🚀 VEO3 est HORRIBLE / 7j pour lancer une app avec l’IA / pubs coréennes.

Frank Houbre - IA & intelligence artificielle

Bienvenue dans ce première épisode Build In Public, où je partage sans filtre les coulisses de mes projets avec l’intelligence artificielle.  Cette semaine :  - Comment j’ai lancé une application en 7 jours pour le marché sud-coréen, - Les pubs générées avec l’IA (et pourquoi VO3 m’a rendu fou 🤯), - Les outils qui marchent (Gemini, One, Eleven Labs…) et ceux qui déçoivent, - Les hacks pour trouver ses premiers utilisateurs et avis (Reddit, SEO coréen, etc.),  Et les galères d’IA pour mon animé et mon autre app Lunesia.  👉 Ici, pas de bullshit marketing : je montre les résultats réels, les fails, et comment rebondir quand une IA vous plante.  ⏱️ Chapitres  00:00 – Intro & objectifs de la semaine 00:18 – Génération de pubs IA pour le marché coréen 01:07 – Taegyo : l’app prénatale développée en 7 jours 02:10 – Traductions & pièges du marketing en Corée 02:48 – Gemini vs ChatGPT pour le coréen 03:15 – Création d’illustrations IA cohérentes 04:24 – Blog Naver & premiers téléchargements 04:46 – MVP et stratégie pubs VO3 05:26 – Exemple de pub coréenne 100% IA 07:05 – Les ratés catastrophiques de VO3 (fail compilation) 09:24 – Montage vidéo & musique IA (Suno, Eleven Labs) 11:22 – Les limites de ChatGPT-5 pour l’animé 12:57 – Test de One, Dreamina & alternatives 15:38 – Projet Lunesia : pubs, avis & conversions 🚀 18:21 – Hack Reddit pour obtenir ses premiers avis 21:16 – Optimisation des vidéos & stories enfants 23:31 – Résultats & apprentissages de la semaine 24:32 – Conclusion & appel à vos retours  🔥 Hashtags  #IA #BuildInPublic #Startup #IntelligenceArtificielle #MarketingDigital #AppDevelopment #AItools #Publicité #Animation #VO3 #GeminiAI #ElevenLabs #SunoAI #Lunesia #Taegyo #SEO #RedditGrowth

Transcript

00:00Welcome to this first episode where we'll see everything I've accomplished this week thanks to artificial intelligence.

00:05We're going to see a lot of things and I don't think I've ever seen a service as bad as Google's VO3.

00:12But nevertheless, through perseverance, we are now able to produce quality advertisements.

00:17So we're going to see how I managed to generate an advertisement for the South Korean market to promote my new application.

00:23Then we'll take a quick look at the launch of the new ads.

00:27And then why I also decide to change artificial intelligence from time to time and I really recommend this one if you want to do animation, especially 3D characters.

00:37We will also get into the heart of the matter a little, see how the application is made, how I proceed precisely throughout the week.

00:45Anyway, here we go, good episode!

00:47So as I mentioned, we just released a new app, we developed it in 7 days.

00:51This application is aimed at the South Korean market and it allows for what is called prenatal education thanks to scientific studies which are quite solid and thanks to a traditional thing called taegyo in South Korea.

01:06The idea is very simple, a pregnant mother will be able to have a meditation session in Korean with her child and the text said in Korean will be repeated in English and also later in Chinese to accustom the fetus to the different senses.

01:19And this is something that is very scientifically validated as you can see right here and has been for decades.

01:26Luckily, my partner speaks South Korean fluently, but I absolutely do not.

01:31So, I have to establish a whole marketing strategy and also create all the visuals and all the content even though I don't speak the language at all.

01:38So you might say, it's pretty simple, we go to the classic sites, we right-click and we translate.

01:43However, as you can imagine, it's not that simple because the translation will be imperfect, there are pages where we won't be able to translate.

01:51So in reality, you have to rely on what you know about websites, UI and UI.

01:58Of course, that will be a button, it is easily distinguishable.

02:01But there are things that are sometimes a bit tricky, like for example, the validated or confirmed button is always on the left and the canceled button is on the right.

02:08So I had to write blog articles in Korean on Never, which is their search engine and the reference for blogs, while accepting and drawing inspiration from the different codes which are very different from what you can find, for example, in Europe or the USA.

02:25For example, in SEO, it doesn't really exist. We're not going to work on our SEO with H1, H2, H3 titles.

02:33We'll just put text that's sometimes smaller or bigger.

02:36So quite simply, as soon as I have a question, I just have to take a screenshot and drag it directly into the GPT chat to ask it to translate the text for me and it works pretty well.

02:47After testing and comparing, it turns out that Gemini is much better at understanding the South Korean language.

02:53So I'm going to use Gemini instead for this one.

02:55Then, for everything that concerns the generation of precisely this type of image that we can see right here, these small illustrations,

03:02I started with a small style that I composed myself and then I will be able to ask the artificial intelligence ChatGPT to also create images that are similar and to be inspired by the different characters.

03:14I just have to load all the images I have taken or a part that interests me anyway,

03:20to give them to ChatGPT and then ask for a new illustration.

03:24I'm going to have a success rate of, let's say, 50%, so one image out of two will be good.

03:31I may need to rework the colors due to yellow contamination of the colors.

03:36If you don't know what it is, I invite you to watch my other YouTube video which talks about AI contamination

03:42and why we have the impression of having a yellow filter on these images.

03:46In this way, I was able to deploy a content strategy while of course doing research.

03:52and by having images that are homogeneous and therefore correspond to the South Korean market.

03:57To make life easier, we also went to the OpenAI playground and created a Naver Blog Writer.

04:06So quite simply a prompt that will allow you to easily generate blog articles by chatting directly with the chat.

04:13Then I just take the result and ask Gemini to check whether the South Korean that was written is of good quality or not.

04:22This way we were able to get our first three ratings and our first downloads.

04:27It's not much and we obviously have to go beyond that because we want to make this application a success.

04:32And we therefore call this application model an MVP.

04:36It's a very basic and standard product that we developed as quickly as possible.

04:40with not all the quality criteria at the maximum but which will allow us to test the market.

04:45So, in order to have new installations, we have to create new advertisements.

04:49That's why I turned to VO3 because we wanted to create funny, entertaining advertisements, with audio.

04:58But the fact is that everything you're shown on the internet about VO3 is perhaps the top 1% of what works.

05:05So, I'm going to show you the final result of this advertisement.

05:09And then you're going to see all the videos I had to generate that were absolutely catastrophic.

05:14And it's all very manual, but I really enjoy going back to basics and starting everything from scratch.

05:22and to be able to create lots of things using artificial intelligence.

05:25The video we decided to make is, of course, aimed at Korean mothers who are pregnant.

05:31The goal is to create a very dynamic ad with a lot of overhead

05:35where we show the mother trying to do everything she can to make sure her pregnancy goes well.

05:40Then, of course, the accent, solution, relaxation, relaxation with the arrival of our application.

05:47I'll just show you a little excerpt.

05:48I'm going to talk to you mainly about this scene which was hellish to do and I'm also going to explain to you why.

06:12So, there's a lot of audio information, but it's something that's done a lot in South Korea.

06:16and alongside that, we generated advertisements that are much lighter.

06:20And everything you've seen here, in terms of videos, audio, and images, is artificial intelligence.

06:25Even the voice was generated by AI.

06:28Here's another ad, much shorter, where we want to do something fun,

06:33where dad is showing off and trying to say the word peanut butter,

06:37which I myself cannot pronounce correctly.

06:40And in the end, it is the child in the womb who will correct the father.

06:43Then you have the whole ad rundown.

06:58So, we could very well say to ourselves when seeing the VO3 results on the Internet that we can generate this in just a few clicks.

07:03Well, I can assure you that, in reality, to get this type of result, you need to generate a lot of videos.

07:09Because the AI does a bit of everything, and yet the prompt is optimized to the maximum.

07:13And here, what I'm going to show you is just a small part of the VO3 generations.

07:17And you can see that I redid the scene, the scene, the scene, where she eats my dish,

07:22dozens, hundreds of times to get the result I wanted.

07:27And the prompt has been validated by ChatGPT, by Gemini and by myself.

07:31So actually when you generate something on VO3 you're going to get this kind of result.

07:37It has nothing to do with it.

07:39The result I want is that the mother is eating her plate of spaghetti.

07:43She is happy, but the cook tells her no.

07:45And replace the dish, precisely, with something that doesn't look very pleasant to eat,

07:50but which is normally healthier for her.

07:51Here, it's a failure.

08:01What is he doing with his leg?

08:04I'm not even talking about this one.

08:11Or from this one, from the beginning, it's already missed.

08:14And it's pretty terrible, because we generate the videos, we expect to see a result.

08:23Each video costs approximately €1.30 to generate.

08:26And we end up with this kind of very fragmented result.

08:29And yet, we are with the ultra and VO3 quality subscription.

08:34So, in the end, by dint of launching a prompt, it starts to work.

08:38You should know that you might have to cut a small piece of the video.

08:42Too bad, it doesn't matter.

08:42This one, for example, makes no sense.

08:50We wanted it to be absurd, but it doesn't make sense for the character to appear directly from a pile of sand.

08:57All these videos were generated with VO3 quality and also with Flow.

09:01Flow will allow you to generate, for example, your first video here and reuse the characters.

09:07But be careful, again, when you are going to make an extension of the video,

09:11you won't necessarily be able to do it with VO3 quality.

09:14He will ask you to go to VO2 because he does not have the capacity to do that currently.

09:19And it's extremely restrictive.

09:22In short, once all that is done, you just have to take the videos and edit the video in the classic way,

09:27which is done relatively easily with this software called Filmora.

09:30This is the one I use and if you are looking for a free alternative, I definitely recommend DaVinci Resolve.

09:36You will be able to do almost anything you want for free.

09:39So we see that I put text elements, we see that I put sound elements that I used,

09:42music that I also generated with Sono.

09:45For music generation, I went to Sono, precisely, and ChatGPT.

09:49And I asked ChatGPT, create me a music that will allow me, for example,

09:54to have a nice little sound, presentation, startup, etc.

09:57Or funny sounds, and so on.

09:59And then I was able to get my files here, and I uploaded them.

10:03Also, as a side note, I tested Eleven Labs Music this time,

10:09and it completely crushes Sono.

10:11It's expensive, but it's amazing the results you can get.

10:14So for example, I was wondering how I could make a nice intro for my anime.

10:19I just did one test on Eleven Labs, and it crushed everything I did on Sono.

10:27However, it has some problems, it's at the pronunciation level.

10:46I invite you to take a test in French, in English, it doesn't matter, in your native language,

10:51and you'll see that it has a very Americanized accent at the moment.

10:54But the result is impressive, and especially for the instrumentals too, I found it absolutely crazy.

10:59In short, for Tegyo, the advertisement is ready, all the visuals have been done,

11:04The Never blog is ready to post articles.

11:08Namely, last little aside, I used ChatGPT5 to generate images and visuals for myself

11:14to advertise Tegyo, and it was really extremely bad.

11:19I don't know what they did with ChatGPT5, there are things I really like,

11:24but by wanting to focus everything on a single model, we have the impression that he is doing absolutely anything,

11:28and that it does not necessarily exploit the right data every time.

11:31Please feel free to tell me your feelings about ChatGPT5,

11:34because for me, for the moment, he's putting a lot of thorns in my side.

11:38Besides, I explained in one of my videos where I was working on the anime,

11:42I'm going to give you a little update here,

11:43that I used ChatGPT to redo the images and characters.

11:47Since the ChatGPT5 update,

11:49I am totally unable to redo my characters correctly,

11:54and to have a style that is consistent.

11:55And yet, I source it with the images, etc.,

11:58but after a while it goes off the rails very quickly.

12:00So, I'll have to find a new alternative,

12:03and I still found one that might be interesting.

12:06What is a bit annoying, however,

12:08is that the anime itself,

12:10the first episode was almost ready,

12:13as we can see right here.

12:15And as you can see, there were some super badass moments.

12:17But if I can't rectify the situation,

12:37Well, I'll have to do it all again.

12:39And this is something that is likely to happen quite often with artificial intelligence,

12:43but that's how it is, those are the rules of the game.

12:45Because even when training, for example, a LoRa,

12:48so really a model who will be trained to copy my style,

12:51I am currently experiencing performances that are not good,

12:54and results that are not good either.

12:57So, recently, I tested Juan,

12:59who made people talk about him precisely thanks to his video generation.

13:03And I was quite surprised by the quality of the video generation,

13:08but also images.

13:09Like, for example, I did some tests for my anime,

13:11and we can see that the animation is rather high quality.

13:13And similarly, I used it to do some testing with advertising.

13:18Here we see the South Korean woman again

13:20who is doing some kind of yoga and meditation exercises.

13:24And the results are pretty good, and I find it very realistic.

13:28Before, I broke my teeth precisely with Ailio,

13:30where the results are not bad,

13:32but nevertheless, as you can see, I had quite a few worries.

13:35And Ailio is much more expensive than Juan.

13:38Likewise, its runtime is much longer than Juan's.

13:42So I started turning to Juan,

13:44and maybe in another video you'll see me turn to another service.

13:48This is absolutely normal,

13:49because everyone is racing to get the best result,

13:53and the best AI.

13:54So as you go to work,

13:56you need to learn a general method of working with AI,

13:59to be able to be very flexible,

14:02and be able to bounce from one tool to another.

14:05Ailio, to give you an example of cost,

14:07it costs about 100 dollars a month,

14:10whereas here you might have the equivalent for 20 or 30 dollars.

14:13So from an animation point of view,

14:14I think it's very good,

14:16and I also went to explore SIDEM,

14:18what is Dreamina now.

14:20It's nothing more than the artificial intelligence of TikTok and CapCut.

14:24It belongs to the same parent company.

14:26And there, the few tests that I did precisely,

14:28with my characters and my manga,

14:30I ended up with some really interesting things.

14:33Visually, it's good,

14:34It's really very animated, very manga.

14:37And there are not too many details,

14:39as we could have seen for example on Mide Journée,

14:41who does anything.

14:42Midday, in fact,

14:43He makes you some really beautiful pictures for the manga.

14:45It has the Niji 7 engine now,

14:48or Niji 6, excuse me,

14:50which allows you to specialize in anime.

14:53But its big problem,

14:54is that it generates so many details in every direction each time,

14:56that he does not know how to be consistent.

14:58So in reality, using Mide Journée to make a manga or an anime,

15:01it's just shooting yourself in the foot,

15:03because you will never have anything consistent.

15:05So there, I say to myself,

15:06there is still something to do with this engine,

15:09and I'm quite happy to take it all back,

15:11to be able to rework my anime.

15:14Nevertheless,

15:14Reusing characters on this service does not work.

15:19And if I give this character back to Design,

15:21an AI that I like for training characters,

15:24or ChatGPT,

15:25well he will always modify the style.

15:28So for now,

15:28I'm a little stuck,

15:30I'll have to retrain some artificial intelligences

15:33through the aura,

15:35with for example Automattic 1-1-1-1-1-1,

15:37I'll explain that to you later,

15:39how it can work,

15:41and that will surely allow me to unblock the situation.

15:44But for now,

15:44I'm completely stuck on the anime.

15:47We will now move on to my other project,

15:49who is called Lunesia,

15:50and which are stories precisely for children.

15:52Quality stories,

15:53with a character that is consistent,

15:55beautiful voices,

15:56and stories that will bring value to children.

16:00And the big problem we had when launching the first advertisements,

16:04it is that of course,

16:05we didn't necessarily have a lot of notes on the Internet.

16:08So people coming to the app page were like,

16:10well, not bad this page, not bad this application,

16:13but there aren't many notes,

16:14I don't install.

16:15Fortunately,

16:16I was able to go and get it,

16:17notes on the left, on the right,

16:19by canvassing.

16:19And there we end up with 23 notes,

16:22and it suddenly starts to work,

16:24and we will see that conversion rates have exploded.

16:27We see that before, I had a conversion rate,

16:29we see that before, I had a conversion rate

16:31which alternated, let's say, more around 2%.

16:33And now it has drastically increased the average conversion rate.

16:37That is to say that all the people who arrive on the page of Lunesia

16:41are now ready to install the application.

16:44And according to Chad GPT,

16:45These are very good conversion rates.

16:47This will now allow us to understand whether people will like the application.

16:50and will then experiment and test it.

16:55This will allow us to validate the MVP again,

16:58and then we can develop the application

17:00and do a big fundraiser,

17:02or finance it with our own funds.

17:04To make this application a success.

17:07This is why our thinking is not about generating money now,

17:12but above all to work on the entire application tunnel

17:14so that it can grow and attract the right people.

17:18So there is no point in us making money.

17:21because we finance it with our own funds anyway.

17:24What we want to know is whether people are actually satisfied with the application?

17:28and are ready to press the “Validate” button.

17:30So currently, rather than asking to pay for example a month or a year's subscription,

17:35Well, we're offering everyone a free year right away.

17:38who are currently registering for the application.

17:41The application is in French and English.

17:44This then allows us to see user journeys,

17:46to tell us "OK, people like this story

17:49or they don't like this application?

17:51Or "They don't like this story?"

17:52This allows us to have data

17:54and then, when we are sure of our shot,

17:57Well, we'll put the previous offers back

17:59and we will be able to see if it converts and the project will be validated.

18:03And you're going to tell me how I managed to find it

18:05the first 20 ratings and reviews?

18:07Because at the beginning, it is very difficult to convince people

18:10who don't know you to test an application

18:12and of course, to give their honest opinion.

18:16You are not allowed to pay for reviews, it is prohibited.

18:19Well, in this case, I simply went to Reddit.

18:22And when you create an account with Reddit,

18:24At first you can't post anywhere.

18:26Because in order to be able to post in certain places,

18:29you need karma points each time.

18:31And these karma points take a long time to get.

18:34And there are some super easy little tricks to do.

18:37There, as you can see, I have 452 karma points,

18:4029 contributions.

18:42And I was able to get all these points

18:44which actually allow me to unlock conversations.

18:47I will be able to ask people to test my application

18:49and give me real feedback.

18:52Except that it's a bit like the snake biting its own tail.

18:54because you don't have karma points,

18:56so you can't post.

18:58And if you can't post,

19:00you can't have notes, for example, on your app.

19:03But to be able to post, you need karma points.

19:06So, you have to manage to find the places

19:08where you can post and get traction.

19:10So, how to have, for example,

19:11100,000 views on Reddit very easily,

19:14have 980 upvotes, etc.

19:17And have the karma points

19:18which will allow you to post on Reddit

19:20and then to launch your start-up ideas,

19:23to have your first customers.

19:25Well, there you go.

19:26Just post a picture of a cat

19:29and ask people to comment

19:31and to find a name for this cat.

19:35In this way,

19:35you're going to have lots of people

19:37who will upvote your post,

19:39who will comment

19:40and who will see your publication.

19:42Or you post something funny,

19:46for example this video here, on this channel,

19:48which allowed me to have 61,000 views

19:50and therefore have a lot of extra points.

19:54It took me 5 minutes to do this

19:56and it unlocked absolutely everything for me

19:57to get the first notes.

19:59And what's super interesting,

20:00it's that now,

20:01I can go to lots of channels

20:02and show the application,

20:04comment on people's Reddit posts

20:06who are looking for history apps

20:08precisely for children,

20:09showing in white paste

20:11because my account is legit,

20:12because he has karma points.

20:14There aren't many karma points

20:15on this account,

20:16but it already allows me to do what I need.

20:18So you've seen it before on One,

20:21when I showed the interface,

20:24I generated quite a few videos,

20:26especially for Lunasia.

20:28In fact, before,

20:28I had generated a lot of videos with Ailio

20:30and I had problems

20:31where the character kept talking.

20:33So I told him,

20:34No, don't talk.

20:36And Ailio made me spend a lot

20:38to have videos

20:39and a lot of headaches.

20:40I had a lot of waste.

20:42While there, with One,

20:43I can put a very simple prompt

20:45and I really have the behavior that is expected.

20:48These videos,

20:49They help me illustrate the stories.

20:51But as you can see here,

20:53the quality is already very very good.

20:55But to be able to increase precisely

20:57the quality of these videos,

20:58I slip them into this service

20:59which is called VMake

21:01and which allows me to have a video

21:02which is almost in full HD

21:04in a much better quality video.

21:07If we look here,

21:08It's already very high quality.

21:09But if we look here,

21:10it's even better defined.

21:12And that's good

21:12because it's curls

21:13of 5 seconds of videos that I make.

21:15So I can download

21:16all these videos absolutely free

21:18and generated a lot.

21:20What do I do next?

21:20with these videos?

21:22Well, we'll have to create a magnifying glass.

21:24So play the video in one direction,

21:26then when it ends,

21:27iron in the other direction.

21:29For that,

21:29since we are on Mac,

21:31I created a folder

21:32with what I call the Loop Builder

21:34and I put a file in it

21:35in .sh

21:36that I will be able to execute

21:37and it will read the source files

21:39and he will assemble them

21:41and optimize the weight.

21:44This execution code

21:45here, here,

21:46in 7 pages,

21:47right here

21:48and I didn't do much.

21:49I just asked ChatGPT.

21:51So these were the O4 Mini-I versions

21:55that I used

21:56to code this for me.

21:57It did it to me in a few moments.

21:58I just fixed 2-3 things

22:00and I just have to execute it.

22:02In this way,

22:03all the videos I have

22:04inside this source file

22:06will be converted into magnifying glasses

22:09automatically here.

22:10So I don't need to put this up

22:12because it's really super simple.

22:14I just then

22:14all that's left is to go to Cursor

22:16which allows me to do so

22:17to develop the application

22:19with my partner.

22:20He's the one who takes care of it.

22:21But to make command lines

22:23which will allow me

22:24to send automatically

22:25all these files in the right place.

22:28This allowed me to

22:28to correct more than 80 videos

22:30where there really was

22:31quality issues.

22:34What could have been problematic

22:35for people

22:36who see the application

22:37and who want to rate it.

22:38So, it was absolutely necessary

22:39rectify all that.

22:41Likewise, we realized

22:42than children's stories

22:43which are in .json format

22:45and which look exactly like this,

22:47you have every time

22:48a video, a scene,

22:49the sentence in English,

22:51the sentence in French.

22:52Well, sometimes,

22:53they were too long.

22:54So I asked Jimmy Knight again

22:56to retire precisely

22:57all books

22:58and he did it to me

23:00in a few seconds.

23:02Which allows me to have

23:03stories that are shorter,

23:04much better indeed

23:05for children

23:06and which are more efficient.

23:07That's what you can do, by the way.

23:08see right here.

23:10I give it to him every time

23:11the json file

23:12which represents a story

23:13and reoptimizes it each time

23:15with the name of the music

23:17that I gave him,

23:18of course,

23:19which was generated on Sono.

23:20Then, the elements

23:21which correspond

23:22to what we have defined

23:23with my partner

23:24in the application

23:24and of course,

23:26all in one English

23:27and in French

23:28which is perfect

23:28and which is linked without bugs.

23:30In this way,

23:31I really have stories

23:33of quality

23:33that bring value

23:34to the children

23:35with a nice voiceover over it

23:37that we generate there too

23:39with Gemini

23:40and I just have to do

23:41these few manipulations

23:43which really allow me

23:44to do things

23:45who might have taken me

23:45months to go

23:47all alone

23:48if I had wanted

23:48to master

23:49all these professions.

23:50So here it is,

23:50I suggest you...

23:52So here's a little bit

23:53for the news

23:53and this is how I work

23:55so daily.

23:56I invite you

23:57to be tested precisely

23:58the Lunasia app

23:59and to make me

24:00feedback on it.

24:01It would really be

24:02super important to me.

24:03I'll leave you the link

24:04in the description

24:05and all the different services

24:06that I showed you here

24:07are also linked

24:08in the description.

24:10I hope so anyway

24:10that this first episode

24:11allowed you precisely

24:13to see a little

24:14everything I have accomplished

24:15in this week,

24:16to inspire you,

24:17maybe to unlock

24:18situations

24:19and to see that,

24:20Well, there you go,

24:20you have to jump

24:21of an AI tool

24:22to another,

24:23you have to understand

24:24its operation.

24:25Sometimes,

24:25we lose everything

24:26like me for example

24:27to animate it.

24:28We must start again

24:29and it's still

24:30for the better.

24:31Don't hesitate too

24:32to talk to me

24:33of your problems

24:34in the comments.

24:34my real desire

24:36with this type of video

24:36it is to do

24:37an exchange with you

24:38and unlock you

24:39exactly

24:39on everything related

24:41to artificial intelligence,

24:42on topics

24:43of project creation,

24:45artistic,

24:46companies,

24:47never mind,

24:48I am at your disposal

24:49in the comments.

24:50Take care of yourself

24:51and your projects.

24:51Hello!

Catégorie

Transcription

Écris le tout premier commentaire

Recommandations