We talked to Freepik's CEO about their pivot to AI, going unlimited, and more!

0:00Everything that we did, in my mind, became opposite. It was going to be opposite, like, in a couple of years. So, to me, it was, at that point, AI crossed the threshold of having creativity.

0:14In this special episode of Denoised, we're speaking with Joaquin Cuenca, CEO of FreePick, one of the largest AI-powered creative suites, serving over 100 million users. Enjoy. All right, Joaquin, thanks for joining us. Really appreciate you coming on. Thank you, both of you. I'm very happy to be here. So why don't we just get the backstory of FreePik, how it started, and a bit of your backstory? Well, the backstory is a little bit longer than the usual backstory of an indie startup. We actually started in 2010, at the end of 2010. A little bit before FreePik, if I can go far down, I made a little startup in Spain.

0:51And I actually was acquired by Google. It was the first startup that was acquired by Google in Spain. I was working on geologated photography. I went to Google and I spent three years there. And then I grew as an engineer, kind of. That's where I learned a little bit how a big American company works, all the titles, how different people, how it's structured. Then I came back to Spain and together with two friends, I started a new startup that's Freepeak. And at the very beginning, it was only a search engine. So that's why we call it FreePick. Okay. It was, we're doing websites, finding images.

1:30It was kind of a hassle. We were looking for like free assets because we were working just, you know, on a budget. So we created a website that was like Google, but to find free images on the internet. That was a big one. Okay. Then, you know, one thing led to the other. Like we started expanding. soon enough the quality of free images on the internet was not great and we started making ourselves images to improve the library and then we did icons and we did more things you know we keep moving on and the business was doing very well it was expanding everybody was happy i guess there were unicorns and butterflies all over the place you know everything was nice until uh 2022 2022.

2:21You can guess where this is going, right? So in 2022, I got like, the first glimpse was DALI, DALI 1. DALI 1 really blew my mind, you know. Before DALI 1, there were a few things going around. You got gas, you got some other technologies, but it was always like quite niche, you know, it was not yet working, really working. And in my mind, honestly, I didn't think it was going to work. I thought it was always going to stay a little bit not as good as humans. Like the AI image quality would never be as good. I thought, and that was my mistake, I thought it was going to be a barrier between human creativity and machine produced images of something.

3:06I mean, when you look at how Ritz is a real photography, with makeup, photos that may be sold to have an impact, and then what you were getting back in the days, it was, to me, there was a big divide. But Dali one, that was the first one that showed, like, my gosh, how can it be, you know, you start picking up, like, from words, you pick up, like, compositions, that it was, like, quite complex, okay, to my, at least at the time. So that really lied about, they say, okay, we should have stopped working on this, maybe this is going somewhere. We started, like, I started hiring some people that were working on AI, initially doing just improving our search engine, doing tasks that we knew that could be solved.

3:51Like no blue sky research. Okay, just, okay, let's start working, so to say. Let's just start learning a little bit about AI. And then Nalito came out. And that completely blew my mind. That was, I have a very vivid memory of what it was. I was with COVID, isolated in my house for one week, a little bit with fever. and so the only thing that I could do is look at Twitter and look at the things that people were posting. It was blowing my mind. I was completely, some people don't like me saying this, but I was completely in panic. Like imagine, like at that point, the company, we were like 500 people in the company.

4:32A future that, you know, it was kind of, you couldn't see the progression of the company. So we knew how to do illustrations, that we did icons and we did photography slides we got the motion figure out now we're gonna do 3d videos you know more things and now this thing was like like it was a big reset like uh the future is not gonna look like the past Joaquin what was the cause of your panic the cause was that at that point there were two things like everything that we did in my mind became obsolete or it was going to be opposite like in a couple of years. I saw, to me, was at that point, AI crossed the threshold of having creativity.

5:15Some of the images that I saw was like, oh, this thing, it really took some imagination to do this thing, okay? And that was, to me, the big mental barrier. So once that fell down, you know, I didn't saw any other limit. It's like, no, this is just going to keep improving until it becomes as good or better. It was clearly superior to the old method. And you saw that back in 2022. Yeah, yeah, yeah. That was, to me, it was like when it, yeah, when it crossed, I saw some images that really took imagination to make them. And to me, that was the big barrier, you know. And the second thing is that I even saw how could I add value.

5:58It's like, okay, this is going to make images. I don't have AI experts. so what's next you know what about the 500 people that working with us like how do we deal with this uh i think it took me up until dali to to realize that this was the future because of the situation where i was in i had to cope with something that was to me i think you know looking backwards is that why did it take me so long kind of i think was in part because of where i was okay it's like having to deal with the company, just not being able to accept the future in a way. I think I still realized before some others in the industry, kind of.

6:42Definitely. But I think I could have figured out like one year before if it was not because of where I was. Did you get pushback or did people think you were crazy for kind of pivoting to AI and so early? Investors, yes. Yeah. Shareholders, yes. I mean, our company is Bootstrap. So we started in Spain. There are no VCs, no private equity. We were just like three friends. I was calling, someone that was doing design. But in 2020, my co-founders, they were quite stressed by the company. It was also their first exit. And they said that they wanted to get out. And we got a private equity that bought part of their stake.

7:23So they got a controlling stake in the company. So that's why we have shareholders, but we didn't have any investors. They thought I was a bit crazy at the beginning. I mean, they didn't. So they were throwing analogies like, you know, but self-driving cars, it was supposed to come like in one or two years and it's not yet here. Like maybe this is the same thing. And in my mind, I was just, okay, whatever, no, this is really coming. But people at the company, we have many founders in our company. So we acquired some other companies. Most people saw it, okay, this is changing. This is really changing.

7:57the big question was okay but what can we do in this world how can we add some value okay and the answer came actually just by working just by iterating in a way you know instead of asking the big question that makes you feel a little bit dizzy it's like i don't know what the future is going to look like but i know what's the next step that i can take today okay the next step became pretty obvious. It's like, we don't have the in-house jobs to do a foundational model, not today. But stability came out. Stability initially with Stable Diffusion 1.4. It was an open source model. Okay, that's something.

8:37We can take that and we can make a website just hosting it. Step one. Okay. And then people started using it. And they told us that it was a piece of crap because they could not easily select a visual style. You know, people were adding to the prompt in the style of some random guy that is an artist and all that. And of course, most people didn't know how to craft their prompts. So V1, we just added some visual styles and we're just appending text to the prompt. You know, it's very basic stuff. And people like that. And then some newer, faster models came out and we made like a fast version. And then we did a sketch tree image.

9:17And doing websites, that's what we do. That's something that we could do, okay? And we just iterated. We just keep working. And we got, like, more needs from the users. Some of them we could do them. Some of them it was not in our hands. So we just send that feedback back to the model providers. And I think the only, you know, the advantage that we have compared to others, that we had already a huge traffic. We were one of the most, we were one of the top 100 most traffic websites in the world. That's amazing. It's a pretty decent traffic. We got over 80 million monthly active visitors. In terms of stock websites, we were the most trafficked one in the world.

10:01Yeah, like the traffic that showed their stock almost. So at least we had traffic. We had people that were interested in images. And we had many people that were paying our subscription. So when we made some AI proto studio at the very beginning, we got plenty of people using it. So that gave us the ability to negotiate a better deal with the model providers where they get very cheap inference. And we passed those savings to the users that gave us more users. And it was a combination of that early distribution with an eagerness to improve the product that led to more people, cheaper prices, and to keep the ball rotting.

10:45Yeah, definitely. In the early days, I mean, the model manufacturers weren't really making money, and you had a revenue stream. Yeah, we actually, you know, we've never been margin negative. So the margin on AI is worse than the margin on stock. So margin on stock, you create the image once, and then you sell it. You know, you can sell it by zero almost. The gross margin on the stock, on our case, it's 70%, roughly. That's correct. We have to pay the author of the image. And then, of course, below that, you have to pay salaries, everything in the company. But it's 70%. On AEI, initially, it was very low.

11:30We're giving it at cost almost. So it was from zero. But then models started to be much cheaper very quickly. So image generation went down from three cents, started going down to one cent, below one cent. It got down to 0.1 cent at some point. Now more like 0.2 because we improved quality. But it came down in price very dramatically, very quickly. So we went, I remember gross margin went from zero up to 95 % on image in a few months. and now we are at 20 % on video, but it's starting to go up. So as technology improves, of course, humans, we don't improve or, you know, we just need to keep our margins.

12:22But with technology, you can get an increasing margin over time. And what are some of the numbers now? I think, was there a tweet recently of like the number of images or videos you're generating a day or a week? Yeah, I mean, this has been changing over time. Actually, in the past, we generated more images because at some point, we enabled automatic generation of images when you scroll. And that generated many, many images, but they were of low quality. We started pushing higher quality, even if the latency is a little bit higher. And now we are at three, four million images per day. and on videos we are getting close to 200 000 videos per day which is biggest it's one of the top three worldwide that's that's a lot of volume that's a lot of gps did this go up recently too from uh running the the free video generation promos like with different models yes yes so Video generation, it increased.

13:22That's interesting. So far, all the unlimited images and unlimited videos thing that we have done, we have only done it to people that subscribe to Premium Plus or Premium Pro. That's a subset of the subscriptions. Premium is the bulk of the subscriptions that we have. So we have on Freepeek above 700 ,000 subscriptions. And then we have other projects. in total is over 800 ,000, but on FreePic.com is 700 ,000. And roughly 85 % of them are premium. And then you get 10, 15%, which are premium, pro and plus. And we offer it to premium, plus and pro because A, they're premium more, so it's only fair that they get first any advantage.

14:11And B, it's a way to A, B test how things are going to behave. Okay, when we expose it. I mean, my intention is to bring unlimited to everybody eventually, but I need to calculate what's going to be the cost, what's going to be the reaction. So on video, we introduced recently some unlimited video, Wix. It's not yet limited forever, like with Image. And since we introduced it, the total number of videos created in the platform increased by 50%, 5-0. That's pretty amazing because unlimited videos are only exposed to 10 to 15 % of the users, just as users increase 50 % the total video generations.

14:57So there are clearly some super users that are just all the time generating. There is people that came to the platform to generate video disproportionately. They go to premium plus and pro because when you generate video, that's pretty expensive. That's pretty intensive in credits. So if you are serious about video, almost all of them are in plus or pro. Yeah, let me just back up for a second, too, because just in case people weren't familiar with. I realize you kind of jumped into the credit thing, but just in case people weren't familiar, it was like a month or so ago you sort of made this move that was um unique in the ai space where you said if you're on premium plus or pro uh you're getting rid of the credit model for ai image generation and it's just unlimited image generation go crazy make as many images as you want you don't have to worry about credits uh and then you've been running a weekly ish promo with different video generation models for like a week or limited time people could generate as many videos as they want on a specific bottle and then that kind of rotates that is that is right of course like you know the universe not being unlimited uh every unlimited thing uh it it always has some kind of limit okay so uh we have not put any explicit limits but there will be a limit i mean we have a limited capacity at the end so someone has their computer hacked and running images 24 7 Exactly.

16:28At some point, computers become slower. You know, there's going to be a queue. But so far, we have been able to handle the load. It was very interesting because when we're thinking about how people use the AI Suite, it's like, how will you feel if every time that you use Photoshop, I always use that example, if every time that you click somewhere, it was like, this is two cents, this is five cents, it will piss me off. It will make me think twice before I do anything. And it was that feeling that people were kind of constrained by that. It's not so much the price, but knowing that you have to pay every single time, we didn't like it.

17:11The flip side I was going to say was the Red Lobster example. If you remember, the endless shrimp strategy backfired and the company lost a lot of money and a new CEO had to come in. I mean, how do you balance your cost with unlimited? And, you know, something that I learned on my time at Google, there was a story when I joined Google that is Google, at the early days, you know, they knew very well their math. They knew very well how much electricity they would consume. Google initially was only a search engine. They work in memory, so they consume an awful lot of memory. You had this story where they came in into data centers, They signed a lease, and electricity was for free, traditionally.

17:55You pay by the square meter. So they packed a rack with servers. It was crazy. And they knew, you know, when they plug in, they had done the math. And those guys, they knew that the data center will go bust in six months. Because by contract, they had to pay for electricity. It was unlimited electricity, you know. And at some point, they were getting with the computers and the racks. they had wheels because they knew that we have to take them out in six months uh it was it was that bad okay so i'm you know that thing really put me a warning always on unlimited when you go for unlimited so first of all as i mentioned if somebody starts abusing the system like there is always an asterisk against unlimited within reason okay if somebody starts using it from 150 different countries in the same day, you know, it doesn't work.

18:51We have a limit. You can only use it in three different devices at the same time. Okay. We did the math and we thought that worst case, the number of images generated will double. And we knew that because there is a model that we are using where we reduced, we divided the price by 10. That was Flagstaff, I think, because our costs had been keeping down. So eventually we just adjusted our prices. And when we divided by 10, usage increased, but not that much. So we thought that was just a psychological barrier that people will use more. Maybe it was going to be double the previous usage. And in our platform, in total, image generation is like 25 % of the cost.

19:39You know, video generation is the lion's share. And we were only offering unlimited to premium plus and pro users, which is like 15 % of the users. So it's 15 % of 25 % of the cost. That's something where you can take a bet, you know, and learn from that. And if it works, great. If it does, I, you know, I became convinced that that was not going to kill us. It's something that we'll be able to stand by, to keep it, and eventually to expand it to other users. and we were pretty close to the reality you know that's roughly what happened so that's amazing it's it got me to upgrade from premium to premium plus and it's it's been nice to um not have because yeah you always have that psychological barrier even if you have a big pile of credits you're like uh you know it's something so but it's been nice to just be like well we could just generate and explore and not worry too much about all the images we're making and just kind of go at it.

20:36And I'm sure even as many as we're making, it's still not really like a dent in the big scope of crazy usage or being too much. When you're also thinking about the, just AI creation in general, FreePick has sort of become this central hub, a variety of AI tools, models, image, video, lip sync, editing, modifying the images you generate. How are you thinking about the interface and the interface evolving with AI image creation and sort of AI video generation. Some people think about FreeBig as a marketplace of models, and I don't like that mental image because I actually think that the UI that we build is the real product.

21:18My mental image is just like when we started having personal computers, you know, the deep technology thing was the CPU. It was something that was very difficult to make. And all the layers on top of the CPU, that looked like amateurish stuff that anyone could put into that computer. It was something simple compared to making a CPU. Or writing the software that ran that computer was something very, very simple. It was just, you know, an application level on top of it. Eventually, people realized that that's the actual product. Like the interface between the user, it's the thing that the user actually uses.

21:58and what is behind the hood is almost an implementation detail. And that's how I think about FreeBik. We are not in our team, we are not just throwing models to the platform. In general, we test the models. We want each model to add something new. So when we add it to the platform, it's because it's better at animated illustrations. It's better at something specific that we know that our users are going to find valuable. There are some models that we don't find better than everything that we have in any dimension, and we don't add them to the platform. The goal is not just to add everything and that result.

22:41And eventually, I think that even the idea of selecting the model, that thing may even disappear. So ideally, the ideal interface is you have a conversation and you explain what you want, And maybe that conversation includes not only text, voice, but also images, sketches. You dump your ideas and your past, your history, and it kind of understands what you want to do and helps you in your task. What model is using at every particular time is not really that important. And you see that of relating the assistant, for example, will tell you what is the underlying element under the hood. And people don't care.

23:21They just want the assistant to understand what they are saying. okay so to me free big is the product is the model is not it's not the product if that makes sense and you're saying too that you kind of filter out what models you have on the platform just not to have every single model but just to have the best in class at doing a specific something and have those models on the platform the thing is that today you still find many models that are good at some particular thing. And that makes the life of the user very difficult. And that's why I think that today Freebeek is more for professionals than for individuals that are just, you know, getting in because they want to make something funny.

24:04So my intuition is that just like the generation of images and the generation of videos is going to be all over the place. Like, yeah, you will have it on your iPhone. You will have it. You will have it everywhere eventually. We are building a product for professionals that want to do something. They have an idea and they want something that really is like their idea. In general, when you use a more amateurish tool, they generate something beautiful, but it's not always what you wanted. So if you're in this inspiration where anything kind of works, that's great. That's a stage of creation, and I respect that.

24:47We are trying to serve users that already know what they want, and we are trying to help them achieve their vision. And that's a different part of the creation stage. In that vein of thought, you clearly have a product or a suite of products that's really geared for professional use. Amongst that giant group of professionals, what sort of demographic is the super user? Where are you seeing strengths amongst that entire marketplace? First of all, let me start by saying that we are biased, okay? Because before EA started, we had plenty of people coming to Freebie that were the two big groups. It's marketing people and designers.

25:30And very often designers working in marketing departments. So that's our core. And of course, many people are still from that core use case. Okay, so marketing for us is number one, okay? Marketing use cases. There are a couple of new use cases, big ones. Like the first one is visual effects. People working in Hollywood, people making films, working visual effects. That is huge. And today it's exploding. It is growing very, very quickly. And the third one is people that want to do product placement. So that's typically companies that have a physical product. they used to fly people all over the uh all over the place to take gorgeous uh photo uh photo uh i don't know photo sessions that used to cost between 20 and 50 even 100k okay depending on where do you have to go and the complexity of the object and now more and more of them are transitioning to ai photography kind of okay so those are like the three the three big groups that we see there are some other uh professions that are starting to use ai for example photographers for back in the day they were not uh ai users and they were not for big users certainly and now we we have more and more photographers using the platform because of our app scaler which that's the only uh acquisition that we made like it's magnificent some acquisition that we made last year and we have many many photographers using it then you have like more like smaller pockets of different use cases you have architects and interior design people that are using restyle they take a sketch you know and they get an image or they take a photo and they change the furniture and the style so it's there we we track like 10 different professions all of them are growing Initially, we were only on a single one, design and marketing.

27:36And now it's just growing over the place. But if I come to big three, it will be design slash marketing, visual effects, and product investment. Thank you. Do you see expanding sort of the tool set as well? And I know we sort of DMed about this and I get questions from the AI community and even my team asking more sort of the project management side of AI generation of like being able to track like better position shop. sort of track versions of shots, storyboard, sort of the kind of nitty gritty project management part of, and more so in the AI filmmaking space. Absolutely. That's actually where I see like the next down value that we have to deliver.

28:16And honestly, it's a pity that we still do such an awful job at that. And the reason is because we were not that in the past. Okay, so that's initially we integrated AI. And as I mentioned, I see FreeBica as the product, so we better deliver value at that layer. So we absolutely want to show versions, help people to collaborate, to comment. I mean, there are more things that we are doing, but a big one is adding that layer of backups, security, integration on enterprise, storage, versions, etc. And I got one other question from one of the AI communities asking about some of the details with like the enterprise plan and that has some like legal coverage or for like commercial use.

29:02Like, how are you thinking about that and what kind of protection you're having with people using, you know, AI and commercial space? Yeah, it's a great question. So first of all, let me say that in our opinion, everything that is in the platform is kosher. Everything that is in the platform is legal. It can be used. there's not going to be any issues. For enterprises, we put together indemnification. So if you sign an enterprise deal, then we give you an indemnification that by default is$1 million. If the deal is bigger, we can go higher. I would say that the legal fear, I mean, you tell me if I'm wrong, but I think the most common legal fear that people have is that if any images were used to train a model, to kind of get sued by the copyright holders of the people that own those images that were used in the model.

29:55And that, in general, I don't think is going to happen. In general, in most cases, there will be probably a few exceptions. There are a few products that didn't do it correctly and are still in the market. But in Europe, actually, what is legal and not, it's now pretty clear. okay in europe you can actually use images that you find on the internet with the exception of images where the copyright holder opted out of the image you know if the copyright holder explicitly say that they don't want that image to be part of a model then you have to take it out but if the copyright holder didn't say anything which is the case actually for the vast majority of the images on the internet, then you can use it.

30:43And the reason why you took that view is because if you ask for explicit permission to use an asset, then you cannot train a model. And I'm not even thinking on images here. Think on text. Like to train an LLM, like Chajic D, or anything like that, you need to use the whole of internet and then some. It's even more than that. And And you cannot get a license for the whole of the Internet. So the only model that kind of rewards authors and allows the existence of modern AI is, okay, you can take anything, but if the copyright holder complains, you have to pay them or you have to take it out of your model.