06:01.089 --> 00:06:03.649 basically trading rendering speed for total democratization. 00:06:03.829 --> 00:06:06.329 Yeah. So we bypassed the massive hardware limits 00:06:06.329 --> 00:06:09.029 successfully. Now we're inside the actual video 00:06:09.029 --> 00:06:11.410 editor timeline interface. Let's look at where 00:06:11.410 --> 00:06:13.730 the paradigm shift actually happens. You usually 00:06:13.730 --> 00:06:16.689 start in the gen space for quick renders. You 00:06:16.689 --> 00:06:19.769 render your clips at lower resolutions like 540p. 00:06:20.089 --> 00:06:23.709 Then you just use the built -in 2x video upstaler. 00:06:23.810 --> 00:06:26.029 You also get all the standard video editing tools. 00:06:26.149 --> 00:06:28.750 You get color correction and auto letterbox formatting 00:06:28.750 --> 00:06:33.189 natively. But the new AI features are truly revolutionary 00:06:33.189 --> 00:06:35.879 here. The first wild feature is the Regenerate 00:06:35.879 --> 00:06:38.379 Shot tool. You just right -click a clip directly 00:06:38.379 --> 00:06:40.620 on your active timeline. It re -rolls the generation 00:06:40.620 --> 00:06:42.879 without leaving the active editor. You also get 00:06:42.879 --> 00:06:45.319 native image -to -video capabilities integrated. 00:06:45.439 --> 00:06:48.199 You literally just drag a static image onto the 00:06:48.199 --> 00:06:50.899 timeline. You add a prompt to create fluid, natural 00:06:50.899 --> 00:06:53.899 motion. You can even mix in external video footage 00:06:53.899 --> 00:06:56.220 files. You can easily bring in clips from clang 00:06:56.220 --> 00:06:58.240 or runway. They all live together seamlessly 00:06:58.240 --> 00:07:00.480 on this unified timeline. The third feature is 00:07:00.480 --> 00:07:02.449 called the Bridge Shots tool. It's currently 00:07:02.449 --> 00:07:05.329 powered by the Gemini AI system natively. It 00:07:05.329 --> 00:07:07.810 analyzes the last frame of your first clip. Then 00:07:07.810 --> 00:07:09.750 it analyzes the first frame of your next clip. 00:07:09.850 --> 00:07:11.949 It automatically generates the missing transition 00:07:11.949 --> 00:07:14.850 footage seamlessly between them. It literally 00:07:14.850 --> 00:07:17.629 fills the empty gap with completely new video. 00:07:17.970 --> 00:07:20.250 Right now, the version one frame matching is 00:07:20.250 --> 00:07:23.269 admittedly quite buggy. Finally, we have the 00:07:23.269 --> 00:07:25.990 native retake in painting feature here. Let's 00:07:25.990 --> 00:07:28.269 quickly define in painting so we're all on track. 00:07:28.509 --> 00:07:31.730 It means erasing a mistake. So AI redraws that 00:07:31.730 --> 00:07:35.290 spot. You regenerate just a tiny, isolated portion 00:07:35.290 --> 00:07:37.829 of the clip. The rest of your original video 00:07:37.829 --> 00:07:41.389 remains perfectly intact. There is a small, annoying 00:07:41.389 --> 00:07:44.569 UI scroll bug right now. It's like stacking Lego 00:07:44.569 --> 00:07:47.189 blocks of data on a timeline. You never leave 00:07:47.189 --> 00:07:49.930 the room to manufacture new bricks. Does BridgeShots 00:07:49.930 --> 00:07:52.449 understand complex lighting changes between clips? 00:07:52.959 --> 00:07:55.040 The current frame matching struggles hard with 00:07:55.040 --> 00:07:57.639 complex lighting shifts natively. You definitely 00:07:57.639 --> 00:07:59.759 have to guide it with very specific prompts. 00:08:00.019 --> 00:08:01.800 So current frame matching struggles with lighting 00:08:01.800 --> 00:08:04.660 without explicit guidance. Exactly. You know, 00:08:04.660 --> 00:08:06.579 it needs specific prompts. We're going to take 00:08:06.579 --> 00:08:09.139 a brief pause here. This deep dive is supported 00:08:09.139 --> 00:08:12.959 by AI Mastery AZ course. Are you ready to level 00:08:12.959 --> 00:08:16.019 up your AI skills? Join the AI Mastery community 00:08:16.019 --> 00:08:18.949 to unlock exclusive tutorials. You can master 00:08:18.949 --> 00:08:22.730 tools like LTX Desktop and more. Learn from experts 00:08:22.730 --> 00:08:24.430 and connect with thousands of professionals. 00:08:24.709 --> 00:08:28.389 Start your AI journey today with AI Mastery AZ. 00:08:28.769 --> 00:08:31.930 All right, we are back. So an AI can bridge shots 00:08:31.930 --> 00:08:34.990 natively on a timeline. It can erase visual mistakes 00:08:34.990 --> 00:08:37.690 with a single mouse click. What actually happens 00:08:37.690 --> 00:08:40.169 to the human behind the computer keyboard? This 00:08:40.169 --> 00:08:42.289 creates serious anxiety in many creator communities 00:08:42.289 --> 00:08:45.370 today. People naturally fear that AI will completely 00:08:45.370 --> 00:08:47.850 automate their jobs, but we really need to look 00:08:47.850 --> 00:08:51.009 at structural limitations here. Models like Kling 00:08:51.009 --> 00:08:53.830 and Veo generate incredibly short clips. They 00:08:53.830 --> 00:08:56.590 usually max out at 5 to 15 seconds total. Right, 00:08:56.629 --> 00:08:58.830 and they completely fail at multi -shot storytelling 00:08:58.830 --> 00:09:01.309 natively. AI lacks an inherent understanding 00:09:01.309 --> 00:09:03.769 of natural visual rhythm. It doesn't understand 00:09:03.769 --> 00:09:05.970 emotional pacing or deeper narrative structure. 00:09:06.309 --> 00:09:08.929 It generates visually impressive... but completely 00:09:08.929 --> 00:09:12.929 isolated standalone clips. Exactly. Imagine chaining 00:09:12.929 --> 00:09:14.750 these short clips together automatically without 00:09:14.750 --> 00:09:17.509 humans. The system simply doesn't understand 00:09:17.509 --> 00:09:20.929 the rhythm of cinematic cuts. It misses the underlying 00:09:20.929 --> 00:09:24.269 emotional flow of the broader scene. The human 00:09:24.269 --> 00:09:26.570 editor still completely controls the sequence 00:09:26.570 --> 00:09:29.990 vision. The AI merely refines and generates the 00:09:29.990 --> 00:09:33.389 raw video material. Even if an AI could perfectly 00:09:33.389 --> 00:09:37.539 chain scenes together seamlessly, It wouldn't 00:09:37.539 --> 00:09:39.500 understand the emotional heartbeat of a complex 00:09:39.500 --> 00:09:42.460 scene. It just doesn't know when a quiet moment 00:09:42.460 --> 00:09:44.830 should linger. Right. And that lingering moment 00:09:44.830 --> 00:09:47.889 requires pure human empathy. Editing is fundamentally 00:09:47.889 --> 00:09:50.549 about feeling the specific emotional weight. 00:09:50.710 --> 00:09:53.129 The software just provides much better tools 00:09:53.129 --> 00:09:55.289 for human editors. Does this mean the editor 00:09:55.289 --> 00:09:57.970 shifts from technician to director? Yes. Editors 00:09:57.970 --> 00:10:00.909 will spend less time managing tedious rendering 00:10:00.909 --> 00:10:03.830 files. They'll basically become high -level curators 00:10:03.830 --> 00:10:06.649 of these generated visual moments. The job becomes 00:10:06.649 --> 00:10:08.690 much more about high -level creative direction. 00:10:09.029 --> 00:10:11.169 Human editors are becoming creative curators 00:10:11.169 --> 00:10:13.820 of generated moments. back and synthesize the 00:10:13.820 --> 00:10:17.919 main takeaway today. LTX Desktop version 1 definitely 00:10:17.919 --> 00:10:20.799 has its annoying software bugs. The interface 00:10:20.799 --> 00:10:23.480 scroll glitches and hardware gates are quite 00:10:23.480 --> 00:10:26.559 frustrating. But the core structural concept 00:10:26.559 --> 00:10:29.480 of timeline continuity is revolutionary. They 00:10:29.480 --> 00:10:31.879 introduced an amazing underlying concept called 00:10:31.879 --> 00:10:35.259 thinking tokens natively. These tokens actively 00:10:35.259 --> 00:10:37.840 look at the entire sequence you've built. They 00:10:37.840 --> 00:10:40.240 maintain character and lighting consistency across 00:10:40.240 --> 00:10:43.220 multiple different cuts. This fundamentally changes 00:10:43.220 --> 00:10:46.039 how we approach multi -shot storytelling entirely. 00:10:46.419 --> 00:10:48.139 You should definitely experiment with this powerful 00:10:48.139 --> 00:10:51.019 software yourself. It's completely free and totally 00:10:51.019 --> 00:10:53.580 open source to download today. You literally 00:10:53.580 --> 00:10:56.000 have absolutely nothing to lose by testing it. 00:10:56.139 --> 00:10:58.639 If an open source community bypasses massive 00:10:58.639 --> 00:11:01.480 hardware gates weekly, will massive corporations 00:11:01.480 --> 00:11:04.000 actually dictate the future of creative software? 00:11:04.139 --> 00:11:06.519 Or will anonymous developers tinkering in free 00:11:06.519 --> 00:11:09.919 time lead us? You jump between five different 00:11:09.919 --> 00:11:12.639 browser tabs today to edit. Tomorrow, you might 00:11:12.639 --> 00:11:14.980 direct everything from a single unified timeline. 00:11:15.259 --> 00:11:18.039 Keep exploring, keep creating, and keep questioning 00:11:18.039 --> 00:11:18.940 the tools you use.