MONETIZACIÓN YOUTUBE 🚀 Cómo Usar IA Para GANAR Más En YouTube

[01:39] Welcome to a new broadcast on the Nichon Nauta channel. This time trying to finish the intensive course on generation content with artificial intelligence, specifically for YouTube or any

[01:52] type of platform on which they can upload long videos or videos in the style of YouTube Shorts or TikToks. This created, Just as we say, as we always do in the channel, with intelligence artificial. I'm going to start by reading the

[02:07] chat and then showing some here little things. all. "I'm actually early this time." Hello, hello. By the way, about your website, Since you're here, I wanted to tell you that I It continues to block by IP address. I don't understand

[02:22] exactly why, or I don't know if it's blockage or saturation or I don't know, because The others don't happen to me either. hosting, but yes, in that one. Therefore, The solution I'm going to apply is instead of try to shift gears starting from

[02:37] an automated script, which is like I always do it, I'm going to make the changes I always do it, I'm going to make the changes locally, make a backup and reinstall WordPress on the server. Perhaps that will solve our problem.

[02:51] Because normal browsing works. He The problem is when I start moving things around administration and all that, and then more since it involves updating titles, contents. update images, delete some, upload others, change the

[03:04] delete some, upload others, change the visual style and other things. Uh, they're a A lot of requests and I'll be making things more complicated for myself by doing making things more complicated for myself by doing directly everything, just like that

[03:16] from some scripts on my PC to update things on the server. I'd rather do everything locally and just have it be a request to the server, reset to starting from said backup. I mention it here also because perhaps someone is interested

[03:29] in the sense of sometimes how We solve problems we can't We solve problems we can't skipping over the problems. Basically, Unfortunately, technology is a very strange thing. Sometimes we don't understand the

[03:41] problem, but if we can skip it better. Gersonet says, "Hello, Rayón says, "Hello everyone." Gersonet says, "Yes, yes. Good afternoon. Runner Bastia Victoria He says, "Hey, are you still offering the

[03:54] pages?" Hey, I'm offering an exclusive deal. I think that I haven't offered any at the moment. But you can contact me at so you can tell me which one you have seen in the past or what is it that

[04:09] You need it, and I can make it for you. valid if you've seen it in the past because these last few days I don't think so I haven't mentioned any websites, But if you're interested, you can... contact. [email protected].

[04:22] Why don't I mention any offers to him? here? Because I don't remember which one it is. Because here we're going to see primarily YouTube channels, which In fact, maybe I will talk about that. They ask for any offers, well I can

[04:36] Okay, thank you very much. And a certain Nichonauta says [email protected]. So, hello everyone, thanks for be around here. And well, before To begin, simply these captures

[04:48] I forgot to show them, believe. I only mentioned to them that I did this, but it doesn't show screenshots like such. This is the 22nd of last month. June, when I was here, I think. They went with 10

[05:03] They went with 10 pods within Run Pod that I was renting or leasing 10 RTX Pro 4500, which I think are 16 or 24 GB of V RAM. I was

[05:20] renting to try to generate more fast, intelligent content artificial. And although in the future, Maybe a live stream or something. or maybe in this very one, We can show you how to use a

[05:34] service like this. Let me mention from the outset that it is Interesting, seems like a good idea, even that you spend a lot of money and almost You don't produce results. Unfortunately, paying for a

[05:48] Online services like these, first We have to put money in, of course. to understand things a little and rent an available GPU,

[06:00] because they're not unlimited either. Then for that one in question and for the Then for that one in question and for the hardware that they provide us with said GPU, try to install, configure and adapt everything we're going to use

[06:14] everything we're going to use already consuming several minutes and several already consuming several minutes and several cents or a few dollars, simply in the part of the configuration to then try now

[06:26] Yes, produce content. However, we can choose the GPU. not the other PC components, with the which sometimes comes with a version From Python, other times it comes with something else, sometimes

[06:41] with more RAM, sometimes with less. And for But that's not the biggest problem. He The biggest problem is the internet connection that has said service, that has said service, because it's possible that I have something on my PC.

[06:56] super fast internet, I mean, in my house, but since I'm renting a of these GPUs, when having to download models and who knows what else to then use them in here, It might work very slowly for me, because

[07:09] The internet for that service is very slow and the problem isn't just in the facility. in the upload and download of each one of the content we want to be doing. So that seems like a good idea

[07:23] until you spend $50 in a day and don't You produced practically nothing but wasted time and fatigue. Very good, I'm just mentioning that here. To the even better in this same live stream we see. In fact, it's more than that,

[07:41] Refresh to see the value updated on the money, but in one of updated on the money, but in one of The screenshots show that he had $7,

[07:54] in another one I had today I couldn't see it, It's three o'clock, and it's still the 22nd. Don't look at it. the time because, well, it can, in fact, here It shows us that they only passed as

[08:08] It shows us that they only passed as 6 or 7 minutes, but that's because the value isn't had updated, that is, hadn't given to refresh. Therefore, these 37, Well, it was several hours ago, but in the same day, we had it anyway

[08:23] same number of GPUs being rented and decreasing the money a lot. Very What are people talking about today? Hello hello. Come on Let's see what we were talking about on the day Let's see what we were talking about on the day Previous adaptations and optimizations

[08:37] of the scripts we have for uh different types of generation of content not only to be that we download or copy them and Let's not just execute it, but adapt it. that truly gets to what interests us,

[08:50] Because automation is one thing, and another is another The thing is to generate content in this case, that is truly useful. No All you need to do is create content because Yeah. Very well, and continuing with this, it makes I had shared this screenshot some time ago,

[09:05] But now I've added an extra piece. Uh, I'm doing this while I wait for a little more time. of people. It's my WhatsApp account. What happens if there is a river? I have a boat and only ones can go

[09:18] two on top. On one side I have a sheep and there's me. Tell me all the steps I need to take to pass all the sheep and me to on the other hand. There's only room for two in the boat, and there are only two of us.

[09:32] There's only room for two in the boat, and there are only two of us. Two, a sheep and me. That would suffice I'll get the sheep and me in the boat and go on the other side. TRUE? Let's see here the things he invents. The artificial intelligence, which of

[09:45] We know that intelligence sometimes It simulates very well, but it's not really an intelligence. Ah, this is a classic. Look, if only Two fit at once, you'll have to several trips,

[10:00] Okay? First you cross with a sheep and leave it on the other side. So, we're done, right? You return only, Then you cross with the other sheep, one

[10:14] imaginary sheep. But here's the trick. When you wear when you arrive with the second one sheep, you go back with the first one. Okay. We went back with the first one. Now you cross over again with the

[10:29] second sheep and you leave her with the first one. I don't know anymore. and you leave her with the first one. I don't know anymore. or even which side. Ah. Uh, finally You come back for yourself. I got into the boat, but I let myself go

[10:43] I crossed to the other side, but I let myself go there, so I have to go back for myself and Okay, all together and safe next to each other side. We know that artificial intelligence Sometimes he's not very intelligent, to say the least.

[10:59] Sometimes he does things surprisingly well, but others surprisingly bad. Now I said, "Oh, Interesting." And if I put it on, I'll do it. I can illustrate it with an image. to understand better, or at least it would be more

[11:14] errors. And we'll see that next one errors. And we'll see that next one part, but for now it doesn't make sense. active. Greetings. Yes, YouTube, uh, Unfortunately, that's how it is. Uh, that's what it's for.

[11:31] directly in the chat, it's always there The WhatsApp channel and link have been set. You join the WhatsApp channel, Nobody can send messages, only me, and Nobody can see anybody's number and I simply use it for

[11:44] notify us when we're going live, because otherwise sometimes not works. I'm trying to use Hermes with Quin 3.59, but it doesn't keep the context of the conversation. Hmm, it could be several things. either

[11:59] You have very little context window. assigned and in the same conversation he forgets to things, so to speak, or what you of the same Hermes agent. I don't know, it would be something also see in upcoming live streams

[12:13] We'll probably see that. Very well, last one Let's begin. In the same conversation I told him, create an image in which you draw each stage and what's on each side of it. This, by the way, is within the same

[12:29] WhatsApp channel. And let's look at the image. I'm in the boat with a sheep. I cross on the other side.

[12:41] I'm still here while I cross to the other side and to the same while I cross to the other side and to the same We have three or four sheep. Then I left the sheep on the other side and I'm going back on my own. I'm not there anymore.

[13:01] And now instead of being four, as in The previous cases now number five. Then I go back alone, There are only four sheep.

[13:14] There are only four sheep. Then I'll take another sheep. Here are five sheep. Then I'll come back for the missing one and the that I left doesn't appear. I'll take a sheep with me at the same time.

[13:30] I'm on the other side and again they are four sheep. And besides, looking at the text, I'm wearing the last one. sheep on the other side.

[13:46] time. Artificial intelligence sometimes with Very simple things are incredibly bad, but also sometimes with very good things complex is incredibly good. Perfect. Let's begin then with

[14:00] today's events. After having done I've been here a while, after not having understood absolutely nothing of GPT chat and its reasoning for a very simple query. Rpod is quite Okay, but it's crazy unless you...

[14:12] when necessary. I'm already broke. a few hundred euros and I only use it, eh per serverless request. Yes, and also Another thing, it can be done right there that you rent some kind of hard drive virtual in which you already have

[14:27] Everything is set up so that next time that you want to run something, you don't have to reinstall everything. But that's not positive either. But that's not positive either. necessarily because you would be paying

[14:40] to always have that virtual hard drive which is cheaper than a GPU, but You might need it, and anyway, there is one. many problems. In other words, it has advantages. but also disadvantages, it's not the solution. Okay, very good.

[14:53] Let's go then. They're already on the channel published as much as the live stream we did previously,

[15:07] both tutorials or parts of this kind of tutorials or parts of this kind of previous courses we have taken. We have a regular YouTube Shores video. Conia, the real method for growth and

[15:20] monetizing from scratch and the continuation which was uploaded 17 hours ago. AI automation creates massive shorts without relying on GPT chat. That's all. the tutorial, the entire understanding of how and what things we need. Here

[15:36] We begin with the automation of each separate thing that we configure and we used previously. We have the same thing in live broadcasts, This is the live stream from here and this one from Here are the previous ones. And today we're going to...

[15:53] with the following part. What is the What's the difference between a live stream and a video? remove any parts that might come loose little of the topic, going only to the important. Anyway, in both cases, if they enter

[16:11] the entire index of each of the parts to understand absolutely everything. earring? Well, despite having seen a good amount of information, a lot of things and have already generated even videos There is also something automated.

[16:26] to keep in mind that we have mentioned a lot, and that's where the most emphasis is placed we have done. In virtually all of those minutes, in all those hours of duration From these explanations, we have not seen almost nothing, or rather, nothing at all regarding

[16:41] a code. We haven't modified any code, no. We have created new codes, we have not I've spent more than 5 minutes talking about code because it truly important It's in what we want to ask of you

[16:54] AI, in what we want it to be deliver as a result. And the code The only thing it's good for is automate these requests and answers that we will get from the artificial intelligence. But it

[17:09] That's precisely what's important, what What is requested is what is received. The Automation can generate a lot trash or a lot of things that could actually be useful of something. So, since we're talking

[17:23] A lot of all that, but barely as a kind of introduction, now in this What we're going to do is... adaptations. Let's see a little more Regarding codes, regarding other models we can use and everything

[17:37] models we can use and everything That, but not exactly with the objective to improve something or with the not with the objective of correcting something, but with the not with the objective of correcting something, but with the objective to adapt if we require

[17:49] adapt something. According to what? to our preferences, to our needs, to the type of content that we want to generate, to the type of audience, to type of strategy we want to follow, both content generation and

[18:05] audience building and receiving traffic as a monetization method, because It's nice to click "run," but The goal of all that is to reach a a very important part that would be monetize, right?

[18:19] And that, while it is the last step, We want to reach that final step. Clear that we are going to invest a good part of the time in all of the above, the monetization is only the result, but Well, that needs to be invested correctly.

[18:34] Previous time, there's no need to spend it. We have to try to reverse it, to do things theoretically in a way that can work. We've already seen a lot about prom, about codes, about models, But that's like a generic way of speaking, and

[18:48] standard. Let's see what things we can personalize, typically why the we would like to customize and for what purpose It can be useful, in addition to how to do it. So, simply show how For example, even on this same channel, because

[19:02] YouTube shorts continue to be published. If we go to Analytics And if we go to, I don't know, income, for For example, let's start with the content.

[19:14] If we look at shorts, we see that from that I started publishing YouTube Shorts, which I believe was from the 18th In June, I started to see some traffic, There was more, then less, and then it goes up and then it goes down. And well, the usual, right? Here

[19:26] uh, it's not that they aren't published. videos from the past few days, but when we are publishing so many and So many videos, it doesn't even show us the little icon with all the correct numbering. In other words, the videos are simply

[19:38] Here, with eight a day, YouTube is... It's impossible that it shows so many here. It gets overloaded, but all the videos are fine. They're still up and we see that In about two weeks, maybe a little more, we have Generated in YouTube Shorts 84,200

[19:55] Generated in YouTube Shorts 84,200 visits. 84,200 visits only from YouTube Shorts. Someone here will be able to to think, "Well, but it's a channel "Already created that has subscribers." Okay. However, let's see where

[20:11] The traffic is coming. It doesn't come from subscribers, it doesn't come from them being Reviewing the user's channel, it comes from people who are browsing among the YouTube shorts. And this mainly because also in each rise of a

[20:25] YouTube short I disable the option to notify subscribers. Because? Because People commonly subscribe to this The channel is subscribed to by up to pure Live stream or long video, not on YouTube

[20:40] shorts, short videos on various themes similar, but a completely different approach different. So, This traffic here is new and different. It's not necessarily the same type of traffic, which is what we are seeing,

[20:55] "Organic growth" Regarding visits and other matters. on a channel It's new, because this is something else entirely. something new. Uh, and this can also...

[21:07] demonstrate that the same channel It can receive different types of traffic It can receive different types of traffic and from different places. Uh, let's go to search, for example, the Scottish Gamer,

[21:22] Well, the name clearly tells us that What's the channel about, right? It's a channel in The person who makes gameplay videos uploads them. videos mainly of playing video games and the same in their live streams when there is, because in this case there isn't.

[21:37] none published, has been deleted. Without However, if we go to YouTube Shorts and In fact, looking at it superficially, but more Even if we go to more popular ones, we we're going to realize that nothing here is a video game.

[21:50] Nothing here, absolutely nothing, is a video game. It's mainly about... historical, historical information of Wars and stuff, what do I know, eh

[22:04] mainly from the medieval period. AND not necessarily the traffic it has Here, then, is the same one who comes to see this. In fact, to such a degree that here We see 6.9K, 8.4K, 133K. And indeed, very interesting, here

[22:20] in a video that received a lot of traffic, That's not even interesting gameplay. Hey? And in shorts, well, we see a lot. more traffic than it has subscribers, right? Then I mention This is also because a misno

[22:34] necessarily trying to earn more money It means creating more channels. He himself you can get more out of it. Clearly not It is advisable to just leave all our efforts focused on a single thing. Because? Because either we win big or we win nothing.

[22:49] because he's dedicating everything to the same thing place. And from one day to the next, if we They'll shut down the channel if it gets hacked. we lose the password or whatever, to happen, because we can go from being generating $000 monthly to zero.

[23:06] It's good to try to get a lot out of it. to benefit from the same thing, but neither You have to exaggerate. Uh, with this what I'm trying to say no out of nowhere, I want YouTube channels shorts dividing up all my

[23:20] efforts, which would apparently be fine for Let's see what works, but maybe in So many things, so divided, no It is convenient. We can do better in the same place try different things. And here too

[23:32] To mention another thing, these results are They may consider them good or bad, what do I know? I, it depends. considering that it is barely the But this shows us something, people.

[23:44] we've been on this channel for a while now And it's not the first time that... Try to get traffic on this channel starting from YouTube Shorts. It does not mean I tried and I succeeded. first. If we're going to see each other from the

[23:58] In principle, I have tried many times Generate traffic with YouTube Shorts. Without However, there had practically never been correctly guessed the shape of generate content for this channel. Because? Because this channel

[24:15] It was mainly intended for traffic in this type of content. Almost all we do is live streams and Sometimes the videos are actually clips live performances or whatever. But then It was starting from scratch on YouTube Shorts and

[24:33] That starting from scratch has another detail. I had never automated before. or had never generated too much content

[24:49] specifically making YouTube shorts. I didn't know how it could work. I tried making short films from clips, of the live streams and the videos. I tried

[25:01] I tried many ways and couldn't get it. By that? Because we're back to the same thing. A The thing is being able to click "execute" and that Generate videos, files, whatever.

[25:14] Whether they work in this case is another matter. after several attempts we have managed to give with something that apparently works. No I don't know if it's what will work best, But at least it works.

[25:27] However, well, besides here that There was a lot of wasted time on my part This channel. Uh, also something else... take into account take into account And it's not about trying things randomly,

[25:41] It is tested in one way, the results and analyzing results is not not only see the visits, but what Things brought visitors, what patterns in There is something common in that content compared to The one who had no visitors, what topics

[25:55] They addressed these topics, what duration of the videos, etc. duration of the videos, etc. etc. By reviewing all of that, we can go fine-tuning it more and more until exactly what we are supposed to achieve

[26:10] that works for us. Here's another one to mention thing, the growth in terms of visits, in fact it seems to be growing There are no visits, but rather the visits daily ones are therefore almost the only ones videos that are uploaded that day. We see that

[26:24] It's not exactly that it's increasing daily visits are increasing, being an accumulation of those of the combined views of new videos with the previous ones. This hasn't happened yet. here. It hasn't happened yet, but the goal is

[26:40] Let it happen. Because? Because the content that is being published is content evergreen. Informational content evergreen that can always work. USB-C versus mini-USB because the keyboard Qwerty is not who knows what. Stne, the

[26:55] virus that destroyed who knows which one. The "I love you" virus, a Windows bug blah blah blah. Okay, it's content that the The goal is not for it to work this week, The goal is for it to always work, that YouTube keeps showing it. Now, that's

[27:09] the objective, not what has yet to be achieved Achieved. But this is not necessarily true. It's because we're doing something wrong, it's because we're just getting started That's how things are. And it's not just that we can keep fine-tuning

[27:22] the formula, but YouTube itself must to better understand this type of content to which people It is truly aimed at, because YouTube is the one doing it showing, the algorithm is the one that is showing.

[27:35] showing, practically everything, all the The traffic we receive is to whom it is showing these contents. But in a I start it, well, actually I always do it constancy and perseverance and Consistency is very important. No

[27:51] continue working in the same direction. Yeah We try one way today, another way tomorrow. Another one tomorrow, and another one tomorrow, well, it's going to be more It's complicated to try to figure out if The results will either come or they won't, because

[28:05] the same algorithm does not yet have the sufficient confidence in the contents of this style. on this channel. And also It's not just about trust, it's about... understanding, he who understands who is it really aimed at?

[28:20] Commonly in YouTube Shorts, when They are done reasonably well, In 24 hours they get something like 1000 visits. The goal is not just for them to stay those 1000 visits, but with the

[28:33] As time continues to increase, keep giving it time. additional impulses, each time trying additional impulses, each time trying of targeting, trying to show it to the target, the target audience, the best potential target audience. And that's it

[28:46] He gets it, right? Daily things different, it is achieved by following a strategy of why we assume that It can work and continue with that change, I won't know what things They work and the algorithm won't know

[29:02] who to show him things. So, This is fine. But the goal is not number of visits. The goal is that there are more and more of them, not because the videos They will necessarily be better made, but because the new videos are going to

[29:16] to have traffic in conjunction with the previous ones. There were going to be videos previous than a single video, it can be that a week is the one giving us more than 50% of daily traffic and that The video may have been published for 6 months.

[29:33] published one year ago. There will be others who They will generate zero visits, others that They will generate few visits per day, but What I mean to imply is that the not only works, but also everything mentioned above. And that's not the goal, it's just how it works.

[29:47] this. Hey, why are there so many YouTubers? over time they become bigger and bigger Less content? Because they get tired, because blah blah blah and because I don't even know anymore They need. Hey, can you upload a video? every 6 months, every 3 months, and still

[30:01] earn a lot of money. Clearly there We're talking about a different kind of audience and other types of content from YouTubers They are very famous and continue to generate traffic. traffic and revenue, but they continue being quite good, to such a degree that

[30:15] that makes practically no difference to them publish new content, because They won't even notice it in their wallet. because what is generating good The amount of traffic is all that

[30:28] content that can be followed always consuming. Okay, let's go always consuming. Okay, let's go See other things here.

[30:44] I could have been getting them from YouTube, I also. Uh, but for each type of theme you want a channel or you put everything in one and you delete what you don't Each channel should be focused on one thing, uh, because of how the algorithm works

[30:59] and because of how we humans function. I When I go into a video on the channel of Nate Gentile, it's not because necessarily talk about technology. but necessarily talk about technology. but because he's talking about technology.

[31:11] because he's talking about technology. When I go into the Bite channel It's because I like the way he talks about video games and in fact that could happen in that case which is done a lot as a personal brand

[31:25] that we would watch almost any content that person come up. But that happens primarily with influencers that show their faces and that we are looking at them, not the content. Without However, we commonly enter and we

[31:40] We subscribe to a channel for content Specifically, in addition to YouTube rewards or better understands content in specific. But that doesn't mean we should always make a new one. Because? Because if a

[31:56] The channel isn't working in one way, it We can reinvent ourselves, but after a time that it didn't work that way or that we believe can be generated better results. So the goal is

[32:10] is each channel with a specific format and with a specific theme. We can vary it from time to time, I don't know, but that wouldn't be the goal. AND

[32:24] AND If the goal is to change, it's because now It will again be in quotation marks like start from scratch, try another one type of content, but because the Previously we saw that it hasn't worked as

[32:36] We were hoping, not so much because we're going to to have all kinds of things in one place channel. Now is the time to... understand something. What is the channel's theme? With thematic focus is not limited to

[32:50] The topic itself is technology. but the way it is explained. It can be a channel for criticism, a a person who criticizes everything. That person can talk about technology, can talk about clothes, can talk about a

[33:06] influencer, but because her main theme The good thing is the criticism, not the technology. Exactly that, because for example Let's look at what I mentioned as something Let's look at what I mentioned as something as a byte,

[33:28] secondary. Or rather, let's go to Lord Draur. dangerous internet, the man with the most Bitcoin of the world, the dark side of the

[33:42] UFC Freedom. Elon Musk is manipulating the market, stole 46,000,000 rise. specific category of type content of specific information, It isn't. It's varied, but trying to to talk about sensationalist topics

[34:00] to talk about sensationalist topics and as if everything he criticized about it bad things. This could be an example of a channel that is apparently like a kind of reports, but what do you know about

[34:14] whoever is going to talk about will to do something negative, he's going to mention bad things about that. So, here the The theme is not crime, the theme

[34:26] It's not about cryptocurrencies, the topic isn't... technology, the theme is this is bad, This is bad, this other thing is bad. Also It has to be understood that way, but it must things. We'll see more here, but I'll continue.

[34:41] considered. It doesn't mean she's the only one. nor definitive. That growth of Of what? That each time it goes doing more. No, it doesn't happen in all of them. is doing more. No, it doesn't happen in all of them. is what should be attempted to be achieved and

[34:54] that when a channel could be said to It's working, it's already happening. Uh, no more get. It's not available here yet. But the goal is to achieve it. And from In fact, mainly everything that I It generates income, because that's what it is.

[35:10] in what is achieved, but not It is necessarily possible to achieve everything. Unfortunately, but yes, it is the aim. The objective. and that is a plausible goal, In other words, it's not impossible.

[35:25] It's quite feasible, But that doesn't mean we'll always get it right. on the first try. I've realized that if you search for something on YouTube at the beginning popular and maybe your videos aren't

[35:37] show even if they deal with the topic I want uh it seems that only It seems that the topic I want is only... Show it to the celebrities if you insist on It performs well and already shows you videos. interesting and related. That's why the

[35:52] interesting and related. That's why the People don't keep looking. No, not Nate Yile I have him blocked. I don't even want to see him. is a creator who It teaches, it only sells covertly. I completely agree. In fact,

[36:06] I completely agree. In fact, He did it with AI. Hey, I'm getting really scattered, but oh well. Uh, regarding SEO, yes, I mean, mainly, but it doesn't mean that YouTube only shows channels

[36:20] big, It's that he grabs onto something to show a content that he considers the best. Try to average it out, so to speak, and

[36:32] calculate an average rating for each one of the contents that appear. And one of the things you can to appear first is if the channel

[36:44] It's very big. If there is no content that necessarily stands out too much as individual content, that does go The most noteworthy aspect is probably who The most noteworthy aspect is probably who It's coming, what channel is it coming from? But if you

[36:59] The content is good, even if it's in a small channel, it's very easy to overtake a video from a very large channel. That's it same as in SEO on Google. Hey,

[37:15] 100% Which video is truly better? by a long shot, because there is something to be had. which one to grab and which one to show first. AND In this case, I would first show the one from a large canal. But if your video already exists Individually, a good percentage is seen.

[37:30] time, even though the channel is new even though the channel is new may appear in first position. But I just said something important. Yeah YouTube sees that people like it

[37:44] amount of time, so in a completely new channel, you're not going to long or practically nothing, so it's going to be Once you have some traffic, something from the community, if the video is

[37:58] truly outstanding compared to the others, as much as content that the people see it complete or more complete than the Furthermore, both in terms of CTR and how many clicks give, title, description, so that attack the keyword correctly

[38:13] With SEO, it will be a different story. But yes, I mean, It's one of the things that YouTube can Take it when there's nothing else to do, to what else to hold onto to organize them correctly. More is not the first thing. It The most important thing is not that the channel is

[38:28] great, but watch the full video. If Gref, if Rubius makes a video that is called CPP, But it's an hour-long video that... People only watch 3 minutes, even if it's just of a channel with 50 million subscribers,

[38:44] For example, it won't appear in the first place position. Because? If he's what I'm looking for, call. CPP. The video title says the same thing. Yes, but if they watch it for 3 minutes and the video It lasts an hour, it doesn't matter if it's from a

[38:57] It lasts an hour, it doesn't matter if it's from a large canal. However, if they don't watch the other videos either a lot, well, a small channel that Watch a little bit of the video, because of that I show less of the one from the big channel. So

[39:12] That's how it works. Here, for example, How much does this channel have? 1.57K. And it's above our... it's above their video of ours, that we are 16K.

[39:26] 16K. Let's look for something else Let's look for something else or call.

[39:42] can appear in first position, no subscribers. This can clearly It affects when there are no other things happening. important factors for determining that one piece of content is better, but this is not

[39:56] It is the definitive, neither the only nor the exclusive. and even less so on YouTube compared, for example, with in a website, because on YouTube each video It works very individually, it gives equals the number of subscribers of

[40:10] channel, that can help if Those subscribers really see it You'll get views on your videos, but not necessarily to keep showing it necessarily to keep showing it other people or whatever.

[40:27] The same account can have effects adverse? Better new Gmail accounts every time. Not necessarily no problem. Up to 100 per account, You can have up to 100 channels. Now, If you lose access to that Gmail, that's where

[40:41] Well, no, there shouldn't be big ones issues. Show the eh more famous. Ah, I see, I already read that, people. No, Anate. Lord R is a bait machine and Above all, he investigates, he doesn't steal content. or at least it's not blatant and above all

[40:56] It has a luxury addition, that's why He's successful and has a great way with words. He's just like you, and it was very difficult for me to get it out if You were one of the few who talked about AI and about quality. Now everyone is an SEO expert. AI company and sells you its consulting services. Fast

[41:09] They've gone too far with talking. And many are Doing that, unfortunately, and that's because It's typical and I don't like it precisely tile. In fact, it is before a type of content that I consumed a lot and I used to find it strange that

[41:24] transformed into a single thing, it's going to lose traffic, I thought. AI, Linux with AI,

[41:39] Linux with AI, AI, AI, RTX Spark, which is AI. I I said, "What?" Because? We already knew because that. Because she launched a paid course, a uh, well yes, a kind of course in the which is very expensive and tells you about it.

[41:56] That happens a lot, unfortunately. I don't know if it's in a mastermind group or somewhere else. I don't know if it's in a mastermind group or somewhere else. side,

[42:09] But yes, unfortunately it happens a lot. Well, to sum up, getting back to the point, uh, Each video on YouTube works a lot in individually,

[42:21] It stands out because YouTube takes into account other things. As I mentioned, as the Number of subscribers, no videos In percentage terms they see it as too much, nor They don't interact too much, nor do they continue browsing YouTube. Mm, well, but

[42:36] This one has 10 million subscribers and This one has 3,000. Ah, well, I'll show you the one from But that's secondary, it's not the reason. major. Anything that could help in a major. Anything that could help in a The beginning is for the initial impulse that

[42:50] The video can have as many views as possible. by those subscribers as of understanding for the algorithm of whether the Whether the video will work or not, we'll have to wait for the next one. from that, decide to continue showing it to more people. Or, in other words, a channel with 10

[43:06] millions of subscribers upload a video and receive 20,000 visits in one day, it's likely that those first days appear in that those first days appear in top positions in search results,

[43:21] because YouTube already has the initial signs to from those initial visits, but if the video is not consumed enough or whatever, it doesn't matter what the channel is. very large. If other content is better,

[43:36] He's going to show the other one. But that's where precisely where the interesting part is. It's not about our content being good or not. The goal is for it to be better. than that of others. This is not it The content is moderately bad, I don't know

[43:48] It shows that it doesn't work that way. Because if that Moderately bad is the best there is. Well, that's the one that's going to be shown. Now, What makes it? a good title, a Nice thumbnail, a good edition, huh? that it's healthy, educational content, no.

[44:04] Unfortunately, to show themselves in the algorithm, as long as it's not harmful, harmful, illegal, to show in the algorithm, what matters is the the algorithm, what matters is the time you keep people inside

[44:17] from YouTube. That's what the algorithm can do. consider a good video for the algorithm, to show it, not for the monetization. If your video is 30 seconds long and only

[44:32] If your video is 30 seconds long and only Come for 5 seconds, but from then on They keep browsing a lot of videos Similar topics, it doesn't matter It doesn't matter if your video is very short. that they hardly see it at all because

[44:46] It's not just about how many people watched your video, but how much longer will they continue sailing in but how much longer will they continue sailing in YouTube. If I owned YouTube and YouTube. If I owned YouTube and I see there's a 5-hour video

[44:59] that they watch for a full hour on average And they leave YouTube and I see there's another one video that lasts 30 seconds, that they see just 5 seconds, but then after that They spend 16 hours sailing between similar videos. Which one would interest me?

[45:14] similar videos. Which one would interest me? further? So, how do we show ourselves? Do you realize that? In other words, it's not the only thing nor The final decision, but we're going to give it a try It has some interesting things inside. right here. Let's see where it is

[45:27] in which it appears, because followed They change all this, content, They change all this, content, all. Here it is. I don't know either, but basically There is something called sales funnels,

[45:42] that everyone follows to promote something, attract a lot of people less and less, less and less, until that you manage to get them in all the way to the last step that interests you. For example, when They offer free courses

[45:56] AI programming, uh, let's do uh five live shows all this week on the that we are going to do a free course on AI programming. A lot of less, less and less, less and less. AND Those who reach the end, oh, and for

[46:10] That's right, before we begin with This last explanation of this course free last day, let's talk more about our master's program artificial intelligence that costs $2,000 and blah blah blah blah blah. Okay, that one

[46:24] It's the typical sales strategy, isn't it? a sales funnel. Well, here we have that sales funnel in which we What is the main interest? Time of reproduction. And this is not necessarily only

[46:39] given by our video itself, but by the next thing they're going to do. that YouTube shows us. The time of reproduction that originated from selected. The time for reproduction of the sources in the

[46:54] The total is shown in the graph of above. Okay. Here, as I understand it, this is the total playback time that was on my channel. Because? Well

[47:06] Others, right? But the part about... also matters. others. In other words, if after my content important. That's not all. When We're talking about one thing, there's no need to Keep in mind that it is not based solely on that thing. It's an average of everything. Hey,

[47:21] trying to to give more weight to what is apparently to give more weight to what is apparently more important for that average, but Well, everything has more or less importance. And here we see that one of the most

[47:33] important On YouTube, it's about getting time to reproduction. That's what makes YouTube more eh

[47:47] That's not necessarily why we're interested. due to the playback time and that YouTube shows it, but because more Sometimes there will also be ads in They watch it for an hour and it takes them 30 minutes. On average, it's better than a video.

[48:01] 10 out of 10 minutes that watch 3 minutes in average. It depends on many, many things. We're not dwelling on this too much, and we're not. Going to the content section of the day today. I'm going to talk a little more about Let's continue.

[48:20] Of the two, it is easier to work with AI? everything that everything else that encompasses. For example, for web pages

[48:33] You need to understand the best of hosting, of domains, of SSL certificates, of of domains, of SSL certificates, of WordPress, plugins, templates,

[48:45] from Google Analytics, from robots.txt, from Ads.txt, of all that kind of stuff. AI isn't the problem, but everything else...

[48:57] that surrounds it. And on YouTube it would seem less, but it's also a lot to understand the algorithm, that there is no way exclusive and definitive way to understand it and how it tries to work, no We will necessarily always guess correctly.

[49:11] which is also inherent to the algorithm school of digital courses. Yes that's how it is. Lord Rur has an academy to teach to write scripts. It costs $300. The type He studied that in Spain. Hey, uh,

[49:25] It's hot, guys. Hello hello. I leave you, teacher. You are my role model as a person water for dispersion. No, that's just how scattered I am. unfortunately. Very well, there we have He explained several things at random, perhaps.

[49:37] Something could have been understood. Continuing with this, then, in this new channel, because clearly there isn't one. a lot of previous content that YouTube a lot of previous content that YouTube Keep showing it and let him understand how to

[49:49] already be well positioned in the algorithm. But anyway, before That's what I wanted to mention, something else. We were talking a lot about the search engine. YouTube is not just a search engine, there is a a lot of traffic sources that can

[50:03] a lot of traffic sources that can reach our channel. For example, in In this case, the search engine gives practically equal. The search engine is only 2.5% of visits. So, our our content generation strategy

[50:18] It's also something you should try. to understand, what type of traffic, what type of audience, what type of source of The incoming revenue is attempting to attack. The incoming revenue is attempting to attack. because

[50:35] complete and maybe they interacted with the content and therefore the algorithm Yes, he just keeps showing it to people with more or less the same preferences.

[50:50] This content was good because people Search for this and the results will appear. compared to the other live streams because it was a trending topic at the time.

[51:06] and that when we were on the same live broadcast YouTube showed it in recommended videos, showed as related This content. So that's something to keep in mind also

[51:21] If we go in this way, we'll look for the video with more visits and we review. This video has generated a lot of traffic, high income and so on.

[51:42] here, primarily through YouTube searches and another part exploration functions, or that is, that they appear as related or Something like that, but probably starting from another video just like it. That too

[51:56] important. Uh, there are themes that They are apparently very complex, and oh, it's that I can't get out in the algorithm in YouTube Shorts and I can't appear either in YouTube searches because already It's very saturated. Okay, but we can

[52:08] to piggyback on the video that appears a lot. If we check around here somewhere, If we check around here somewhere, that's how it is.

[52:20] of the content in which this is suggested video, there is a video that gives us 9% of the visits, in which for a long time, we went out many times as recommended, as related

[52:34] when the person was watching this video. So if we can't show up in the algorithm starting from being positioned, if we can't appear in a YouTube short, if we can't appear In one way or another, there are others here

[52:50] YouTube gives us all the information, So all of this can be useful to us. quite. And this single video of 13 minutes, which are seen on average by 3 In one and a half minutes, it has generated almost 3,000

[53:04] In one and a half minutes, it has generated almost 3,000 pesos. 3,000 pesos, that is, $171.

[53:23] a 13-minute video that was uploaded several years ago and to this day which is not the time when most traffic is received. It seems to have a faster growth, but just as we say,

[53:36] that we mentioned snowball style, that what is new is not necessarily what is new. which will continue to give us more income. That is to say, That is to say, Let's see

[53:49] This video here is from June 7th. It's just under a month old. Let's look at others. 30 visits and this one is over a year old.

[54:13] 16 visits and has plenty of time, more 2 years old. We click on another one. 54 visits in the last 48 hours and

[54:27] It's over a year old. The same as we say. The main objective here It is therefore trying to follow some species of strategy to continue receiving traffic. And anyway, following a strategy is just because.

[54:40] try to do them in a way that apparently have senses. But no implement a strategy that strategy is working. Of all The strategy did indeed work. initial. What was the strategy?

[54:55] initial? create content that people I was looking in the search engine, evergreen informational content that

[55:07] people search. How to get a VPS and Free hosting with Google Cloud, how to get a free domain that works In 2026, how to connect a domain to your WordPress on a VPS using A records. How to program with AI 100% locally for free

[55:21] and limited in Visual Studio Code. Are things that people can search for and that Our video may appear to them, being a new channel or not, and that we will always be able to continue receiving some traffic. The traffic

[55:34] that we have on this channel and the revenue It's not because of the new content. Ready, to YouTube Shorts, which is what we're going to Let's see, and we'll move on to code, which I already know. I spent a long time on this. So, Here we are seeing this beginning, what do you think?

[55:48] As we say, it's the beginning, but because At least as a starting point, it's working. account. Doing things in a certain way and not wanting to change them It doesn't necessarily mean that in the future

[56:04] The algorithm will show it. We may be doing many things wrong. and that they never end up working. But if from the beginning we see that something There is a quick reaction, and not only a quick reaction. not rapid, but sustained over time,

[56:20] This is susceptible to further complications. while this continues to work. And indeed We can do something else here. It's because starting from the premise that a bral is a good value or No It doesn't matter because we're not Mr.

[56:34] and we are not someone else. Rather, from according to other things I have done, This is working better or worse for me. Based on the things I do, Which one is working best for me? And in For the last 28 days, shorts have

[56:49] For the last 28 days, shorts have given, for example, 185 subscribers. In the last 28 days, to those who already It shows, I think it's all together here, yes, The long videos have given me 247 subscribers and in the last 28 days

[57:06] The transmissions have given me 79 subscribers. I am not comparing against the threshold of other people, because I am not another doing things and how I am which one is working for me?

[57:19] better. And here, clearly, the YouTubers Shorts are not outperforming videos long. But in the last 28 days, barely more or less than half the time is when there is YouTube shorts, while videos

[57:34] There have been long ones for two or three years, which in about half a month we'll be achieved in two or three years with videos long, this is showing me that

[57:47] Here's something interesting, more It's interesting that even doing the live streams. For example, now, why Should I do live streams? Because I like it mainly, but also for other reasons thing, to build loyalty.

[57:59] If we see them more as if they were pure numbers, not people, is the goal. to retain these numbers. Those who They can come from anywhere, they can become More fans of this type of content, no with a 30-second video, but to

[58:15] starting from seeing the person talking and all that. Therefore, apparently, at least in subscribers, is where least It is available new,

[58:30] but where they are most maintained the same. We were saying on YouTube shorts, no se they necessarily maintain the same,

[58:42] but with shorts and with long videos We bring in people who then stay. always in long videos. And well, we have subscribers and blah blah blah blah blah and we've seen a lot of things in the previous live shows, such as

[58:54] For example, the main thing here is going to be, what thing? Playback time, Don't let too many people skip the video, the more people stay to look better and come in closer to or greater than the duration total video length, average duration

[59:08] The good part is the visits. Why is this bigger? Because when one A person watches a complete video and then swipes through it. until it starts to repeat itself. With the which, that a short YouTube video It will work excellently for those who

[59:21] the duration per visit longer than the total duration of the video. Even so, that's not going to happen, we won't We'll start with the first ones days, with the first 15 days, with the first month and with the first video. By

[59:35] that? Because YouTube didn't yet know who to show it to. This video. It was the first video that was published on this channel in that form to generate content.

[59:50] Initially he showed it, more or less. He understood, and he showed it more. Not at the moment He has shown it more. But what do I want? make out? than in the first days make out? than in the first days These results are going to be pure.

[1:00:05] They are going to be wrong in the sense of it that even YouTube doesn't understand yet to show it to what apparently could be the right person. And not even when it takes longer. By that? because it's an average of everything, of

[1:00:19] This initial step against the little that is new, which doesn't tell us much. Compared to this content, if there were It has just been published, probably the The results were better. Because? Because apparently those impulses that

[1:00:32] traffic has given us would have been better oriented to truly the type of traffic that this wants to consume content, but that in itself has been the first content uploaded in this way, that almost all the

[1:00:46] traffic was initially and that more than half of the people have stayed to watch it when YouTube isn't even out yet I understood the content, it's very good. AND that the average is to watch almost the video It's complete, it's very good. So that? for

[1:01:00] The first video will be released in 15 days. Stop at 15 days to be the first video. And even better this. Oh, it's just that she hasn't received anything anymore. new. Well, in the last 48 hours four visits. If we look at it that way in each one, because the total sum is in

[1:01:15] not only continue to provide traffic, but also former. Another thing to keep in mind here as well taking into account, Hey, just as there was growth here Crazy, there might still be some afterwards

[1:01:30] in some of the videos, not all, Not in all of them, but there will be videos that They're going to take off a lot, and if we have 100 videos, maybe 20 will be the ones that they're going to be giving us 80% of visits, to give an example.

[1:01:47] Okay, let's move on to automations. Two years ago I bought an SC 2 TB for 120 and now it's at 380. If prices have changed a lot with the Let's get to the important stuff. We've already talked a lot. of the algorithm, which is not so simple and

[1:02:01] Anyway, another thing, what I'm I'm speaking from my own experience and my assumptions as conclusions to from having experienced a lot of things, but I don't have the answer definitive and nobody has it, not even the

[1:02:13] creators of the algorithm. Because? because it is something so complex and evaluates so many things that we as simple people We mortals are not going to understand 100% of the patterns, all the little differences meticulous that made YouTube show

[1:02:27] meticulous that made YouTube show 100% one video instead of another. seemingly more predictable, but not We have the absolute answer. Just like

[1:02:40] that with a language model, neither the creators of the language model know who is going to answer because she analyzes so many things that even the one who created it doesn't know which one it is. to be the answer. Here we are in the same, but we know more or less the type

[1:02:53] of answers they can give, right? So, that's what's important. Now, after being here for exactly one hour and not having done any of the things we had What to do today, since I barely have anything to do Let's begin. What had we done in

[1:03:06] the live stream before? Well, we had already a set of scripts among the which we could generate and we did generate certain automated content. in In this case, doing practically everything

[1:03:21] locally, with the exception of generation of the voice, which in this case was with AI, We saw why we had it that way and why We want to change it. And this generates the YouTube shorts, in this case, starting only from images generated with

[1:03:36] artificial intelligence on our PC, but we don't create videos as such yet, or that is, intelligent video clips artificial, but they are only images artificial, but they are only images with transitions, with movements, eh of

[1:03:51] following. Let's remember a little bit about how were planet, right? It was a brief period, but a hegemony of 165 million years

[1:04:05] during the Mesozoic era. While human civilization is barely a breath in history, they shaped terrestrial life for almost the entire existence of the continents current. But the most surprising thing is

[1:04:20] that did not disappear completely. Each Every time you see a bird, you're seeing a small dinosaur that managed to survive in time. There we see that one of the minor inconveniences in this case are the voice, both for its quality and for the

[1:04:35] the way in which he is narrating. Quality in the sense, not that it is listen with noise or anything like that, but not pronounces 100% correctly, and that perhaps she's not very lively

[1:04:48] the voice to the type of content that We're doing it, it's kind of very relaxed. In this case, that model, the GPT TTS that we are using here, which is the best current model they have The people at Open are doing this online.

[1:05:03] It inherently has different tones of voice, although we can't use just any, only those that they have. It's not that we we can create one, it's not that we can voice clone. And another thing they have, no It's only the tones of voice that we can

[1:05:16] give them instructions on how we want them to be the narration, trying to make it be the narration, trying to make it more terrifying and suspenseful or whatever more terrifying and suspenseful or whatever more cheerful or more sarcastic.

[1:05:29] That's fine for that type of content. which interests us, but there's a detail which is precisely the one we had as mentioned before, and that is because This model is paid for and we don't It offers the best quality compared, because

[1:05:41] It offers the best quality compared, because For example, against 3 TTS, which is not only It's not just free, it's much better in quality, but also in versatility. Qen 3 TTS is small, fast and with

[1:05:54] Much better quality than this, too. for free. But it has something else, like As I mentioned, versatility. Can clone any voice from a an example of just a few seconds, a file MP3 of a voice sample to clone it or

[1:06:09] make it quite similar. You can design voices from a prompt and also You can give instructions at the moment of that he narrates it so that he narrates it with a certain specific tone, without changing the basic voice itself, but making it more

[1:06:25] sad, happier, that kind of thing. So let's change this set of scripts to use Qen 3 TTS, not because it's free, that's something Additionally, fortunately, But the main thing will be

[1:06:42] But the main thing will be in versatility and quality. Let's go to change everything we have here to that it be local and in this case what We were missing voices. Let's go use accounts as in the first video

[1:06:56] of this series of three videos. This is the third. Uh, let's use that account STS like in that first video, but now here in an automated way and it will simply improving by changing model to have better quality,

[1:07:10] making it so that we are also free, but something else entirely consistent over time. The GPT model TTTS, like the Gemini TTS, we have used for a long time and in We have both suffered from the same problem. EITHER

[1:07:24] that the service goes down or that no matter how much Pay today, you can no longer use it on Pay today, you can no longer use it on that account or worse, the model is He keeps calling the same thing, It doesn't give the same results.

[1:07:37] It doesn't give the same results. We updated the Yemini 2.5 TTS and now You'll be enjoying a higher quality human and blah blah blah blah blah without changing nothing in your code. That's not necessarily true. It's positive. Because? Because they are

[1:07:50] changing a formula, in quotation marks, winner that we already had from to make it work, because maybe we won't We realize that they updated it until we generated the voices of 100 videos, to name a few,

[1:08:06] until we reproduce it, we realize Realizing that it went wrong, we lost time, we lost money, we have to adapt everything again so that it looks more or less how we liked it to look

[1:08:18] And maybe we'll achieve it. or not, because the The model is up to date, He's going to understand things differently and he's going to to function differently. Maybe Yes, apparently it produces better quality. but perhaps they removed a tone of voice that

[1:08:32] It was the one I used. Maybe not now It tells a terrifying story well, but conversation, but I didn't want a conversation, I was telling stories terrifying. So, given that we They change the model from one day to the next,

[1:08:46] fortunately we have for use locally. So we have more quality, more versatility, consistency in time and to top it all off, for free and fast. Therefore, Qen 3 TTS, Qen 3 TTS. In the first video that

[1:09:01] We did about it on this channel, Regarding these topics, we were seeing how to install it, how to do it to function within Confi UI, which is something like an Oyama LM Studio or Yama CPP, but not for diffusion models

[1:09:15] of transformer architecture, that is, primarily generating models images, videos and audio. Therefore, Since Q3 TS is an audio generator, Well, it can work within Confi UI good. It is not the only way nor the

[1:09:29] definitive nor the best way to do it It will work, but it's a way that, given Confi UI for other things, because of one Once we put this in there so we don't have everything watered in different ways. I went we use to generate the images with

[1:09:43] Image Turbo. I went, we can use to generate videos like with guan. Well, let's use Fiui for that too. Generate the voice right there with Qen 3 TTS. So, as I said, I already know this. we had done

[1:09:59] And here I have this Config UI, in the which is where we had installed those things. If I go into any folder, because For example, in TTS, here I see that I have For example, in TTS, here I see that I have the Qen 3 TTS, version 1.7B.

[1:10:12] So, given that we had already done this installed since the first video, in the installed since the first video, in the Second, we automate, but without accounts. So now we're going to adapt the automation to certain details that

[1:10:25] First and foremost will be It will allow two things. We're going to open configure. The first step is to make it 100% in local and free and the second one we have

[1:10:38] higher quality. Uh, so given that Everything will always be local, daily I can click on run and daily I will be sure that the content will to end up in a similar way, right? That Oh, it's down, oh, they raised the price,

[1:10:51] Oh no, I can't use it anymore today, oops They updated it and it's not called that anymore. shape. Ah, they updated it, but I didn't... They changed their name. Even worse, because not It no longer works like it used to. So, putting this into local voice, it

[1:11:07] The next one will be, for example, ah, Well, at this moment it is Now we're going to make sure it's not like that. only images, that makes clips of video, and there's a whole world about that there. inside. But anyway, I've already opened Conf UI here

[1:11:22] and in this case despite having The execution has finished, if indeed it has. It finished loading, I don't think so. That's how it is. But even though it has finished loading, I don't think it's going to open in the browser.

[1:11:34] browser? This is specifically because I had changed one thing inside of all your files. I don't remember exactly where, because for certain I was interested in automations that would open and close automatically

[1:11:47] when I was making several uses with the GPU. So, if every time It would close and reopen, it would open for me a tab in the browser, I was going to End the browser with 100 tabs open ones that were of no use to me.

[1:12:00] So I just changed something, I don't know whether here or at point P which opens, I think it's in main.P, so I don't... Open it in your browser. The server It starts, but the tab doesn't appear in the browser. Even so, if I want to get in,

[1:12:13] So I press control and click or I type that directly into the browser And so it appears here. TRUE? So, within Confi UI, which in that configuration of both Confi UI and Qensts, here we will be able to

[1:12:27] We will use it, and in fact we will. as an initial introduction and then We're going to adapt the code to use this. So I have to find where this. So I have to find where That's where I specifically had

[1:12:44] the workflow that can be useful to us or that We were using for all this. I don't know exactly where I have the exact same one, but Well, I have a very similar one, so it's going to be almost the same.

[1:13:03] Good afternoon, Juan, it's a pleasure to greet you and community. It had been a while since I'd logged in. that you are using and if so, a brief summary. I just got in. Queen in local For advanced things, gem for writing, Z image Turbo for generation

[1:13:16] writing, Z image Turbo for generation of images. One oan video for some kind of video. This TS is for voice generation. sometimes R Code, sometimes Open Code, sometimes quen Code, sometimes cloud

[1:13:30] always with local models, but with paying online. Good afternoon. Hello. Hello. Great, we have Confi UI. to open our workflow, which in In this case we have this workflow, which is not

[1:13:45] practically the same as the one we had seen in past live streams, which allows us to start from select accounts

[1:14:00] a small fragment of voice, to be able to generate the narrative for us about what we write with that similar tone, that similar form of narrate. Which means it's not that we have

[1:14:16] Doing a fine tuning isn't like that We don't have to download a parrot, it's not that we must choose one of the shades of voice or voice models that come from default, but directly we We can put anything here. Of

[1:14:28] In fact, we could even clone more or except our voice. Now, the fragment The voice doesn't necessarily have to be very long, but what it must be is very representative, not only having the tone

[1:14:42] with which the tone of the person that He's talking, uh, but also the rhythm of the narration, how he pauses and others. If I say, "Hello, how are you?"

[1:14:54] And I use that as an example, because I'm not going to to expect that it will sound more natural afterwards if I spoke in a very robotic and slow manner. So, that should be very much the case here. similar. However, it is important to keep in mind that it's not just with a voice

[1:15:06] generated by a voice recorded by us, but it can be a voice generated, an example of voice generated with another voice synthesizer or with this one the same starting from having designed the voice prior or whatever. There are so many things that

[1:15:20] It can be done. And in fact, well, I... I already have an audio example here. Just keep in mind that we don't Let's plant their voice, let's not use the voice of another real person just because Yes, uh, not only because of legal problems,

[1:15:35] but for simple ethical reasons. Let's be people aware Thanks, Juan, it's a pleasure as always. Thank you, thank you. Okay, very good.

[1:15:49] file to upload and I'm going to select a small voice file that I have a small voice file that I have For example, which sounds like this. Great opportunities come to who knows how to take them, so if you get them

[1:16:03] One, you can be sure that It's really for you. Okay, that's the example we have and This is the model we are going to use, This is the model we are going to use, the Q3TS in its 1.7B version. We can use

[1:16:16] version 0.6B, but also version 1.7B It's very small. This could be executed on GPU, Nvidia mainly, but AMD or Intel, but even with just CPU. We already know that Confio has that

[1:16:30] Versatility, not as simple as it sounds, But she has it. Then we had seen Also in that first video, this is the Third, what I did here when I had This activated it was that by itself to time to hit run

[1:16:44] time to hit run the text we would like the text we would like to switch from text to speech, from text to speech, to switch from text to speech, from text to speech, I was going to take this file first and then

[1:16:58] I was going to transcribe it to understand what it is what it says. and with that be able to later use a fairly similar tone to the exit. That's very fast, very

[1:17:10] automated, highly transparent, without However, that doesn't change the fact that it's a process. additional work is done. Therefore, if instead of having to do it by itself try to transcribe each time, each time that we're going to execute, if by

[1:17:24] Yes, we already have the file and we write explicitly the well-written text that It says there, so what we're going to "Getting" doesn't necessarily sound better. not necessarily, but rather save ourselves a step

[1:17:39] which takes practically no time, but Ultimately, it's an accumulation additional. Uh, let's skip that one step by step and with that it will be faster result. It could happen that writing it ourselves manually

[1:17:53] generate a better final quality. When? When I don't understand correctly the narration and transcribe it incorrectly. But then The goal is for the input audio be easily understandable. We're not going to to make nothing too complicated. We're not going to

[1:18:05] put a tongue twister or something, or a voice of very poor quality. So no because I should be able to transcribe without problem, but yes, a little bit in speed because we are skipping this step we'll just send him the audio and

[1:18:20] the text of what it says audio. So I write hello here. And when you click on run, well, what's going to happen? happen? That after it loads model and those initial things that it does The first time, it will generate a

[1:18:34] audio file that says what We write here. However, Here's a detail. It may sound very similar or very different from what similar or very different from what We had at the beginning. Hello. Hello.

[1:18:49] It doesn't sound very similar, although well, it is a small piece, not very representative nor to understand properly. Hello. Let's put something, let's put the same input text. It doesn't make sense, but oh well.

[1:19:03] What's going to happen here? which is clearly going to be longer. This will help us realize what It is so similar, and among them we will find It is so similar, and among them we will find account of, "Ugh, it doesn't sound identical."

[1:19:16] There are several things to keep in mind here. The first is that it must be completely representative, that's enough for him and about to fully understand the tone and Furthermore, but another thing is that a The same tone of voice can be altered.

[1:19:32] Each person is born with a voice, right? But we can do it a little differently. But we can do it a little differently. So, that's something to keep in mind. This is the base voice. You can do a not very different. And in this case, with

[1:19:46] that? With specific instructions how we want it to be narrated, among them even the accent in question. Of all ways, unfortunately, for the Spaniards, commonly, given that they train this type

[1:20:00] of things like, ah, in Spanish almost They leave the accent aside a bit and given that they are more Latino than Spanish, commonly the form of narration remains more, in quotes, like a Spaniard neutral, but more than neutral Spanish,

[1:20:13] neutral, but more than neutral Spanish, It looks like generic Latin American Spanish. AND then it could be that practically We will never achieve 100% a narrative with Castilian accent, with a Spanish accent, exactly as we would like, but from

[1:20:27] that we can make more similar and more Forced, yes, it will be possible starting from the instructions we give, as well as other details. But to mention everyone ways that, despite everything we do instructions and all that, given that

[1:20:40] These are unpredictable models, all of this already We know, artificial intelligence There are going to be phrases that will... pronounce better, just as we are indicating, but others not. With the which I can tell you in a Spanish accent

[1:20:55] specific details and everything that I want to, but maybe someday I will to say, "I am the owner of such and such a place." You can mention him as Uruguayan or Argentinian or whatever, there might be things of that style. It very rarely happens

[1:21:08] deviate a little from the instructions that We'll give it a try, but it'll never be 100%. perfect, and even less so when they are going to be, Let's imagine, 8-minute narratives, Well, clearly a part can leave. a little bit of what interests us. Without

[1:21:21] However, despite all this, They seem like excuses, it's just to give context, because anyway this is going to sound much better than what we It generated, for example, the GPT model. Just something to keep in mind here.

[1:21:34] we can't even see here doing things correctly visual, but this doesn't matter twice over reasons. The first one is This is the browser and the second one We're not going to use the browser, so it's going to be...

[1:21:48] Here we are just recapping how This model worked, but oh well, if we can see all the exits that have There have been, because I have a lot of here generated content that I have to return to be eliminated, by the way. But anyway, one

[1:22:03] That file should be recent. a voice that should be called a voice anything, voice and something else. By What else? Because if it already happened in the past I did this because there was already a boss, one boss two, one boss whatever. Although

[1:22:18] Organizing 29,000 files, well, as They will understand, It should take a little while. Not too long, but it will take some time. Otherwise, there will be a voice. Phew! Well, too There are a lot.

[1:22:36] In case anyone doubts that I use things in local. They'll probably be sorted by now. Phew!

[1:22:52] This is very long, Hey? This has to be it. That's how it is. who knows how to take them, so if you get them One, you can be sure that

[1:23:04] It's really for you. This is the result. Great opportunities come to who knows how to take them, so if you get them One, you can be sure that It's really for you.

[1:23:17] Great opportunities come to who knows how to take them. So if you get it One, you can be sure that It's really for you. It sounds higher-pitched, more energetic, like more

[1:23:32] epic. Because? Because despite being, we do we realize that it is the same voice base, well let's say the instructions They can slightly alter the rhythm, the style, to the mood we want with which

[1:23:46] narrate. I'm going to delete all of this so I don't have all of the above. Perfect. So, oh, what should I use as the prompt? How do I make a good prompt for that? Well Why complicate things? That AI becomes more complicated. We're entering the world of AI

[1:24:00] with search capability or even a Locally, I don't know, but I can search the internet to tell him, Investigate using official sources. Investigate using official sources. What should the prompt be like?

[1:24:15] for give correct instructions give correct instructions narration to quen 3 tts? And well, it's not going to be what we assume or what we understand from a single

[1:24:31] answer or a single result of internet, but here it will search for everywhere and with all of that it's going to do things like you do important things, the common patterns that the different online recommendations. Then with

[1:24:46] This is not just that we will be able to not only to better understand how to do it, but also do it. Therefore, I myself tell him, I want it to sound like this, but like that. prompt and done. Here you'll see how Because it's going to have all the context, right?

[1:25:00] Because it's going to have all the context, right? And if with a prompt, We can tell him, "But it turned out "It's very muted, I want it more like this." And Well, he's going to keep iterating it. Here it is. TO starting from having investigated who

[1:25:12] He knows how many sides, because he's giving us the information on how it should be more or less the prompt prompt for shorts most viral. Interesting facts, most viral. Interesting facts, My final formula for your shorts would be,

[1:25:26] My final formula for your shorts would be, And when did I say shorts? oh well. Okay, very good. Let's go tell him,

[1:25:40] tell him, Write a prompt with the following details. Let's reinforce it first. Male voice, serious, cheerful. Well, male voice,

[1:26:03] serious, cheerful, what do I know. Cheerful, animated. sensationalist. I think that's more or less what I would call myself. attention for uh an animated YouTube short. I put it on

[1:26:19] animated, perhaps that's already understood. animated, perhaps that's already understood. bit.

[1:26:42] promoting Bitcoin. Which of the Which TTS accounts do you use? Uh, this, the 0, the Which TTS accounts do you use? Uh, this, the 0, the 1.7B base or 0.6B base? Hey,

[1:26:54] 1.7B base or 0.6B base? Hey, The 0.6 base doesn't sound worse, but narrative style. Even so, it's still from better quality than GPT, but yes it is Because it's flatter, it's not as versatile, so animated, it doesn't sound more robotic or more

[1:27:10] Noisy, no, not at all, it's just less. versatile in terms of narrative style, at least in my opinion. To count between this TS I would use this Prom in the instruction field, version one slightly more intense. Oh, well, if one doesn't

[1:27:24] It works for us, the other one works for us and We glued. Ah, we didn't say here that with Latin American Spanish accent, for example. Is Furthermore, let's tell him, he adds that it must be in Castilian Spanish from Spain. I clearly wrote everything in lowercase,

[1:27:40] oh well, to test. Random question, what do you think of the team? MD Ryen AI Halo? which is usually more marketing, more than anything else, because there's a lot of it. integrated memory in the processor, yes,

[1:27:55] but it still has the speed of a RAM, not a V RAM, so no It is neither the most powerful processor nor the More powerful RAM. More powerful RAM. You can run it there now, yes,

[1:28:08] but more or less the same as in CPU and RAM, No? It's not the same as a GPU, so which, given that it won't be that fast, You won't want to carry things too much large things, because if you carry large things, They're going to be slow because they're big and because in

[1:28:25] That equipment doesn't work very fast. So that you're going to try to load more things or less small than those that don't work as well Slow, small things that fit in a GPU that costs the same and that works to work faster and that you're going to

[1:28:38] VR RAM. Furthermore, in that case, being from AMD, unfortunately, almost all AI generative is primarily intended to envy, so if you want ready. I finally see you live. Me I would like to know if you have already purchased the package.

[1:28:52] I could just pay the $60 for the monetized channels. Can you Yes, contact me at [email protected] We can see it. I don't want to overdo it either. We can see it. I don't want to overdo it either. [email protected].

[1:29:08] What is the maximum length of STTS for Will it work well for you? Uh, or You cut it yourself or he cuts it for you, with which will result in a size practically nothing about that was very long, oh it was very short. Uh, no, it doesn't happen

[1:29:23] nothing. An anecdote regarding the data of neutral Spanish or closer to Star Trek was only translated into Spanish Latin, that is, the Spanish They had to see her like that and they're more of a fan. over there, they obviously prefer their

[1:29:35] dialect. Uh, not only in that one, in many old Disney movies, like The The Jungle Book was folded only in Latin American Spanish, that more than Latin American Spanish, it seemed Mexican Spanish because I know almost everything

[1:29:49] Created by Mexicans, so even sometimes in different types of references or sometimes in different types of references or They used humor that was more Mexican than generic terms of the Spanish language. And in others movies, in the 90s and early 1990s

[1:30:04] In the 2000s, they did it less and less. They were doing There were two variations of the dubbing that was pretty much the same thing throughout the whole movie or series, but certain phrases adapted to Mexico and others are generally supposed to

[1:30:17] Latin America. For example, in searching In one part, Nemo says, "There will be snack And in the Latin American version it says, there will be snacks.

[1:30:31] And in the Latin American version it says, there will be snacks. And in the Mexican one it says a pachanga And in the Mexican one it says a pachanga And in the Latin version it says a party. So it's always been like that, and that's why less, that is, to a greater or lesser extent and

[1:30:44] At least in this case, everyone is free to do as they please. apparently it has more or less its dubbing, although even so, sometimes in the Latin American dubbing includes a lot of things Mexican women, like in the Shrek movies, for

[1:30:56] For example, or Srek, uh, mainly everything the donkey dubbing. Very good, here he has already adapted this a bit. We paste the prompt and we're going to put another one The difference is slight if we...

[1:31:10] We more or less agreed on what it was like. Now Let's remember something, we're not going to use it like that. But first we're doing it this way more manually with the aim of knowing what things we need to integrate as input In the code, which file should be uploaded?

[1:31:26] What instructions can we give, how We can separate the audio if we want. only that it goes from text to speech, but with some specific instruction. Or to The best thing about it is the great storytelling and we want one part to be in a certain tone,

[1:31:40] another part with another. All of that is going to... to be of interest. Very good. Uh, let's go play this. So, Great opportunities come to who knows how to take them, so if you get them One, you can be sure that

[1:31:53] It's really for you. The ending really did sound strange, and in fact He doesn't speak much Spanish, but he does It sounded different than it did a while ago. Great opportunities come to who knows how to take them, so if you get them

[1:32:06] You can be sure that really It's for you. Well, the thing about it still playing quite Latin American, But overall, yes, the style did change.

[1:32:18] other details. We can try more telling him that it wasn't enough with the accent thing or whatever. And if not, Anyway, and unfortunately There are more Spanish-speaking people than there are accustomed to Latin American Spanish than to

[1:32:33] Spanish, eh, that is, the Spanish of Spain. With which, even if I were Spanish and was used to the Spanish of Spain, if my Spanish content is going to be consumed by anyone Spanish speaker, I would almost prefer in

[1:32:49] Spanish speaker, I would almost prefer in Latin accent, because there are more people and many more people thanks to the fact that there are more countries and so on in Latin America that only Spanish from Spain. By the On the contrary, if it is 100% for one thing in

[1:33:01] Specifically, Spanish from Spain, because another That could be the thing. But even I would do it. I would leave it like that. But it's clearly something that We can change. Very good. Let's go We can change. Very good. Let's go Try something else here.

[1:33:20] I'm just going to tell him to write the prompt to narrate in a style terrifying, slow

[1:33:32] and for reflection. Which label? things. What you said about the translations and what didn't convince me about

[1:33:45] The MD Halo is the sluggish bandwidth which it has and the high price. Yeah It would cost 900, one might think. That's how it is. At those prices, you'd be better off with a GPU. Furthermore, another thing, both Latinos and We Spaniards are bothered when it's a

[1:33:59] It doesn't bother anyone, even in podcast. Yes, that actually makes more sense. more accustomed to seeing all kinds of accents and in a movie or series, another That could be the thing. But there it goes depend mainly on

[1:34:12] our preferences, because if we need is addressed with an emphasis on In concrete terms, it will be something else entirely. And then our preferences. Very And then our preferences. Very Okay, let's try another text.

[1:34:39] narration terrifying, slow and reflective. Why is 5 + 7 25? He continues to seem very cheerful. Because? By the start.

[1:34:54] 15. Okay. So this also matters a lot the base. Well, but there anyway We've more or less understood all of this now, No? So, now that we understand how Does all this work?

[1:35:07] an example where the accent is more noticeable Spanish might also work, But anyway, leaving that aside, let's go Leave it as is. And then this to incorporate it into our workflow, in our YouTube shorts stream,

[1:35:21] our YouTube shorts stream, TRUE? So to the promptil for YouTube Shorts. And this is exactly how it is.

[1:35:34] We have it ready to go. We're going to save it and start it. to put into our workflow, into our Python code pipeline. That's going to be the voice file, the That's going to be the model, the best example.

[1:35:51] The voice transcription is this. How do I always want it to be narrated? Goes to be so. The only thing we're going to want that it's changing the code, well, it's going to be the text itself that will be passed from Text-to-speech, right?

[1:36:05] It's changing the other values, but It doesn't make sense if they're always going to be equal. The one we're going to want The change will be this one over here. So, we left that one empty, the target text. We'll leave it like this and

[1:36:20] target text. We'll leave it like this and We're going to go to file > export API. It's practically the same at the end practical, with the difference that in Exporting weighs much more compared to export API, because in exporting APIs they are

[1:36:35] export API, because in exporting APIs they are the nodes, their values, and their connections. the nodes, their values, and their connections. While in exporting it's what I said in conjunction with the exact position, the a little color of who knows what, etc.

[1:36:50] It's for the most visual part that remains exactly as we had it organized, but when we want it for a single code We want the logical part, the connections and the values. So we give it a go export API file and save it as

[1:37:06] be. There it is. So, And we're going to put this little file in our set of scripts. Here I go to save. We have one called api.jon, which is the that we used for the images. It

[1:37:20] We're going to be able to rename, in fact, we're going to to do it. Moreover, all of these their name start uh with an underscore and then a letter with an underscore and then a letter guapi.jon

[1:37:42] Turbo. That's what they call it. set of scripts. Oh, but I don't know how to do it

[1:37:58] ask AI for the change that I interest or to do it myself. It doesn't matter necessarily because even though the artificial intelligence, the models of They won't know the exact language. of no specific model

[1:38:10] specific details, what they do know is understand the logic. And simply by watching They will be able to understand how each one is used thing, what is the entrance, what is the output, in addition to Z image Image Turbo de Cuentras TTS, of those models and

[1:38:25] those specific quantifications and those They won't have ways to execute it. information, but yes, some Confidence UI, they understand more or less how it works Confi UI, so that would be it enough. But if we also have another one

[1:38:37] enough. But if we also have another one something here like this, a code in which We already used a confidence example in which to base it on, but for that we'll tell you that it is based on this. Anyway, We'll take this opportunity to update

[1:38:50] the name of the JSON file with which we load the image model. Perfect. And now then we're going to do the thing stories. We can use Codex App, we can use

[1:39:04] Open Code, we can use what we happen. I'm going to open Open Code. Only I want to review why Open Code, because Last time we used Q3TS, I mean Q3, We use the Codex app, so just because change, not for any other reason.

[1:39:23] Hey, if we don't have an example or what AI to be based on that, to adapt the code, we can tell him to investigate. What we did a while ago, not only in GPT Chat,

[1:39:41] development, if we even have it. connected to some kind of search in Internet. Here, for example, Dipsic V4 Flash and other models in Open Code are free, with very little use really that We can do it, but they're free.

[1:39:58] The advantage we're going to have isn't that much. not only to be able to use that model, but also Since it is already part of your subscription They already have them connected to the internet. On the contrary, if we weren't using something connected to the internet, if

[1:40:11] For example, if I were using llama CPP, Well, I must have, for example, an MCP that allows me to use the browser or search in certain development information libraries or search directly in a browser like

[1:40:25] in Brave. with some MCP we can give that functionality, but as I say, if the environment we already have is not enough It is integrated, for example, in Codex, Well, you can now search on the internet for them. GPTs, in the Cloud you can now search in the Cloud

[1:40:37] Code can now search the internet for cloud models, I don't know, well in Open Code can now use search models that Open Code itself provides. So that dipsic V4 flash and we could I would say to investigate, although I wouldn't

[1:40:51] lack. Because? Well, because it already We already have examples. Also, here's another one thing, it seems he says no It works, who knows what, but Okay, let's try it. I can tell you, analyze and explain to me how

[1:41:08] I can tell you, analyze and explain to me how The connection works The connection works From Python to confiágenes.pi. Here, to help us understand, we can give it multiple instructions at once.

[1:41:23] passed. He's going to analyze it there and he's going to tell me. explain it to me, but at the same time it He's going to explain it to him. It's not that here conversation, the whole context of that code, but what it does have is this and

[1:41:38] the reference for where you can check further. So now I can tell you, further. So now I can tell you, now analyze

[1:41:57] And tell me what changes would need to be made so that it works with the conf and the file so that it works with the conf and the file Qen 3 TTS.

[1:42:11] things, we are asking you to analyze. But I reiterate something else. If we don't have different things in which one can to ensure that AI doesn't reinvent the wheel,

[1:42:23] What we can do is tell him that Look, let him find out how to do that. Looking at example code online, Well, you'll be able to find out. how is the model in question made, but so that you understand more or less how

[1:42:36] All of this works at Confiju. In fact, I can tell you, In fact, I can tell you, research on the internet to validate your plan.

[1:42:50] It's not necessary. We already have examples here. And it's not that complicated, but... You can also search on the internet and ready. He's going to look for examples of Conf UI, Python, Wflow, Cuentrast TS, more or less With a little bit of everything, you can make one

[1:43:03] With a little bit of everything, you can make one idea.

[1:43:16] channels, one in Spanish from Spain and Latin? It depends. Uh, I have a channel called Nichonauta in which I, with my Mexican accent I'm talking about artificial intelligence and I They come from Colombia, from Argentina, from Chile,

[1:43:29] from Paraguay, from Uruguay, from Spain and from anywhere. Depends. So, uh, now yes I was talking about a service in specific that is offered in Spain for Spaniards, For example, well, then preferably.

[1:43:44] It would be better specifically in Spanish. from Spain, but not another one in Spanish Latin, but only specifically for that place. But, I mean, in terms of content The internet doesn't necessarily have to be 100% the exact accent. That will depend.

[1:44:00] exactly the specific case, but I wouldn't see much point in creating from Spain. Better one in Spanish and another in English, Because I, being Latino, see on YouTube Spaniards.

[1:44:15] Awesome. They play the quality of Quen The Spanish TTS left me amazed and Hello everyone. Interesting to integrate TS account. I often ask Cloud that automates with FI and creates the workflow is getting lazier every day. Uh, yes, to

[1:44:28] mainly for the reasons we discussed, simply by understanding a little Confidence UI and a little bit about the nodes and other things that We have the flow, because with that we can Trying to adapt, it doesn't have to be Know everything perfectly, but the more

[1:44:41] context, then less likely that it I made a mistake. Ryzen AI Max Plus 395, 256 GB bandwidth. Uh, I don't know what how much will the card I have cost? 50 Ti, 896 times more bandwidth and the models to load almost the same. Exactly.

[1:45:06] memory. But it's not just the card and plus you have to add components of quality between $00 and $000 more. No? When you run artificial intelligence and that it will practically fit completely in the GPU, the other components of the

[1:45:20] PCs, while important, are secondary. Because? What are you going to do? secondary. Because? What are you going to do? Does the CPU load when loading an account? To ask that the GPU processes what we are asking it to, But the main thing, the very important thing, is the

[1:45:33] GPU. Cape Verde was eliminated. That's how it is. Internet. Your plan is confirmed online generals. My plan, his plan, I don't know it I did it, but oh well. Corrections

[1:45:47] necessary. Who knows what he wants to do. Okay. Ah, Does he want to put it in another folder or something? So. That was his correction. move it to another folder. But anyway, in Yes, the model already had the

[1:46:02] do it. So now we're going to tell him then implement the change in voices.pi. Do not modify any other files. Eye,

[1:46:18] Do not modify any other files. Eye, I just want the code to adapt I just want the code to adapt the next step in the workflow. The rest leave it as is.

[1:46:32] Random seed, voice of uh entrance voice

[1:46:49] textar. To switch to voice.

[1:47:02] The name of the voice file will be hm

[1:47:15] and will be at the root of the project. So So we're going to... our voice example, Put it here and give it the name you want. We said that he's going to try to use the model.

[1:47:29] model. That's how I left it. was. That's how it is. And with that, it will already be taking

[1:47:41] That code will always upload this file to trusted, so that they are not always exactly equal

[1:48:09] The same person, for the same entrance, will come out. It's going to be exactly the same exit, but natural, that he doesn't always narrate it exactly the same. Uh, so that's why We're going to want this changed. random form

[1:48:32] Flash version and free, so let's go It's even better when quantized to 8 bits or to Four, I don't know, so no It is necessarily the best model, but neither context nor a lot of knowledge. He has the knowledge

[1:48:46] starting from these contexts and it won't require understanding a lot of a moderately simple task, it's not being working on 50 files at the same time time. So, in itself, a model in quotation marks small and bad between

[1:49:01] Quotation marks can work very well. Okay, I made the changes there. Uh, now We came this way and now our voice code is no longer in any

[1:49:13] our voice code is no longer in any The side should have the use of Openi The side should have the use of Openi and neither is the use of GPT clearly present, but which will now use Confi UI, that The URL that you saw in the example of

[1:49:26] images With that file, Jason, with that file of voice and this is where they will be values ​​that will change in the node of

[1:49:39] values ​​that will change in the node of audio upload, which is the ID1, reference. At the other interesting node where

[1:49:52] target text says and at that same node where it says it will go and at that same node where it says it will go change the seed to a random one. change the seed to a random one. So, if we for this project

[1:50:05] We're going to create new YouTube shorts from scratch to remember everything. If in this project we generate a YouTube short, now I should generate it with trust and accounts based on this example of voice.

[1:50:20] So Another thing, I also think something here additional thing he did, No, he didn't. Uh, I thought there was of syntax, but it doesn't seem that was it from a previous execution. But also

[1:50:34] if there were no errors or if he executed it Just to try it out, I don't know. So, let's go Just to try it out, I don't know. So, let's go to create a new YouTube short.

[1:50:54] And in fact we're going to change more things in the path. artificial intelligence already knows that it is not necessary to give it the real context right here, but There are other things where we need add more information, I don't know, but

[1:51:09] Typical general knowledge stuff, you know. You'll see them, for example

[1:51:35] the origin My surname is Baladés. This is something that the model surely... I could know about language good enough. If it's a model medium large, like a Gema 426B or

[1:51:50] 12B or 31B, the youngest ones go to invent the information. Uh, this would be It's better to look it up online, but... Yes, it's something, since there must be a lot of it. time, so a model of a medium-sized, good, because

[1:52:04] It will always be about giving it more context, no only this, if it cannot be done invent, but oh well, it's something moderately typical that I know the answer, so we'll be able to see it. We saved. So, now that we have

[1:52:16] that, when executing scripts with CPP call when executing scripts with CPP call open, 412B generates the script for us, like We had previously done it in the other

[1:52:31] directly, from this prompt in Specifically, we are only going to try to do a project, a video. So, Gema 412B will receive this prompt. AND

[1:52:43] Here's the topic at hand, and with that it will try to follow the instructions. So we click on run and here Apparently there would be a problem. Which? No, well, in fact it already happened, it More CPPs have already closed.

[1:53:00] That's what I was going to mention, that if already We have Confi UI open, the problem is not It's that Confi UI is open, the The problem is that it's loaded in Conf UI Q3TS, so it's already consuming GPU resources. If the

[1:53:13] model that we were going to load into flames CPP was smaller and copied, not whether there were problems or if we had them adjusted to also be able to use CPU and RAM. But since I don't have it with anything That's it, he wants the GPU free so that

[1:53:26] It doesn't fit because it doesn't fit. It's a model medium-sized, being 12B for a medium-sized, being 12B for a GPU as small as a 3060. There it is. both things wouldn't fit as we are seeing.

[1:53:41] Then, we click on run and go to generate the script to then generate metadata, that is, title and description and After that, generate the voice. And there TRUE. They are already starting to generate tokens there.

[1:53:56] based on the information that he we request and so on. It's going to take a while because it's not due to the generation not from the script, which is very short, but because of the Another one, no, but this is better. Uh, but

[1:54:09] I rewrite it like this, ah, then the other thing, the reasoning is helping us not to riddle, not a problem of programming, but to write the script fast enough. Despite being a simple GPU, being live and being

[1:54:24] a dense 12B model, even so, because It worked very well in terms of speed. If we review then, So now we have the next column. Here, for a single video, we see that no It makes perfect sense, doesn't it? The

[1:54:37] when there are a lot of things. Despite That's not bad. Because? Okay, Well, it was just a response. Yeah, In the next one there will be two, but in The following will be several, they may be up to 10 each part of the video that

[1:54:50] Write the prompt for what should to be displayed on screen to generate the project, automation is coming deluxe. We're going to review the script more or except how it turned out Also because we're going to see something here

[1:55:03] interesting for this content. What is What will it display on the screen? Are we going to leave the model alone or what? Know

[1:55:18] Baladés? Even though you carry it in your identification, this name originated in the land of Spain. It's a surname. toponymic, which means that It was linked to rural areas of Castile and Navarre, identifying the families

[1:55:32] who inhabited specific valleys or areas delimited by fences. Your last name doesn't It's just a name, it's a map of a story that defines your roots and also Placencia, although it comes more like from center of the Republic,

[1:55:44] from Jalisco, from Zacatecas, from Aguascalientes, which is where I'm from. Uh, well, So that's where the script ended, right? Let's look at several details. Ah, this execution follows.

[1:56:02] Ah, this execution follows. And I was wrong. I executed the following without having executed the previous one. Oh, it gave mistake. Yes, the mistake is not having mistake. Yes, the mistake is not having Open with FI, but not really here

[1:56:14] Nothing would have happened if you had open trust and was requesting this next part. Because? Because this is what's going to The process of converting script to voiceover is as follows:

[1:56:28] generate title and description, which gives The same applies to the next step. Here it is The same applies to the next step. Here it is Our next step is... Because? Because both will occupy the CPP flame for not to be closing and opening like crazy

[1:56:41] every now and then, which we already are. I am doing it, but it's for grown-ups just so it takes 2 hours to do One thing, then two hours later the next. I don't know, we might even be able to automate the opening and closing,

[1:56:54] But anyway, I mean, basically here We can run this part and it will work. Because? because only It requires the script, and the script already does. We have generated. However, then First we have metadata.

[1:57:07] It's not loading yet, it's not showing up yet. The server starts correctly. Hey, We don't have it yet, Or rather, we first have metadata, because it is already occupied by cpplama cppa, I trust.

[1:57:25] other. It seems that the file name It's not exactly the same, or it's not It's not exactly the same, or it's not finds. That's how it is. Guos.mp3.

[1:57:41] The way he did it was good code, right? Let's move this on to the If we can't tell her that she does execute it and realize the error. Uh, the

[1:57:53] execute it and realize the error. Uh, the The execution shows me this. The execution shows me this. Uh, cmd. Let's write

[1:58:17] errors, but for other reasons dependencies and all that, well, let's go dependencies and all that, well, let's go tell him to just focus on the correction tell him to just focus on the correction of

[1:58:31] of the shipment correct audio file. correct audio file. I don't know what happened here, but oh well.

[1:58:47] black thread and 48GB of server SCA RAM, That much isn't necessary. I understand the priority, Juan of the components, but You're not going to put a toy model in it, are you? one outside of poor guataje, huh? A source of poor quality if you want the

[1:59:00] Team lasts, huh? No, mainly there. the source of limitation. The motherboard It doesn't matter which one you use. The motherboard It doesn't matter which GPU is connected. Well, Except for its weight, eh, but Why would it make no difference? It's just one

[1:59:12] PCI Express communication, it's not that the the motherboard has to make the GPU, that's what will make it work to be electricity, so yes, one the processor asks for things apparently a little slower, either

[1:59:25] This means that it will be damaged. processor, motherboard, and GPU. By The processor is going to do something very simple, request something from the GPU. So the that makes sense, but the rest doesn't necessarily. And how much did it cost you in your

[1:59:40] drivers? A source like this is always good. So is. And it's Sonic 850 eh plus bronze and the motherboard with a Chinese motherboard. That's it. A few years, eh, price I think about cent and something. Motherboard with 32 GB of RAM and the

[1:59:55] None ever. Excellent. Good information. Did you choose that SEO for a particular reason? Or was it something random? No, not at all. 10 cores, 20 threads. The issue that those Anyway, it's cheaper than arming yourself. a brand new team. More or less. He

[2:00:10] processor and on the motherboard, it's in the RAM and the GPU, but yes, more or the RAM and the GPU, but yes, more or less. the way he tried to climb the things. Let's see if it's supposed to be with

[2:00:24] That fixed it correctly. If not, here He's already giving us another clue. We could have passed it to him route from where Confiui was so that He will review the files. Not in this case It was necessary because the logs themselves

[2:00:37] They probably mentioned details like that. That's how it is. that. That's how it is. Very good. now correctly because it continues in execution. Continue with the error, then

[2:00:50] because it was the previous code. We're closed And we opened. And it seems that now it's going to work. There it is.

[2:01:06] what we are doing. Maybe not We're seeing a lot of the logical side of it. we are doing. And besides, another thing is not a code for us to run. for a banking app or whatever I. So it doesn't matter. A

[2:01:20] application for us locally, There's not much of a problem. So what happens then? well that the The GPU has accounts loaded in the

[2:01:32] The GPU has accounts loaded in the version 1.7B full precision base. Overall, not just the model, but also things like, for example, Confi UI, which we don't know if it already occupies Some GPU, if not RAM and CPU, something there

[2:01:46] can occupy. Anyway, something to keep in mind, How much RAM will be used? The more The more RAM we have, the more RAM it will use. The RAM It is there to be occupied, to be used. With which, if we have eight, both the same

[2:02:00] operating system, like programs, They're going to try not to oversaturate that. Yeah We have more and they detect a good amount They'll be free, and they'll try to take advantage of it. Of They'll be free, and they'll try to take advantage of it. Of What shape? Cache, for example, instead of

[2:02:13] recalculating things with the processor, temporarily leaving things loaded into memory so as not to reprocess them all the time. Okay, that's it. We already have our code then prepared for

[2:02:28] that generates the voices of each of the projects that we are going to do. Here only Automation doesn't have much of an automation feature. sense. Even so, it's better than being clicking everywhere a lot of programs, but clearly between

[2:02:42] more projects because automation It makes more sense. Let's review then the project. We have the original files that it generated for us, uh Confiy UI with TS account and this file What it does is compress the audio, no

[2:02:57] not in terms of size, but in terms of compression of soundtracks. eh the volume peaks to make it sound more evenly, and then increase the volume so that everything sounds loud and even. strong. Do you know where it really comes from?

[2:03:11] Is your last name Baladés? Although today it wear this name on your ID He was born in the land of Spain. It's a toponymic surname, which means It was originally linked to areas rural areas of Castile and Navarre

[2:03:24] limited by fences. Your last name doesn't It's just a name, it's a map of a story that defines your roots. There it is. Okay, then There it is. Okay, then It offered higher quality, it's more versatile, it's

[2:03:38] It's free, it will always work the same way, and I don't care if it doesn't get updated. I don't care if it doesn't get updated. Honestly, it lacked quality. No, nor We don't even need a 100% new one version.

[2:03:53] Do you know where your Baladés surname? Even if you wear it today In your identification, this name was born in the land of Spain. It's a surname. toponymic, which means that describe a physical place. Originally

[2:04:08] It was linked to rural areas of Castile and Navarre, identifying the families who inhabited specific valleys or areas delimited by fences. Your last name doesn't It's just a name, it's a map of a story that defines your roots.

[2:04:21] a story that defines your roots. Okay, here we need to have three Things to consider about how it will turn out the final narrative. One is the script. Yeah We make the script emerge from a prompt indicating that it is very

[2:04:35] slow, short sentences, with many commas, with lots of semicolons and everything Therefore, it is clear that the voice generator model will do many pauses. On the contrary, if the the script becomes more fluid, more streamlined, with

[2:04:50] longer sentences, because that's how it will narrate. That's the first thing to have in mind account for the narrative. The second one, the example of the voice we give as input. And third, the instructions that are written here with the details of

[2:05:07] how we want the narrative to be. Without However, these details are not enough, yelling, no matter how much we put here Someone whispering, well maybe not The difference will be noticeable because of Yes, the entrance is of someone screaming and

[2:05:22] if the input has background noise, It has an echo, the output will be the same. Because? Because we are cloning. So which is something to keep in mind. The The entrance must be good, otherwise the The outcome is going to be just as bad. Very

[2:05:35] Okay, that's where the voice was generated. If we had A lot of projects, huh? we could then stay longer in This part, which here was only one, We could have been giving it to him, he We would run it and it would take a while

[2:05:50] now doing it from text to speech for 100 YouTube shorts, for example, to say something. Or if we want a longer video, Well, change the promo so that it's... a script for an 8-minute video in long format, mentioning who knows

[2:06:03] What, let him mention who knows which one. Come on Let's see more in this same live stream. Now, alas, we're skipping metadata. Phew! And I made a mistake and went to subtitles. Will it give an error? No, not that either. What takes up to subtitle the audio file you

[2:06:17] to subtitle the audio file you Have we already generated it? Therefore, here too metadata is not yet needed. Of all ways, at some point it will need. In the next part, in scenes, you need to understand, for example,

[2:06:30] the title to start from the title and the the script that will be here, it's already there script, but the title is missing, understand 100% the context of the video regarding it describe each scene, just as we had seen in the previous live stream, in the

[2:06:43] latest video on the channel. But anyway, I mean, here they are numbered, not because it has to be 100% definitive that order of execution, but rather Well, based on preferences and recommendations. according to the order in which we do

[2:06:58] things. I use Llama CPP, because I still With CPP flame, why wait until With CPP flame, why wait until after. Okay, he already did the voice "eh". text. Whisper took 12 seconds. AND So here we also have for this

[2:07:10] So here we also have for this unique project subtitles.srt srt. We have word for word in what It is at this moment that it is mentioned, and not only the It ends, but there are other things it does. Very good, Whisper. Ah, it's said with a tone

[2:07:25] of question. puts the symbol of corresponding question. Mm, here he paused. I'm putting a comma. Whisper is so good that it even does that.

[2:07:37] Just as the text-to-speech model is It's good to pay attention to these things, so that the pauses are noticeable, the tone of question, statement, or whatever, Whisper is also good at understanding it. back.

[2:07:52] We had seen the metadata thing. leapt.

[2:08:09] Cloning. Yes that's how it is. That's why We choose. We mentioned all of that in this live and in the two previous ones of this same topic. It's not just that it's free, You can choose from voice tones that it already has default. You can start from a

[2:08:24] prompt to design a voice, that's not it We've done it because it doesn't make sense. too much, and you can add an audio of entry as an example for cloning the voice. So either way you can choose infinite different voice tones,

[2:08:38] so to speak. It produces good quality, It's fast, and it comes in two sizes. Even the "Big" is very small, 1.7B. Hey, the little 0.6B 6B is still working pretty good and it's not just that versatile in tone of voice, but in the

[2:08:52] instructions for the tone and rhythm of the narration. not for free, but because it's better. Have gem in local, because it's better, not only

[2:09:05] for free. We have local accounts because it's better, not just because it's free. better, not just because it's free. Here we are Subtitle the times correctly completely accurate specifics compared with online services that won't help us

[2:09:20] not allow uploading any duration or any file size. Uh, and it's This is very accurate. Uh, we commented in the direct previous that when I used this The same model online doesn't work the same way. good.

[2:09:34] Follow scenes. Perfect. So, I don't know why I opened Confi, we don't need it this. doing. Well, basically we need gem

[2:09:49] specifically here because First, what we need is to generate the title and description that not yet We have generated and then because for We'll need those scenes too. so that from this prompt the

[2:10:03] language model generate the prompts of what should be displayed on the screen. AND Here we'll be with another adaptation which is not changing code and it is not changing of model, but it will be changing the instructions, not because they are

[2:10:17] not bad, but perhaps because for this We are interested in another type of content stuff. Before that, let's try exactly how will it, on its own, to do things, but let's remember Something from around here. Despite these

[2:10:29] type of content we are interested in creating and so on, that it's a good script for a short, make it a good description for a pair of shorts and that are a good scene description for a YouTube video short and the Z Image Turbo model, despite

[2:10:42] short and the Z Image Turbo model, despite However, there is one detail about this, and that is that in this We're doing it too much, in case we're doing it too much. versatile in the sense of the type of image that it can show, but at the same time We are limiting other things in time.

[2:10:55] that we are not usually interested in and We are providing enough versatility. that maybe we won't be here either I'm interested. Here you'll be very free to choose the image you want to display most or less. Because

[2:11:10] Do we want a clean exit? No elements of text, not symbols, not letters, not numbers, no logos, in something that speaks of the word trivialities, from the surname Balades, which means it is It is likely that we would be interested in some

[2:11:25] in the scene someone is seen writing word balladness or see the word written on a sheet of paper or whatever, on a page, whatever. So, Perhaps for this content we don't It's interesting to be so strict, and perhaps

[2:11:41] We are not interested in it always being by itself I tried to create a realistic style and people, from places, who knows. To the We're more interested in it being more stylish. drawing and let it be an open book, let it be a

[2:11:57] and let it be an open book, let it be a Spanish flag, whatever, I don't know. So, it's not that it's 100% necessary Change it, but maybe for a certain This type of content may be convenient Let's do it. We already generated then

[2:12:12] metadata, which we had skipped over and We continue with scenes. So, Here in scenes, for each part of of the video, every 3.5 seconds or more

[2:12:25] At least, he's going to decide what to show in screen. Based on these instructions, from that title and full script of that section specifically from the video. We had already seen in the previous live broadcast

[2:12:40] that there is value precisely here, which in this case is 3.5, which we are trying to put words together until blocks of at least 3.5 seconds or a

[2:12:52] blocks of at least 3.5 seconds or a A little bit more. 3.5 seconds. 3.5 seconds. 3.5 seconds so that roughly every that amount of time appears one image. So, with that, Gema is going to decide what things to show on screen.

[2:13:09] This at the same time being a species of prompt so that in the next part of prompt so that in the next part be Z image Turbo within Conf UI the that generates said images from those prompts that it generated for us in this case

[2:13:23] those prompts that it generated for us in this case Gema 4 12B. And here's another detail, to be displayed on screen according to the type of content. There's another detail, the structure that that particular model It generates images like Z image Image

[2:13:38] Turbo and according to the values ​​that we are of interest for that type of content. Okay? And if not, they'll just be images. if we are then going to transform them into Video, we can do two things. The First, have a single generic prompt

[2:13:51] that generates the image for us and that with that same image and prompt generates the video. A single generic prompt that From that prompt, it will generate directly from text to video. Or we can have a prompt to generate the image and

[2:14:04] a prompt to generate video, which Use these same things as context, but overall it should also have that is like an initial frame, as a scene, what was done as an image, so that

[2:14:19] From there, decide what will happen in that image once it has animation. Ultimately, it will depend on exactly what interests us exactly what interests us do.

[2:14:40] audio base And then we get into trouble And then we get into trouble legal. important. First, for ethical reasons, one must be careful with those things and then

[2:14:56] because the world isn't as big as we imagine. And despite seeming unattainable, like a certain person noticing and file some kind of complaint or whatever whether on the internet,

[2:15:08] What's happening? We want our The content is supposed to succeed, this is assumed to mean we are doing. So, yes, we are. Starting with zero visits, okay? But We want her to have many. If it were up to us, I would have

[2:15:21] millions, that is, if we achieve the aim aim that we tried at some point what we're doing Bad things have consequences. So

[2:15:36] primarily for ethical reasons, but not only not for ethical reasons, but because on the internet we They'll be enough when we're doing something bad. Because? Because we want it to work, right? Yeah It works, it will reach a lot of people and

[2:15:52] then many people can get involved If you're in trouble, be careful with that. But it's so simple, let's not be a little boy. Sometimes I've put this example. If we ask a small child

[2:16:05] Why shouldn't we steal? He's not going to say it's because he's bad, he's going to He's not going to say it's because he's bad, he's going to to say that because they put him in jail. But it's supposed to be when we already understand The world is better, not only because of the

[2:16:18] Consequently, it is because we should not do it. We shouldn't steal, it's not right. our. Well, the same applies here, let's not steal a tone of voice, however bad it may seem to us do. It's that simple, we can do it.

[2:16:33] cost the same amount of time, because Let's do it right. In other words, with something that make sense, something legal, a voice designed with AI, one of us and

[2:16:45] modified in Audacity so that it has another tone or whatever, anything. Perfect, that generated the CSV file scenes.csv. And this is where Let's start looking at more advantages of automation,

[2:16:58] There is an advantage to not watching and Closing GPT chat, copying and pasting some things to one side, sticking them on the Another one, saving files, now doing We already see an advantage even being just one project. Since there are many more, then more

[2:17:12] Automation makes sense because I run it and it works for me at 1000 projects that you have in the queue. But despite that, there are things here in and of themselves a great advantage with automation and It was that, not counting the head of the group,

[2:17:25] It was that, not counting the head of the group, eight lines, eight prompts, eight prompts sent automatically and answered automatically by Gema who were armed from all this with title and a hyphen pasted here and then with a

[2:17:38] another, then with the other, and then with the another, even though it is a single project, a The only video we're going to make is the one we're going to make alone with this part. Automation is already desirable and therefore much, otherwise, to top it all off

[2:17:53] save it somewhere or whatever, do So much manual processing makes no sense. for a single project. So if we enter this way... projects, let's look at scenes.csv now in this first project, which is the only one

[2:18:08] that we have. We have the beginning and the end of each time fragment of more than 3.5 seconds. As soon as it completes more than 3.5 seconds stops adding words. And so, more or less, this is it.

[2:18:21] number of words per fragment of time. And here we have it. Gema answered us, which in this case is the prompt we're going to use

[2:18:33] case is the prompt we're going to use I trust in him. that not necessarily the prompt that We have here to generate those prompts

[2:18:47] image turbo. We shouldn't think that way. Here It's not such specific details of the concrete. It is a generic form that It works like this with Z image Image Turbo and Furthermore, we can fine-tune it even better. depending on the model we are going to use, both

[2:19:02] for the text it will generate as well as for the image that will be generated from of that text as for the type of content that we are going to generate. So, This works, but clearly it can be done and it should be changed. The better

[2:19:15] adapted, better adjusted to what What we want is better. We're closed more CPP And now we need to configure the UI. And that's it. It's the end. In this case We opened Confi UI

[2:19:29] We opened Confi UI and then with the other file, which because True, the names perhaps should be be upside down because the voice one is That's the least of it. Hey,

[2:19:42] That scriptim images.pp will use Z image Turbo. In this case, just as in the previous live stream, which we don't have neither quantized nor optimized in any way neither quantized nor optimized in any way form, is the complete model

[2:19:55] and yet our GPU still works, But we will be able to test other models, We're going to do it, and quantify it. I only mentioned something important here. I only mentioned something important here. What does Cohen use?

[2:20:07] text to try to transform into Image, right? A language model He understands text and something else too, They usually have vision, so for Editing images can also be useful and This helps a lot. This model is very

[2:20:20] full 16-bit, which means it weighs like 8 GB. Let's see what good things there are It has good things.

[2:20:47] One advantage we have with these models so powerful today very good at drawing, but they have, in quotes, a lot

[2:20:59] intelligence. Here's why, because For example, it's a pretty good model. For example, it's a pretty good model. So with incomplete prompts, Nevertheless, there is something here that previously reasoned our promptar the

[2:21:14] image. So it's very likely many sometimes the result is the same is correct, not because the model image generator knows how to add, by for example, but because there is something here that Yes, he's smarter, a model of

[2:21:27] language. So, let's try this. just by reviewing this example to understand that it's not 100% either obsessing over perfect proms. The more accurate, the better, but if not Well, there's someone here who takes care of it.

[2:21:41] So there's not too much of a problem. necessarily with that.

[2:21:53] TRUE? It was only for testing, not stealing. only for the rich. They're confident of winning with better lawyers. I don't know if it's worth it.

[2:22:19] to get to do it. In fact, it's more about this.

[2:22:32] Notebook with the result complete of the sum of 2 + 2. complete of the sum of 2 + 2. Insurance. Make sure the number is there Insurance. Make sure the number is there of the result after the symbol of

[2:22:46] equal. In other words, we're here, but this isn't the goal. No? But it's just to give you an idea. which in itself can help us quite. In fact, it's more than that, just because... in all models, uh, in some that

[2:22:59] Yes, they have enough in themselves to function one language model, others not, more ancient ones, for example. Uh, let's see, They'll probably even give us some. Examples of that. I don't know, but Probably so, right? Well, it has a

[2:23:11] language model inside.

[2:23:23] not necessarily better at PROM We wrote all that down. Ah, here with reasoning. Here they are being presented as certain

[2:23:37] mathematical problem, but it is not telling exactly what to put in each partly, but simply given that by Yes, it has a model that in this case Yes, it has a model that in this case He can reason that it is Q in 3 4B, because

[2:23:49] simply with that data alone we may give the apparent result That's right, without needing us to give it so much detail, but this isn't the best way to do the prompt. Better that everything is already completely in place

[2:24:03] well explained to avoid possible errors. equals sign. But oh well, it's He understands the idea, but we've already seen that Nor is that the main objective, TRUE? But yes, it can anyway

[2:24:17] to produce those kinds of results. Right there we saw in the official documentation, even with a more complex task than this, But hey, we can generate images, However, we don't want to generate them. We're going to generate them here. Now

[2:24:30] The objective of this is supposed to be Direct means adapting things to work with different models and with different adaptations in the code. The only one The adaptation we have made is in the voice part. Really everything else

[2:24:44] voice part. Really everything else They're like recaps or nuances, huh? from previous things we have seen. So here we'll be able to do another adaptation. In this case, Z works. image Turbo. Let's do a

[2:24:58] adaptation to use Z image Image Turbo. However, Z image Image Turbo Turbo. However, Z image Image Turbo in GGUF format and quantized because in at this moment the Z image Image Turbo that We are loading,

[2:25:12] the model weighs almost 12 GB almost 12 GB together with Qen 34B which weighs almost 8 GB. together with the BAUE which weighs 319

[2:25:28] megabytes and we're trying to load everything That's on a 12GB GPU. Fortunately, she's 12, so when it tries to upload the file that It weighs 11.4, but it still fits. Herein

[2:25:43] With a GPU, we would have the problem with an 8-bit one. But what's more It's not just that in other places it might not It fits, we're pushing the limit here, but it fits. but It's uploading and downloading too many times.

[2:25:56] things because not everything fits at the same time time, thus making there more reading and writing to storage, which here At least it's an NVM SSD fast enough, but more neck

[2:26:10] still from the bottle and that surely in RAM if we have enough. Save in each of the files. If we don't have That much RAM is fine, but if it has That much RAM is fine, but if it has He takes advantage of the opportunity. So,

[2:26:25] Yes, a single complete file, like for opening and closing one and the other, at the same time time so that they weigh less, they are lighter downloads, whether apparently faster execution and

[2:26:39] that fits in all kinds of hardware, so let's see how We can load the model, for example, in GGUF format. and also quantified. images, but perhaps also the same Qen 3, 4B and other things that are

[2:26:56] use in a model. This is exactly what we're going for. to serve not only to execute it quantized, which will lose a lot of size and some quality, but a lot of size goes to lose, not only for that, but for execute

[2:27:10] faintune models that are more optimized, that perform worse quality, but maybe instead of taking one minute, taking 5 seconds, or what do they produce certain types of things better, or what I don't know, so that it's not exactly like that.

[2:27:23] not only the native model, but more things that we can vary. Come on Let's see then how this same We can use either model or another one. And there are Another thing here, it can't just be C Image Turbo, it could be some flux, some

[2:27:39] stable diffusion or whatever. And from In fact, some of these models not only generate images, They edit images. It may be that in everything

[2:27:51] We want the same scene in the video, but different things happen. or that it's the same character in different scenes. That is achieved by starting with image editing. This

[2:28:05] starting with image editing. This character now seated, now uh in the character now seated, now uh in the Eiffel Tower at night, now in a kayak, Eiffel Tower at night, now in a kayak, uh whatever. Then it's not going to be

[2:28:18] model of different forms, but because We can also load other models that even have image editing. As one is used that has editing Images? Just like TS. That have? Find this TS text and audio

[2:28:34] to begin with. Input text and audio to generate an output audio. It's like if we were modifying the audio to starting from that text. Well, this is going to be it starting from that text. Well, this is going to be it input text and image to generate

[2:28:47] that output image, that is, editing a image, maintain certain things, but modify others. And indeed, the edition images are not only useful for

[2:28:59] do consistent things with variations, but also to maintain the style. TO the best thing for what we want to have a entry image, it's not meant to be like that image, but maybe it

[2:29:11] That's the style we want him to adopt. TO The best thing, despite the context, is the prompt. indicate in a cartoonish style like this yasá and colorful and with lines lines thick and blah blah blah, maybe with That's not enough, but that together with

[2:29:25] a real visual example of how we want The result will be the same as The result will be the same as find TS. We can design a voice, but maybe only with a prompt the text doesn't look exactly like this

[2:29:37] we want, but starting with one of For example, yes. So we can use that too. both to use this optimized model as to use others and even to be able to use others

[2:29:51] functions within certain models such as image editing. Uh, anyway Here we're going to review and within our project, while we desire all this, The images have now been generated And here it is, precisely for a video.

[2:30:04] Where does the surname Baladés come from? These were the images. Spain, But be that, it doesn't represent us, the great

[2:30:17] So, the images are fine. drawn, but perhaps they are not the best for this type of content. So we can change the prompt,

[2:30:29] a lot of things, and that's not even before we've finished. We don't even see the generation of clips from video, but oh well. Well, if that's the case We have seen it, but not automated. Let's go and assemble it then, just to see how the final video would look. This is not

[2:30:43] It requires IE; this only builds things with FFMPG, as we saw in the two live streams FFMPG, as we saw in the two live streams previous. Perfect, I'll check it out. So if we come to the end, here all the final projects would be included,

[2:30:58] in this case only one. Let's take a look. Do you know where your Baladés surname? Even if you wear it today In your identification, this name was born in the land of Spain. It's a surname. toponymic, which means that

[2:31:13] describe a physical place. Originally It was linked to rural areas of Castile and Navarre, identifying the families who inhabited specific valleys or areas delimited by fences. Your last name doesn't It's just a name, it's a map of

[2:31:27] a story that defines your roots. This is already better than it was a while ago with the GPT voices, clearly. However, There are several things we can modify. Well, besides the tone of voice and blah blah, If we want to continue doing it, we can

[2:31:39] Add some background music too, not just the voice. But here I mainly think that what I would do most missing before adding other things like video or music clips It would perhaps be changing the style of the

[2:31:52] Image, right? Ah, so that's why we're going Time to change models, right? Before that, By changing the model, we need to adapt. our proms. Maintaining the model We must change, adapt our prompts anyway to the type of

[2:32:06] content that we want to create. And for For example, here this same model can generate other image styles, Therefore, simply

[2:32:18] We have open code, which doesn't even look good, oh well, we're going to tell him to modify scene punt which is the prom we have

[2:32:31] so that it generates the promises for the images and indicates in general and in detail specific that the type of image to be generated should be

[2:32:45] that the type of image to be generated should be style comic or cartoon with thick strokes, flat colors

[2:33:00] with thick strokes, flat colors and simple, uncluttered images information. Also, that it can Also, that it can display text when necessary,

[2:33:14] display text when necessary, but just one word to choose the wrong model, as we saw in the first video.

[2:33:29] And with something like that, well, you can already... in case it changes the prompt it will create the prompts. So there are two previous reasoning. The one who's going to put it together This is better and then that from this point on it will

[2:33:42] and then that from this point on it will to assemble, it will settle, it will arrive better to exactly the type of final prom that we are interested.

[2:33:55] such, such, such. Perfect. What do we have to do? because with our language model that Let's use generate again Let's use generate again from scenes because scenes are the

[2:34:09] The best prompt descriptions are the descriptions or prompts that we are going to use to generate the images maintaining the model. They only changed maintaining the model. They only changed the proms. We click on run

[2:34:22] And we wait. Thanks to those who are here. Many thank you. These live streams tend to be bored in some way. But I think they're interesting, huh? Very good. Just in case, and while

[2:34:34] They're running out of ideas about what things to talk about while we wait. Only Remember that for $100 they can hire so that I Do the same thing to you that we are seeing.

[2:34:51] In this case, it would be the channels of YouTube

[2:35:03] with 100 shorts each on whatever topic you tell me on whatever topic you tell me including codes. I deliver, supposedly in about a week. It can be more, it can be less, and me

[2:35:18] You can contact us by [email protected]. What do you gain from this? Hire someone. I'm not going to do magic, I'm not going to

[2:35:31] Probably yes, if you're just starting out beginning, but I mean, it doesn't mean that bad or whatever. It depends on each person person, but simply if someone me wants to hire so as not only to have everything the explanation and experience that

[2:35:46] discussed in these live streams and videos and looking at the codes and explanations and Furthermore, you also want me to himself will do something for you I would consider it well done so that Continue from there, providing examples.

[2:36:02] real and without needing to start from zero and also hiring that great zero and also hiring that great initial impulse, well you can do it. I repeat, it doesn't mean it's going to be most excellent and definitive. I'm not

[2:36:14] saying that it will look bad, it will So, how do I do it? Uh, with The payment is not because it will be the not because I am the best existing product, but because I I'll work on it for you. I emphasize a lot in this, not because it delivers results

[2:36:29] I do consider that, otherwise they might even sue me. daily. Uh, it's not because of that, but because Sometimes we get confused and when we see a We believe that the price is simply for paying It's going to get better. And you're paying for Someone do it, someone who

[2:36:44] Apparently, he knows, in this case I do, more It doesn't mean you're paying for a state secret, a unique form and definitive way of doing things. I'm going to Use the same example I used last time. It's like paying someone because

[2:36:56] paying? Because I cleaned your house. No necessarily means that he's going to do better than you. It just means that You paid him to do it, not you. The advantage here is that I say This, so some advantage may be possible

[2:37:09] to have for someone new. If not, well, all the videos are there. explanation for this fragment of time. Brilliant. Football has everyone entertained. True, true. Very good. Let's see the results.

[2:37:28] Where are they? And there are no images. images. We made the scene archive with the new proms, which we are going to use to generate the images.

[2:37:40] We don't use any fine tuning, we don't use parrots, we don't use any of that. We changed the prompt. Let's look at that same 6B model small, open weights,

[2:37:55] not necessarily the best, nor within those with open weights. How good is? And well, quite good, despite generating the images, right? mega resolutions, we're generating it a 720

[2:38:08] And with that, they look good. Hey, Then we click on run the same number of images as it makes a while. Because?

[2:38:24] subtitles, they remain the same mathematical calculations of time. What's going to change is what... It will show in each fragment of time. Despite that, we were able to to have changed

[2:38:40] how long each scene lasted, making it whether more or fewer images are acquired, that change more often or less often. It generated an image, then it will generate The next one, then the next one, and we'll see.

[2:38:57] that in total the first one took almost 33 seconds to generate. This is one of The important stages, yes, but There are other things before and after, 33 which means that already in here We should have the first one there

[2:39:12] It's more of a cartoon style. In fact, It turned out very well. However, in this Regarding the text, we see that it doesn't It included many details. And here the The problem isn't necessarily that there is

[2:39:25] including many details the model of language, but he asked for too many details generics on air without detailing them, and then the image generator model was invented

[2:39:38] image generator model was invented the text that's here. precisely because of the prompt, because the prompt does not explicitly mention what prompt does not explicitly mention what text,

[2:39:58] text, even though I had been here one-word position, that is, Since this doesn't work, it had been as very specific, but despite That wasn't enough because we It would be useful for the prompt to say the

[2:40:12] screen and in what form. Not here He doesn't mention any of that, so he did basically random text. But we see that style, for example, In this case, we got it right the first time. It looks pretty good. Only it

[2:40:28] What we failed to emphasize was the part of the text. And here precisely there is someone else who can help us. Who Is it AI? In what sense?

[2:40:44] What was he saying? All because of reading the phone. Ah, yes, he can help us. In what sense? First, this is not a drawing. Hey,

[2:40:59] language model, which can navigating between files, between and so on. We can do it, if this model has let him analyze the images himself generated and compare them against the prompt

[2:41:15] generated and compare them against the prompt and try to modify it, update it to try to correct the problem that There is something in the text that requests text, but in a generic way. It doesn't say explicitly what text to display. Just in case

[2:41:29] Only by understanding the whole context, he Even looking at the pictures and everything, it can better adapt the prompt. us telling him what things to check, not whatever comes to mind. Probably It checks codes that have nothing to do with it, but yes

[2:41:42] There we are seeing that it is a change of total style based solely on what, had you indicated another prompt. With the which model are we going to change for others reasons, not necessarily because this It doesn't work

[2:41:55] uh, for the style that interests us, but also for the model size, the speed of the I don't know, or for other functions such as editing, such as we were saying. Perfect, they were done.

[2:42:10] as we were saying. Perfect, they were done. all. TRUE? about the text, but because the Prom doesn't say it, not because of the

[2:42:23] model. So, now that we have this So, That's not the goal. Uh, the The goal is to see more adaptations as well in terms of models and all that.

[2:42:36] So, in the confidence that we already we had, we had had custom nodes and In fact, in the first live broadcast we had Z Image Turbo is installed, but quantized, if I remember correctly.

[2:42:51] What we don't have is the Wflow for adapt it in our code and that adapt it in our code and that work with that model. Hello. Hello.

[2:43:04] Install user Z image Image Turbo Confi UI language model. That's how it is. We were using it quantized up to the Quen model, for example, in FP8.

[2:43:20] So we already had that. What not What we had was that SD academy workflow adapted so that our code works with that

[2:43:32] another workflow, not the one we have. Here, which is the original. So, what do we need to do? Basically, giving the AI ​​context, this file, this one here and the workflow we want

[2:43:48] in exchange for this because the files and installation are already there We had. So, First, we need that again. SD academy workflow that does not We have it here.

[2:44:02] We have it here. Well, let's go look for SD Z Academy Well, let's go look for SD Z Academy Image Turbo

[2:44:21] Wordflow, which in this case does not Let's skip straight to AI. We see that It's huge. This one weighs 2 KB, this one weighs 17. Why? Because this one has all the visual aspects that is, the exact coordinates of

[2:44:35] where it will be located, what size and The rest will have the nodes, I've compared against our tiny workflows. So, the first thing that interests us

[2:44:49] same, correctly specify the files, in This ZM Turbo case in Q4 km, just as it we had previously downloaded. complete model, but I mean, no quantized, but here we want it

[2:45:03] execute quantized. We removed that from once. We have the complete 34, but We also have quantization and BAUE. It is not quantized, it weighs 300 megabytes. Hey, So with something like this over here We've got it more or less ready for the

[2:45:18] That interests us, right? That because It is true that it takes eight stages, not six. We had seen that too. And once we set 720 by 1280, the vertical resolution that we have indicated. And now export file API

[2:45:33] We already have it with the values ​​that we It's interesting, but not only that, it's also interesting We're switching to a format that's very heavy. less. So there are several ways to make this work.

[2:45:45] First, recommended, new conversation. This previous context does not It's useless to us. And we're going to open and closing because it didn't look right. Uh, that one The context was for other things, so Well, a new context.

[2:46:01] We're going to tell you, analyze it and explain it to me how it works Zimagurbo.jonjon in conjunction with images. Pi. With that

[2:46:15] Well, he'll understand. Now, how does he Shall we transfer the other file? We can put it directly into the same folder and mention it. We can give you the route from the other side in where it is or we can depending on the

[2:46:31] drag it and have it right here. So we're going to tell him, analyze this So we're going to tell him, analyze this new workflow that I'm sharing with you and respond with a plan of what changes will be made they would have to do

[2:46:48] zimageturbo.jonjason and and images.options with this new workflow and models.

[2:47:06] We're interested, he knows how it works, he We're passing on the new stuff, which we want We're passing on the new stuff, which we want that now use this here stuck in This point Jason, and so we'll get to it. to give the possible answer. Perhaps the

[2:47:20] At first I sometimes failed, but well, We say, "He made this mistake, correct it." It no longer delivers monetized channels. I have never delivered them monetized. Never. Because? Not because I say so yes or no, but because it is impossible

[2:47:36] a lot of things. First of all, what How is it monetized? With the same YouTube or externally? Using YouTube itself, how within YouTube. But Besides that, that's what the person is for. You need a Google AdSense account

[2:47:49] name of that person, with details of that person, registered from that country person and that in time he may also receive the PIN. from Google verification to its home. So I've never sold channels

[2:48:05] monetized, not because I don't want to, but because it can't be done. the option to skip them. Good afternoon, Mr. Juan. Hello. Hello. And that's despite being... deactivated. Ad automation disabled.

[2:48:20] YouTube and its stuff. I don't have the blame. When I want to put up ads I'm supposed to click here, I which I never do, by the way.

[2:48:32] So, we're going to tell him, Here it tells us the changes, replace The complete content of Z Image Turbo for the new one and do some adaptations in the code of which nodes which ones. So let's say,

[2:48:46] implement those changes. Meanwhile, let's look at other things We are making small adjustments, But this isn't the only way. adapt, modify, but rather enhance, Adding some things, removing others, just like the

[2:49:02] generation of video clips. But There is much more we can do. Let's go There is much more we can do. Let's go Show an example here.

[2:49:22] which is already over, so we removed the images and we click on run to see that works with the new model quantized. It should clearly weigh less And it should work. Hey,

[2:49:39] versions, one local and one online for certain parts of the process, leaving interesting things we have, such as example but after the images generate video clips from the

[2:49:53] that? Because a generating model of images makes better images than a that the initial image, the frame initial be with a generating model of to be able to check before even transform it into a video. But also

[2:50:08] This generates a thumbnail; there are prompts. also for this purpose in the other parts of the process and all this because this is designed for long videos, therefore designed for long videos, therefore The prompts changed.

[2:50:23] video generation. and another one which is image generation, But for a miniature, not for a part from the video. And there is another very important detail It's interesting what I added here, which is that the same script, besides being much more

[2:50:37] long, the script that will also generate It includes certain details, certain labels that I give possibility to writing model so that each part The narration in the script should indicate what

[2:50:52] The form must be narrated. with the objective that once the The script will be generated from all these details and the topic that we're going to give so much of Information however we want, uh then

[2:51:05] There will be a label, text, label, text, label, text so that when Let's generate the voice so that it is generated each one of those blocks of text by

[2:51:17] separated, but starting from a corresponding instruction as I the corresponding instruction as I the I want that label. in question. So this for a long video that will have more changes

[2:51:30] of rhythm, of tone, and of things to say, He can say things sarcastically, things of ironically, normally, in a documentary, I don't know. And here's another one. It's a very interesting camera. This is not just takes a tone or a rhythm in question

[2:51:47] how to narrate that part of the script, but that also the other parts of the code What they allow is that when you... camera label, it's not just that it goes to generate video clips from a prompt in text or prompt and image

[2:52:01] prompt in text and image to video, but that also from that file of audio and a reference image such as This one's around here. making it so that when you are

[2:52:15] label, here we see that it no longer occupies as much as a while ago, when you are camera tag It will be used as input for that video transformation section, the prompt in text, which is a specific one for

[2:52:31] video, not for image, it's another additional. The image is from around here. additional. The image is from around here. the voice file so that with those things generate that video clip of the person speaking to camera

[2:52:48] saying that part of the script. With the Which is for certain types of videos, certain types of narratives in which there is no We just want background voiceovers and clips generics, but also a person speaking to

[2:53:02] camera every so often, because it we can have. And this image, because For example, it could have been generated, there edited with what, with another of the same model, edit, as some model might be such as Flux Context or Flux Claim, which

[2:53:17] Let's see. We'll try it right away. Flux Claim. Because? Not only because you can edit images, but because it's smaller, it's lighter, it's because it's smaller, it's lighter, it's faster, but you can make more mistakes.

[2:53:29] Version 4B. There are versions 4B and 9B. We are currently using the Image set We are currently using the Image set Turbo 6B. Let's look at that other thing then. So there are a lot of things that They can be done, right? Uh, perfect.

[2:53:48] much has changed. There are about 30 of them seconds on average. They used to be seconds on average. They used to be Like 32 or 33. Uh, at least really low. these same adjustments and so on. Nor is it that the duration has decreased significantly,

[2:54:01] the time required by each generation image, But what is true is that, in addition to that It decreased a little, you don't have to be loading and unloading things like it does It hasn't even read from the hard drive for a while, and that...

[2:54:14] It will fit with a smaller GPU. No problem, because we're occupying a while, keeping practically the same quality. Let's review the quantization It makes you lose a lot of size, but little

[2:54:27] quality. We're not even going to talk about quality. not necessarily notice. Well, maybe here not necessarily notice. Well, maybe here Yes, the text makes no sense at all. we had discussed. In other words, they are still the same proms. The proms not the

[2:54:40] We changed. We see that in terms of image quality, no It changed, but the file size did. They are 4 bits, they weigh a quarter more or less. the model files

[2:54:54] compared to originally It did not generate exactly the same images. Because? Because it changed the seed and that kind of thing, but in following instructions, quality It's quite similar, only we went from

[2:55:15] a model that originally weighed almost 12 GB half. The same applies to the text encoder, instead one that weighs almost 8 GB, one that weighs

[2:55:29] one that weighs almost 8 GB, one that weighs 3.24 GB. adaptations we can make. Uh, yeah So we have here in this case everything Locally, everything optimized, everything ready to operate locally,

[2:55:43] automated and light and quantized. Gem 4 Uh, quantized 4 bits, in this case the Uh, quantized 4 bits, in this case the version 12B, the same Q3 this TS, which of

[2:55:57] which doesn't make sense here. The model of It is very small by itself, 1.7B or 0.6B, It is very small by itself, 1.7B or 0.6B, although it could also be quantified. Whisper also unquantized because not That makes sense, it's very light.

[2:56:10] Gem quantized again. Z Image Turbo quantization and the arming of final video. Everything 100% local and optimized. And to press run and that will generate hundreds or thousands of videos,

[2:56:25] images, uh, automatically and for free in local, always the same, and of good quality. consistent. We use it locally. Already Not because of, but because of quality, because consistency, speed, and because they are

[2:56:40] the right models for what we want to do. So now we're going to try another model. Let's look at, for example, flux clain and then we went on to do video clips.

[2:56:57] Diffusion 3.5 Large Turbo is one of the faster and the SNEL flux is still there, But both of them are getting a bit old-fashioned. No They were the smallest in size, although they're not exactly very big either, But they were fast enough, just

[2:57:10] that make too many mistakes compared with more current models. Hey, Flux Context came out, which allowed us edit images. We currently have Flux Claim in two versions, one version 4B and a 9B version.

[2:57:25] It's not just that 4B is less than the half the size, but this It has an Apache license, it has a very permissive license. The 9B

[2:57:37] Theoretically, if we follow the rules, we won't. Commercial use is permitted. However, it doesn't exactly make us 4B is more than enough, it's just that He's going to make a mistake. They are both wrong. A little more than Z Image Turbo, only

[2:57:53] image, not just text, therefore We can edit images, just like here. We are watching. Images may be being edited, combining the one thing of an image with Something else from another image, what do I know. So

[2:58:10] which is smaller, faster, you can edit images, but He makes a few more mistakes. It's something to take into account. Uh, actually, we're going to Try here and see if it's true that we They allow you to do something quickly.

[2:58:35] to see that the entrance is there. No is it not just text like a prompt, but There may also be optional an image. This is how you find this TS.

[2:58:48] There must be a file for the entry audio in our workflow. of our project, yes, because we are using it as a basis, as an example of tone of voice, but one of them could perfectly well be used the ones that come by default with everything

[2:59:03] certain prompt. We can have the same here. input image, generate one from scratch or input image, generate one from scratch or We can edit an image. Now instead of a black background we're going to tell

[2:59:16] in his rocket. Let's see if he continues. correctly, because the promo, The PR makes no sense. We are running online without adapting it in And because maybe even PROM should be in English and with other details

[2:59:30] specific. So let's try it this other model. To achieve that style, I thought a Lora was necessary or That's what the AI ​​recommended, but those It is necessary to change model. Even if we change

[2:59:43] model, anyway we have to change the prompt. Because? Because each The model understands things differently, it They ask for different things. So if everyone I have to change the model, then Let's start primarily by changing.

[2:59:56] If we need to change the prompt, Let's start by primarily changing only the prompt. If that wasn't enough to change the prompt, we change the model, but usually the prompt is something too important, but not to such a degree that

[3:00:10] He wasn't trained, but anyway He was trained for many things with Lots of different kinds of images, so it's something to keep in mind. Okay, let's go So try this Flux 2 Clain now, the same quantized.

[3:00:25] This model is 4B, the Z image Image Turbo is 6B. images. I want to see how many of these stages work.

[3:00:38] To make matters worse. This one's going to be several times faster architecture, but oh well, there are fewer. billions of parameters. Instead of 6 Billions is 4 billion, but also in Instead of eight stages, it requires four.

[3:00:53] It uses less memory, it's faster. It also edits images, but it doesn't offer much. worse quality. When? when He makes mistakes, but everyone makes mistakes and When it goes well, it goes very well. as in the others. In other words, it's not a

[3:01:05] such an extreme problem. Except that we want to make 1000 images and not Review them, because it's best to use Z Image Turbo. because he will make fewer mistakes. Uh, with We won't notice the difference with just one. With 1000, there will be several wrong ones. Hey,

[3:01:18] With 1000, there will be several wrong ones. Hey, Let's go to quantized

[3:01:31] 2.6 GB. Perfect. Now, what about the workflow? And What other things does it need to function? Will everything be explained here? No, not here It's designed to run in Python, the complete non-quantized model.

[3:01:45] Here we have two options, just as we have done it on previous occasions. Can Install it, test it, and prepare it for us. workflow, It remained. After she herself did evidence even she herself seeing the

[3:01:59] correctly or watch a tutorial with us. In other words, either or watch a tutorial with us. In other words, either Does AI see it, or do we?

[3:02:14] the only one, but certainly the most important, Well, it's going to be on a UNET. Commonly when The diffusion models are in the format ggf, that is, they are quantized, they go in the Unet folder. Now we'll search for the file there.

[3:02:31] Furthermore, and now we're going to try with Codex app again. app again. We're going to ask the AI ​​again.

[3:02:44] Open Code to see that there is a There are many ways, there isn't just one. definitive. I just want to see that there isn't one here. folder names that I'm not interested in show.

[3:03:04] And another thing is using GPT. I mean, We can use Codex App, but with another local model, for example, But in this case, with GPT. Let's take advantage of it.

[3:03:20] our project, because here there is no We don't want to change anything. We don't want to change anything. What we want to do is in confidence download the corresponding files and try it.

[3:03:39] It's better to start directly on that one. folder and let him fend for himself from there from here, although honestly not here The information is there, it doesn't link us to knowing which model, knowing that it is quantified and with that you can review it

[3:03:53] quantified and with that you can review it others. Hey, come in

[3:04:06] I downloaded that model in Q4 km and I put it in the Unet folder within models, here in Configure it.

[3:04:27] Research the other things on the internet. that I'm missing and compare with that I'm missing and compare with my UI configuration.

[3:04:41] everything and proof. To test you will need To test you will need run configuai.

[3:04:57] install correctly and make it work Flux 2 Claim Flux 2 Claim 4B

[3:05:16] and have generated at the root of the current route. the workflow to generate an image new and the workflow to edit a

[3:05:29] image. There is no need, I repeat, for the model of smarter language. What it does The missing element is that it needs to be connected to the context. and necessary tools. It's no use to me.

[3:05:42] I will use Cloud Facible 5. from cloud.aiaiaii. In other words, it's a very powerful model. Yeah. but executed in a place that has no access to my PC components, nor to to be able to execute or anything else. So, here

[3:05:55] GPT 5.5 is very good, yes, but The main issue is not the GPT, but the to be able to access that URL, to be able to navigate among my files, the ability to search in internet, the power to run and test

[3:06:09] things on my PC, being a model with vision to review the images generated and the ability to write files to generate the two workflows resulting. It is clearly required that the model is sufficiently

[3:06:22] intelligent. So that? For giving it those possibilities But the most important thing is Give him the chance, because if not,

[3:06:37] It gives them the possibilities, neither too little nor too little. He won't be able to do much because, well, no has access to my system. So, The main thing here isn't the Codex, eh, it isn't The main thing here isn't the Codex, eh, it isn't GPT 5.5, but Codex App,

[3:06:49] which would even allow me to configure Codex App to work with my models Locally, but it doesn't make sense, no That's the goal of this live stream.

[3:07:05] generates the workflow. That's right, but... We want two, one for editing and one for Generating from scratch, it's not going to be the same. And anyway, it may be that in Compress the image or what do I know about it? until we eliminate it too

[3:07:18] metadata, it will no longer have the workflow including. But yes, commonly when we generate with I trust UI something

[3:07:30] The metadata includes the workflow. workflow. I don't see it here.

[3:07:48] We dragged Confi UI, huh, what's the point? including?

[3:08:00] I'm going to check first what you already have in models. I've already seen an important point, it's in the portable version container. I also confirmed that the GGUF version of I trust. In other words, he checked that the file

[3:08:15] question. Now I'm going to work inside. from configuring the gufq4 km. Now, if it's look for. Existing models by text encoder and installed nodesf. Your model is located such and such. You also already have

[3:08:30] the custom node with fgf and two quen 34B encoder candidates, the complete and quantized, at 16 bits and at four. What remains to be verified further Careful is whether you have the right buy flux 2, which I don't have. This is it

[3:08:45] flux 2, which I don't have. This is it have. Uh, this other file, right? The bu already contains specific nodes. ta It has CUDA running on an RTX. 3060 12GB. That's right on the edge. He This model requires approximately 13 GB of V RAM

[3:08:58] small tests with minimal download to validate operation without forcing 1024 by 24 input. Yes, it requires 13 GB to go comfortably the model without quantify it, but we are running it quantified up to the 6B quota of Z+ Turbo.

[3:09:13] This one even more so. The workflow Confi official for Claim 4B. Use bla blah blah blah blah. The encoder is already in your folder. I'm not going to download the buy only that official buy. The official buy is It's going down. The download continues. The buy

[3:09:26] It's at 61 MB. I'm going to look at the names exact input before editing files. Uh, I'm going to create two Wflows APIs small at the root, one of text to image and another for editing with a entrance image. The two will remain

[3:09:40] I left both workflows. In When the by is over, I start with FUI. Start with I went and try both workflows that he did. Once the

[3:09:52] Try checking the results, which everything works correctly. Otherwise, it will correct it until it is correct Okay, that's it. Uh, so here only I want to check something quickly, taking advantage.

[3:10:14] What I wanted to show is that I also arrived at the same Codex App Modify Codex App so it works for me with local models, although it doesn't Codex App is needed to modify Codex App, but that's also possible.

[3:10:31] We see that the download isn't going very well. Quick, to be honest, but oh well. I don't know how much that buy weighs, but it says that Almost ready. Number 3 won't download it

[3:10:47] Almost ready. Number 3 won't download it Because that's it. I will reduce the resolution if there is a lack of memory. It's not going to happen, there won't be one.

[3:11:00] lack of memory. There was no room until 6B without quantification. less than a 4B without quantification. less than a 4B quantized.

[3:11:12] The download has finished. It's not downloading My PC does almost nothing. By completed. Now I lift up, I will trust which exposes the corresponding nodes.

[3:11:24] We're almost ready. Meanwhile, We can check right here that we have the two workflows that we He did, which at least as a basis are Okay, but clearly we'll be able to modify things. This one, for example, is

[3:11:36] text to image. This initially only has text. Indicate the files corresponding. That's right, including my file. And I already had this one too. Not this one He downloaded it.

[3:11:49] He put a prompt there. It will generate the image at 512 by 512. And once I generate the result, well Then he will download and review it. AND The other workflow is for image editing. It's going to be practically the same, but

[3:12:04] also having as an input a image. image. It's already undergoing re-escalation.

[3:12:17] Because he didn't put it here, he didn't save it. as with an example image, a route temporary, because we don't see exactly Where is it going, but here simple name that is supposed to be there where the image would go. So

[3:12:32] Theoretically, it's fine for him to try and see. If that's true, and if it gives an error, then look for the If that's true, and if it gives an error, then look for the mistake. confirms that he is not trying to treat it like a normal safe tensource. Perfect.

[3:12:45] Text and image generation completed. I'm now sending the editing workflow. What he didn't do was check the image. resulting, but let's do it ourselves meanwhile. Yes, there it is. This is when the text attempted to generate

[3:13:00] an image. Now I suppose he's going to use that same one to edit it. Oh, no, he used the which comes from an example. An image of an example that's drawn haphazardly within Confi and here he gave it a style realistic. Surely someone has seen

[3:13:13] This image, but the original. Well Okay, it worked. I'm going to locate the exits, uh, validate Wordflows exist in the roots and leave the session clear. Confirmed, there are generated images. I close the settings and

[3:13:26] perfect. And just like that, not even We necessarily have to complicate things. us. We see that things can be done with Codex App, with Open Code, with whatever, simply by giving it the correct context

[3:13:38] to the model. It is usually more than enough. Perfect. So,

[3:13:50] downloaded and things set up necessary to function. We can then load confidence and in this case of these WFlows

[3:14:02] edit. We can try the editing one a little bit, eh, from Confi, from the a little bit, eh, from Confi, from the browser, but add to to our code. In this case, no. We need an original image that

[3:14:15] Let's always be editing, at least not for this current project. With the which we're going to replace, but the one from Text to image, convert from Z image to Image Text to image, convert from Z image to Image Turbo to Flux 2 Clain 4B unedited.

[3:14:28] We will edit it because I only went from Confui to test it a little more. Perfect. Here we have configuration And then we're interested in image editing. And then we're interested in image editing. For example,

[3:14:44] original, the one I mentioned, a strange drawing. And the final was over here, that only that example image put a realistic style.

[3:15:03] rescaled, Because if we upload a 4K image or... Whatever it is, it's not necessarily what we want. It is of interest as an example, it reduces it. And here We see an example of a prompt. This example prompt is surely among them

[3:15:15] This example prompt is surely among them all the URLs that Codex App visited do it. We still don't know how it will be They use prompts to generate an image with They use prompts to generate an image with This model, and even less so for editing.

[3:15:28] But editing image one doesn't make sense. Image 1. Image 1 as if there were more than one. Okay. Four steps. What do I know? I'll be here. final image.

[3:15:50] We're going to put a GPT chat. Investigate how How are the prompts made? How are the prompts made? for flux 2 claim 4B for flux 2 claim 4B that allow you to edit an image.

[3:16:03] Then, once I understood how It's done correctly, that is, We can do it right, or we can do it If it's bad, then do it right, right? Here goes information that will help us. and we're going to In other words, write a prompt in English that

[3:16:20] allow you to do the following. Hm. Change, edit the image to give it a style edit the image to give it a style realistic and well detailed. With all and

[3:16:40] and sparkles. What do I know? At this moment the entrance image Well, it's like a little drawing. We're going to want it to be practically the same, but with a realistic style. And one more thing,

[3:16:55] We're going to make the resolution of the an image that makes sense, It won't be completely different Makes perfect sense. It is, and where is the resolution?

[3:17:17] total resolution amount. Broad, high, high, obtain. Ah, starting from What is the entrance like? Is it how does it work? exit. Even if we could disconnect

[3:17:30] We'll enter this manually. as I understand it. So, if I remove this, could select. Ah, well, it's there good. That's how it goes. linked, because not Nothing unusual happens with the resolutions. strange. Perfect. So, once

[3:17:44] who investigated everywhere here and already He understands how it's done, because he's the one who helps us. give the prompt. And this silly thing, in Instead of writing a prompt, it seems that I wanted to create an image, but oh well, In the end, he did.

[3:17:57] We insert the edit promo, the image of entry to have a realistic style, highly detailed in appearance, huh improvements, blah, blah, blah, whatever.

[3:18:11] negative to indicate what we don't want in the image. However, not all Image-generating models do In that case, we would have to pays attention to what we don't want it to. appear, to what yes and to what no we

[3:18:27] Little ones don't have that. And there it is. We see that it is practically the same We see that it is practically the same image, practically the same, only in This case with a more realistic style and with more lighting, just as we requested.

[3:18:40] It has a sparkle, it has a shadow, what do I know I. And this being a model, in this I. And this being a model, in this Case 4B, which took a while to edit the image 14 seconds. The other one was taking a long time It takes 30 seconds to generate an image. This

[3:18:54] Editing it is taking half the time. But Because I just realized something, the final image does not look the same resolution that the input image, stay with the resolution to which I pass here the

[3:19:08] things. But Here are two details then that we can take into account. If we are going to adapt the entrance image, if we leave it very The entry image is tiny, if you read it

[3:19:23] She won't be able to see it anyway. correctly and the final image will to look bad, because from the outset the model He really saw a terrible image quality and output we can also change. Maybe the entrance is

[3:19:36] the exit is larger. For example, 720 isn't really necessary. These are Image Turbo, only in reverse for vertical,

[3:19:48] Except that there it's for generating, not for edit. ask the Codex app itself, which was

[3:20:00] adapt, because here it seems that with that It wasn't enough. Anyway, it It wasn't enough. Anyway, it It brings resolution from elsewhere, But hey, there we see that it works. No? Both for editing and for

[3:20:14] trigger. A while ago. Let's try Again, the generate button and that's it. But Again, the generate button and that's it. But Yes, the editing one works too.

[3:20:27] thank you so much. Edgar mentioned it, that if we drag an image It also opens up the workflow that was used based on the image because it remains in the metadata of the same.

[3:20:49] But they have it by default. Uh, very good. Let's try the one to generate good. Let's try the one to generate images from scratch. from scratch. To begin with, there is only text.

[3:21:04] of the proms we already have prepared to generate images. Let's see with those same prompsa, because It's a change of model and it should deliver the proms. Anyway, here another thing. We see that we can

[3:21:20] further reduce the size required in GPU because it's running this model Complete accuracy, but we can manage also quantized, that is, it can be like this We won't notice any improvement in quality. difference.

[3:21:43] This image. It's clearly a different model. because even though he is making the drawing cartoon style, like we said, huh It doesn't do it exactly the same way, But it's more or less the same style. Let's try 1280 * 720.

[3:22:01] Let's see how long it takes to generate one. image compared to the other. 2.89 seconds. Why the first time Did it take longer? because we were barely there running and we had changed the model, the tex encoder, with which it

[3:22:15] He had to open it, he had to read it. In This case was already open and we see that We're in the same boat. We need to look into that. But in this

[3:22:30] We need to look into that. But in this In this case, a 512 by 512 image took less time. In this case, a 512 by 512 image took less time. 3 seconds. execute.

[3:22:42] Prompts would work for us. The problem in This is the moment to listen to the This is the moment to listen to the resolution.

[3:22:59] something. Oh, and here it is in two places in the that needs to be changed. Basically, it's like the canvas and what it transforms truly in image. Let's wait and see just with that. We see several

[3:23:12] These are more technical things. By For example, this is so that in a Technically, follow the prompt better, but no It is necessarily good. We see that it generates images very well, a exception of the text. This is not a

[3:23:27] Very good model in text. If Z Image Turbo isn't, this one even less so. Z Image something to say in the text, but this no longer It's very good. What is true is that the style The visuals follow it quite well. It's not exactly like Z Image Turbo

[3:23:41] because it's not the same model, but in itself The images are good. and much faster, so it's something to keep in mind.

[3:23:56] use the root image as the only option which is going to change scenery in scenery? The way you've done it, no It's even more amazing. Hm. It depends on exactly what it is. whatever you want to do. Like, why?

[3:24:08] You would always be sending the image initial? It's just a bunch of extra computing that can be unnecessary for certain processes

[3:24:24] root that will change scenario in scenery. Seeing, but there isn't just one way in which it can be done. Can do it with an initial image that serves context to edit it. It can be with

[3:24:37] Prom is based on the previous ones, can be totally proms independent, as in this case, there are many ways. Okay. So, now Since we have this, let's adapt it. exactly how we'd like to put it

[3:24:51] to our workflow. We want that model, that quantification, that buy. Yes, we want to This resolution, but in this case in 720 by 1280, In other words, it has changed, reversed vertical.

[3:25:07] Maybe this is even convenient for us right now Doing them in full HD works so fast. like this. File, Export API, Confirm. And for

[3:25:22] For example, let's go back to Open Code. Hey, given that we already have here the Our project folder, right? Of trust

[3:25:41] how they work. Then we'll pass it on this one here and we tell him to adapt.

[3:25:53] and we tell him to adapt. images. Pi and Z image Turbo for images. Pi and Z image Turbo for work with this new workflow that you work with this new workflow that you I am attaching it.

[3:26:10] Basically, we'll have to replace the Z+ Turbo content. and at the same time it's going to do some adaptations in images.p

[3:26:27] according to the workflow of this new model. What is the content or the path of the New WFlow Jason, what do you want to use? Already I pasted it above; it's in a file. Ah, Okay. He literally wants to copy and

[3:26:39] I don't think it's necessary. You can write literally all the content, but literally all the content, but Well. There it is. That's it. Maybe it's just It will copy and paste.

[3:27:07] consistent, face and clothes. Ah, in that one If you do require an initial image, Just like we saw a while ago with a robot, I, a cyborg me. Yes, it depends. In this type of The project we are carrying out is not

[3:27:22] It is needed. In fact, more than being Positive would be negative, let everything be Positive would be negative, let everything be equal. But for something like this, yes. But it depends. We don't need it here. to a character throughout the entire video.

[3:27:36] In fact, it would be too monotonous video in that way. That's how it is. Yes, in the case of an avatar, but Not here. Okay, then

[3:27:50] Both files updated, replaced completely by the new workflow and adaptations. adaptations. Now you have a fixed negative,

[3:28:02] But I don't think even this model... It works with a negative prompt. It doesn't matter, We're going to delete the generated images and then we click on generate. And now should use, even though the The file is called Z image Image Turbo,

[3:28:14] What's really inside is flux. That's how it is. We can clearly change the That's how it is. We can clearly change the