# I Think GPT-6-Astra Just Changed YouTube

Channel: Nick Saraev
Video: https://www.youtube.com/watch?v=MujaqZS9wtM
Duration: 21 min
Language: English
Words: 4797
Transcript page: https://viewrankai.com/tools/youtube-transcript/MujaqZS9wtM

---

[0:00] This entire video is going to feature me in an animated style. The reason I'm doing this is twofold. First, I genuinely believe this is the next level of creativity for creators and people on the internet like myself. Just like we used to wear different clothes or adorn ourselves with jewelry, so too is modifying the actual style of a piece of media going to be how we express

[0:24] ourselves in the future. Two, I wanted to show you guys how to do it. I'm going to open-source my approach to doing this in a very scalable way. I'll show you a consumer method that you can use first that's easily available and requires just an API call. And then afterwards, I'll show you guys a pipeline that you can use to optimize with GPT-6 Astra and other cutting-edge AI agents to allow yourselves to generate these more economically. I'm also going to give you everything that you need to do this, all the templates and the skills and everything down below, uh so you guys can just copy and

[0:55] paste this approach into your own business or your own media production pipeline, whatever the style is. So, how exactly does this work? Well, the actual flow all starts with a normal recorded video. In the past, you had to use complicated motion capture systems and stuff like that in order to achieve the stylistic transfers that I'm showing you. But

[1:15] today, you don't have to do any of that. You just record a video without a lot of noise in the background, ideally. Uh as you guys could see over here, I'm going to be feeding in a source video that looks something like this. My hair's going to be a little bit different. And then the background of the video is for the most part, um you know, just like a

[1:31] white wall. What we're going to do is somewhere between the original recording and the ultimate stylized look, we're going to provide a style image. And you can think of that style image as basically multiplying, okay, you know, my core recording times the style in order to get the output. We've been able to do this for a while now with, um images. And actually, one of the very first things that I worked on back in 2019 was how to apply like image style transfer uh to a variety of different things. You know, like abstract art, pictures of myself, and so on and so forth. Uh but right now is the first

[2:04] time we're actually able to do so meaningfully with video. In a nutshell, here's how it works. The core video is going to supply motion. We're then going to generate using GPT image 2.5 or another model a style reference based off of that initial image. That's very important. Needs to be the same character sort of just with

[2:23] some stylistic additions. And then we're basically just going to multiply the two, okay? Uh pass it through this Genjutsu model. And I'm going to show you some alternative models that you guys could optimize on your own hardware as well for cost purposes if you wanted to do this at scale, like 1 3 turbo for instance. And then basically you achieve this stylized clip. The way that I've set up the system, and there's probably a bunch of people in the comments that can come up with a simpler and easier way, is Astra basically acts as the

[2:50] production manager. Um I've tried using a variety of other models to do this. I find that you need to provide it a lot more like oversight if you want us to do a good job. So, I'm just going to be using Astra. But keep in mind maybe if you give it the exact SOP, you guys can do away with a cheaper

[3:05] model that might use fewer tokens. Essentially, what GPT-6 Astra is doing is it's going to plan, run, inspect, and that's important, as well as retry any failed gens because this technology is not 100% yet. So, the flow looks like this. We're going to feed in a piece of footage, okay? The core thing. We're then going to use a model known as

[3:25] Genjutsu to apply the style transfer. We're then going to stitch the outputs together. And then we're going to post-process them using a tool called Solero voice activity detection and then FFmpeg. We're going to do all this at a specific frame rate to, in my case, give it like an animated sort of style. And then at the end we review and we export. As mentioned, there's a fair bit of splitting going on. And so, really what's happening because of the current limitations of these models including Genjutsu is you can't just feed in the whole video in one go and be done with it. Uh the models just, you know, aren't

[3:58] that good yet. Maybe they will be in 3 to 6 months from now, but certainly not right now. And so, in order to achieve this, you need to take, you know, let's say a 5-minute video, and then you need to divide it into chunks. So, this might be kind of your first 20 seconds here. This might be your second 20 seconds here. This might be your your third 20

[4:15] seconds here, and so on and so forth. And the good news is you get the original audio. Um the your model, uh KGPT6 Astra, basically is just going to inspect every join to make sure there are no major discrepancies between the gens because, as you'll see, if you're feeding in multiple runs to this Genjutsu model, uh and I'm going to show you guys an alternative to that in a moment as well, uh you know, sometimes the the stylization is going to look a little bit different from person to person. Now, the first thing you have to do in order to make this work is you need some form of style transfer. And

[4:46] so, what I've done here is I've taken a single frame of my face, and I basically had GPT image 2.5 make a bunch of modifications to it. So, I'm going to run you through an actual practical 0 to 1 of me, like literally making the prompt and then generating the image in a moment, but first I just wanted to show you what all of these styles that you guys saw earlier started as. You could see that one was a 3D sort of painted style, and that's the one that I I chose just cuz I think it's pretty cool. This one here is a prestige hand-drawn anime. You could see the

[5:16] original image was this, and this is sort of the stylized one. Next one was that sculpted stop motion effect, which I think is super cool. Obviously, it kind of changed the fundamental nature of my sweater, but it is what it is. This one's the ink and graphic novel. We have oil painted animation, which I think it was really cool, too. Uh watercolor {slash} graphite layered

[5:36] paper relief. Premium feature film CG. Lean Claire animation. I don't actually know how to say that, but I think the style's really cool. It kind of reminds me of that Takayuki Terries movie. And then charcoal and chalk as well. Okay, So, we need to actually generate these. That's sort of step number one. So, what I'm going to do first is actually walk

[5:52] you guys through how to do this. Probably the simplest and easiest one-stop shop for all of the stuff is Higgs Field. These guys basically have just aggregated every image and video model known to man and stuck it all in one place. And so, that is what I'm using here, more for my own ease of use than anything else. I'm going to walk you guys through how to sign up and then set everything up um right now, too, so you can do the same thing. And what's really useful is they also host their Genjutsu model, which is the model that I'm using initially as part of this

[6:22] demo. Uh I'm going to run you guys through some other software that you could use at the end of the course, as well, but this is definitely probably the easiest way of doing so. So, they have some sort of discount up. I'm just going to click get my discount, and then um I'm going to go continue with email. Make sure to actually agree to those dang terms. Let's put in my email address. I was previously running a bunch of influencer campaigns with uh you know, user-generated avatars and stuff like that, marketing, you know, the skin care brands and so on and so forth. That's what all of those are. Um

[6:52] I don't think there's any way to remove that, or or maybe there are, and I'm just being silly. But, uh the way that this interface works is basically over here, you have all of your previous generations. And then right down over here is where you feed in the image that you want to modify. And so, I'm going to be using GPT image 2.5 sun burst. That's the model that I want. And then what I want to do is I just want to feed in a frame of my face, okay? Which I've actually already done over here, as you can see. I'm looking kind of silly. Um

[7:17] you can get this frame any way that you want. I mean, you could literally just take like a screenshot of any video that you've recorded or something like that in order to get it. But, basically you just need a frame, okay? And then afterwards, you convert it to uh some sort of style. So, I don't know. I think that like I want a cool 3D style. Let's say, convert this to a cool 3D Avante-Garde stylized you know, style, while making sure the man's features and everything else about

[7:44] him are recognizably him. Okay, so the idea is, you know, I want this to be recognizably me. What you'll find a lot of the time happens when you do stuff like this is they come out and then they're not recognizably you for whatever reason. You know, they like try and make you look prettier than you really are. I don't want to look prettier, man. I want to look like Nick Surive. So, let's give it a go. After a few moments, an output like this is going to show up. As you guys could see, it's stylized with really big eyes and I kind of look like

[8:10] a cartoon character. My hair looks pretty dope though, I'm not going to lie. Let's get some ideas. Um and now you have everything that you need in order to actually do the style transfer assuming you also have the core video. There's also this one here where I'm kind of blocky and I really like the blocky one actually. I think it looks really neat. So, I'm going to give that a try. So, what I'm going to do is click download in the bottom right hand corner and now I actually have this on my computer. So, I can continue with the next step. And now you need some sort of

[8:33] video of yourself. So, I have a 50-second clip that I recorded kind of just as a demo cuz I wanted to show you guys how you can do this on short clip. Obviously, short clips also keep costs low. But I'm going to run you guys through what that whole process looks like here. The only issue is I don't have my my headphones, I think. So, that's one thing I'm going to have to change, but that should be okay. The model to choose is video and then go to Higgs Field Kenjutsu. This is the thing that will allow it actually to transfer the motion. And then what we're going to

[8:59] want to do is we're going to want to feed in that video with the still that I showed you guys a moment ago. My case, that's right over here. I'm just going to feed that in and we do have to trim it to 30 seconds in order to use it just because what we're doing here is a test and I just want to show you guys

[9:14] something as quick as humanly possible. You just move it to where I start talking. There we go. And then after that, we're just going to add this here which should add the stylized image after it's all ready to go. You can see we now have that sort of 3D version. That's what I want to select and it's the same thing that I showed you guys earlier with like that cool 3D effect. Last thing we have to do is add a prompt. And so, the prompt I'm going to use here is transfer the exact same performance from the reference video, that's this, to the same man in the

[9:44] style reference image, that's that. Preserve his recognizable identity, face shape, curly hair, stubble, gray sweatshirt, and black microphone. Now, because I used this for a previous prompt, it says the reference image defines only the painterly 3D rendering style and character design. I'm just going to remove the painterly and I'll just say 3D rendering style. But, the video is authoritative for head movements, blinks, mouth shapes, lip articulation, shoulder movement, and so on and so on and so forth. So, you need So, yeah, I mean, I'm going to give you guys the whole prompt below. You don't need to read it all, but um we're basically just saying, "Hey, keep it as

[10:18] simple and as close as humanly possible to the original gen." One of the reasons I think this is so cool is because you can use another model to actually generate the video, and then you can stylistically flip it or style transfer the generated video with a generated image style. And so, obviously, I'm doing this here on my own footage. And the reason why is because I think this is genuinely like a step towards all creators having their own distinct

[10:43] visual style. And so on and so forth, but you don't have to, you know, just use your own footage for this. You guys can use whatever the heck you want. This is extremely powerful um for ads because a lot of the time human beings, according to where they live and their culture and their preferences and their markets, you know, we like different things. And so, I'm imagining a future where you basically just have one ad, you know, a video ad, and then you flip it such that it's like 20 or 30 different styles, split test all of these futures, determine which one delivers the larger return on investment, say it's like a

[11:14] like a 30-second ad or something like that, and then just put all your money in in in that basket uh instead of spreading them all out like most people do. And you know, somebody that's run a fair few number of ad campaigns myself now, I can see something like that working really well. And that's where the economics of video production start to actually make sense. You You right now, this costs uh what, 32.5 credits in order for me to do this on a 5-second run. I think mathematically, if you're to do this for like a 20-second run, it'd be like 130 credits or so. You

[11:41] know, 130 credits is not free, right? Like obviously you're spending money for it. But if you do this across, let's say, a number of assets, and then it helps you determine what one specific ad that has, I don't know, 1.5 to 2x the conversion potential because that style just happens to be in or it's much more visually appealing, uh, then you make your money back on that in the first few clicks that you got. And it's worth noting you can use the exact same approach that I talked about here, but then use Waifu 3.0. Uh, Waifu 3.0 is basically just another model that's very similar to Genjutsu and probably related

[12:11] to what they're doing underneath the hood, but um, it's just the core intelligence. And so, what I did here is actually ran a test of Waifu 3.0 where I fed in this reference image as style, and then this video over here as the actual motion. Then I said, "Transfer the exact performance from the reference video to the same manner as the style reference image." And you can see for the same length of time, rather than it taking, uh, over 32.5 tokens or credits, I think it was before, now we're down to nine at, uh, 720p at 16x9. And so, the quality is not the exact same, but it's

[12:39] still pretty good. Occasionally, a transfer will fail, and so this was one that failed with Higgsfield Genjutsu over here. Uh, when that happens, you have all your credits refunded, and then, you know, it's just because of usually some mathematical elaboration, not because you got censored or whatever, unless you were using copyrighted character, in which case you probably will get. Um, but then anyway, you can just click retry and it'll re- run it. By the way, I'm going to show you how you can do this not just manually, but also with GPT-6 Astra in a minute. Want to make sure to build out the logic so you guys know the steps

[13:07] that you have to follow in order to get it done. Okay, and you guys can see we've actually now finished it, and the results are pretty cool. He animated just like this. And the reason I'm making it is because I think this is a the Right. So, an important thing to note here is we just did this for 5 seconds. Um, aside from that though, it did pretty well. I mean, it captured virtually all of my features. It it like everything about my apartment room, and, uh I also had some pretty cool sexy lighting. What I did now, if you think about it, is I just showed you guys how

[13:32] to go from zero to an actual generated version of the output. Um that's really cool and pretty straightforward to do. But if we want to, you know, go beyond just generating like a 5-second or a 10-second or a 20-second stitch, we kind of have to level up and I could sit here doing it all manually or I could just

[13:48] give it to GPT-6 Astra to do it for me. So, how do you do this in Codex? It's fairly straightforward. Um all you need to do is, you know, open up the Codex app, which you guys can download from ChatGPT's website. It looks something like this. Then give it a folder to work in. In my case, I'm working in the example folder. And then just, you know, make sure that it's it's inside and can hear what you're saying. Um my model

[14:10] picker here is set to GPT-6 Astra High. Obviously, you guys can go as high as you possibly want. I'm not going to go much higher than that. Then, if you guys head over to MakerZero, which is the second link in the description, um it'll have links to full prompts with everything you need in order to stylize your videos, just like how I showed you guys a moment ago. The one that you guys are going to want is this Genjutsu prompt, uh assuming you guys want to do the Genjutsu workflow. And then from here, it'll say you are Astra, my video production operator. Help me turn a

[14:36] recording of myself into a polished animated or illustrated video using plugin and Genjutsu from the work of the available tools and so on and so on and so forth. So, all you have to do is just double click this and then go back into Codex and then paste it in. It's already reading the pasted requests. And what it's first asking for is for me to attach my video or to give it its accessible local file path. So, I have my minute-long video here. I'm just going to drag and drop it right in and I'll say here's the video. Now, it's going to use that to come up with that

[15:03] cool file path uh as discussed earlier. Once it's done, you then have a little snippet or a frame that it's been uh taken from. It defaults to paintly 3D animation, which is the style that I'm using for this video, but obviously you guys can do it however you want. Um I'm going to say yes, use that default just because I think that style is super dope. And now it's what it's going to do is check Higgs Field's available models, account status, and pricing before preparing the quote. It's then going to send the media to Higgs Field, and then the benefit is it's even going to ask you if you guys

[15:33] are open to spending X credits for X result. The reason why I put that in there is cuz I just wanted to make sure that there was a verification step before you continued. Now, if you don't have Higgs Field connected as a plugin, this is where you'd have to add it. The way to do that is really straightforward. Just head over to plugins up here, type Higgs Field, click the that puppy, and uh click the install plugin button. So, if you don't already have that in, what this will do is install the plugin for you. After you run that OAuth, it'll then verify what account you are part of. Uh in my case,

[16:02] that is my account. And now it's actually going ahead through the uh production range. So, here it's giving you some options. It's saying, "Do you want to do the full 60 seconds, a short test only, or specific range?" You know, what sort of resolution do you want? Do you want 1080p, 720p, 480p? Do you want to keep the original pauses, cut pauses longer than 0.8 seconds, 0.4 seconds, or a custom threshold? Uh what I'm going to say here is, you know, let's do the full

[16:27] 60 sec. We'll do 720p. Actually, I don't want this to take forever. Maybe we'll do 480p. No, we'll do 720p. And then I'll cut anything longer than um 0.5 seconds, okay? The idea there is, you know, it's just what you need to get one of those cool YouTube jump cuts, and that's obviously my my style. And it'll get back to you with information on how long the uh tests are going to take, credit-wise. Um it's going to run the painterly 3D still, just using Nano Banana Pro to save on credits. And then as mentioned, I mean, this is just one way I could do it. I'm using uh Genjutsu here, but you

[17:00] guys could also use Wand 3.0 if you wanted, and then cut that down in order of approximately three. So, it'd be a little under 100. If you wanted to do that, you go down to Wand 3.0 prompt for Higgs Field. Um so, I'm just going to copy that in, and then I'd paste that back into Codex, just like I did a moment ago. It's gone ahead and actually generated that style for me. As you guys could see here, style is a little bit different from the example that I

[17:22] provided earlier. Um obviously, every time you uh generate an image, it's going to be a little bit different. But, that's still pretty cool. Um so, now I'm going to say, "Yes, go ahead and do the whole thing. All 60 suck then join and surf." Now, worth noting, you could actually download a slightly dumber version of this 1 2.2 animate 14 bill and run it locally on your computer. Um and so, that's actually what I have done with uh this inference pipeline. Because I've tried to optimize this to run basically as quickly as humanly possible. So, while these big providers are obviously fantastic at getting you up and running

[17:56] and then showing you guys what is possible, you know, if you did want to serve like a dedicated pipeline and do this really quickly across all your videos, you can just direct an agent over to this link right over here, huggingface.co/wand-ai/wan2.2-animate-14-bill.

[18:10] And uh you guys can actually legitimately download this and then do whatever the heck you want with it. You can run it on other hardware as well. I've experimented with a variety of different ways of doing this. If you're trying to run it on your own, it could take a a fair amount of time as well. Just keep that in mind, but yeah, that is that is how you do it. Um it basically just allows you to animate and then replace characters. And you can do whatever sorts of optimizations that you

[18:31] want to it to to make it even faster. So, obviously, I don't want to make this super ridiculously technical, so I'm not going to include information on all of that here. But, um wanted to make sure that you guys understood that was an option as well. Okay, and first clip is in. This whole video is going to be

[18:44] animated just like this. And the reason I'm making it is because I think this is a glimpse into the future of what most media will look like. Not this exact style, obviously. I'm not going to be painted in this exact way. But, pretty soon, we're all going to have the choice to consume media in format and as you guys could see, there were some minor issues with like mouth movements and stuff like that. I find that this tends to happen when um you're using kind of more stylized artistic forms like this. You know, I'd say the one that I'm using right now is also quite stylized and artistic, but this

[19:17] one seems a little bit softer around the edges, and I think that's probably impacting its ability to like track my mouth. Um, also, I don't know, maybe I'm just not really moving my mouth around that much. Still pretty reasonable quality, I would say. So, I'm going to wait for it to continue and uh make the other two segments and then combine them all. So,

[19:33] that's one reason I wanted to do that. The second reason is I want to show you guys how to do it. Naturally, this was done using GPT 6 Astra, as well as a couple of cool Cool, cool, cool. So, hopefully it's clearer. Um, you know, this is now interpreted and inherited the style that I trained it on. And uh yeah, you guys could do something like this yourselves. You guys could upload this to YouTube. You guys could clip

[19:54] this in a variety of different formats. I genuinely think this is probably one of the simplest and easiest things that anybody could do right now to make themselves a fundamentally stand out. Although, obviously, there's some packaging things that I think people need to get right. Do you have any questions whatsoever? Just leave them down below. I actually get my ideas for future videos from comments from people

[20:11] just like yourself. Um, really excited at I think the next phase of, you know, human creativity in media. We're seeing some extremely cool breakthroughs in AI virtually every day, and I thought this would be a cool way to expand that conversation. [gasps] If you guys like this sort of thing and want to learn how to monetize your own media production automation business, AI service business, or anything that has to do with the technology that you saw today, definitely check out Maker School. It's my 90-day accountability program that shows you everything you need to do in order to land your first customer. By the end of that 90 days, if

[20:46] you have yet to acquire your first customer, I actually give you all your money back. So, it's 100% money back guaranteed. Anyhow, thank you again, and I'll catch everybody here in the next video.

---

About this transcript
Read from YouTube's own caption track and laid out by ViewRank AI (https://viewrankai.com).
ViewRank AI finds the videos already beating a creator's own average on Instagram, TikTok and YouTube Shorts, transcribes them from the audio itself in more than 60 languages, and turns what worked into new ideas and scripts.
Free transcript tools, no account needed: https://viewrankai.com/tools
How to read any video this way: https://viewrankai.com/llms.txt