CV Deep Dive

Today, we’re talking with Joaquín Cuenca-Abela, CEO and Founder of Freepik.

Freepik is an image-generation platform that enables users to generate stock images for professional and consumer use-cases. The company was founded by Joaquín in 2011, after his previous startup was acquired by Google and integrated into Google Maps and Google Earth in the 2000s. In early 2023, Freepik fully embraced generative AI - specifically diffusion models such as StabilityXL and others - to become one of the leading image generation platforms in Europe.

Today, Freepik boasts over 100 million monthly users across a range of use-cases, from professional freelancers doing creative work to generalist consumers looking for entertainment. The platform is averaging 5 million image generations per day, putting it far ahead of many of its peers in the image-gen space. For its next wave of growth, Freepik is now focussed on offering the entire creative suite to its users, similar to Adobe or Canva.

In this conversation, Joaquín walks us through the founding premise of Freepik, incorporating AI into its products, and his goals for the next 6-12 months.

Let’s dive in ⚡️

Read time: 8 mins

Our Chat with Joaquín 💬

Joaquín - welcome to Cerebral Valley. First off, just give us a bit of background on yourself and what led you to co-found Freepik?

Hey there! I’m Joaquín, Founder and CEO of Freepik. By all accounts, I'm a geek. I started coding when I was eight years old in 1984. I began with an Amstrad and went from there. My first company was Panoramio, a community of photographers sharing geolocated photos. It was acquired by Google in 2007 and became part of Google Earth and Google Maps. I worked on the Geo team until I quit Google in 2010 and went back to Spain. There, I co-founded a few companies, including Freepik. So, that's who I am—co-founder of Freepik and now CEO, but my heart is still in engineering.

How would you describe Freepik to someone who has never heard of your platform? What exactly does Freepik do?

Let me give you a little bit of background. We started Freepik because we were making websites for different startups and needed to find images very quickly. It was a slow process, having to find free or affordable images. That led us to create a search engine for free images. A few years later, we started creating our own images and building our own library, eventually shutting down the search for third-party images.

Freepik then became a stock image website. We expanded to include icons with Flaticon, presentation templates with Slideshow, and videos. Recently, we've moved from just having content to providing tools on the same platform. This includes online editing and a suite of AI tools to help people create great designs very quickly—the same goal we had when we started the company.

Today, we're more focused on professional creative people. Our number one users are freelancers, and number two are people in marketing who work in design. That's our core audience. We have close to 100 million monthly active users on the platform. So, even though our core is professionals, we have a big reach. We have, for example, some people working in cinema who use our upscaler to improve background images for their films. We also have users who create basic images with our text-to-image tool and then edit them further, either with our tools, Photoshop, or other software, to change illumination or add text.

In general, we are generating close to 5 million images every day. A significant portion of the Freepik user base has shifted from downloading regular assets to creating new ones with AI. There are many use cases like these, showing the versatility and capability of our tools.

You’ve talked about “innovative UI” as a core pillar of Freepik - could you elaborate on exactly what this means?

Our initial technological advantage was having faster models than others, and we leveraged that. For example with our sketch-to-image tool, which allows you to sketch an image from scratch, it draws a fully detailed image in real time based on your sketch. We did the same thing with our text-to-image generator. Besides, thanks to its infinite scroll-down capabilities, it continuously generates images as you scroll down, providing more options and helping people find the perfect image for their needs, even though image generators are not perfect.

We've also utilized our images with Reimagine, a tool that takes any existing image, not only from Freepik’s portfolio, and understands it to generate other images that are similar. This real-time generation of inspired images is unique or almost unique to Freepik.

Walk us through your tech-stack - how are you leveraging generative AI to build Freepik’s image generation engine?

We use models that are derivatives of Stable Diffusion XL. Sometimes we start with fine-tuned models available on CVTI and modify them to make them faster. Every week, we release a new model and A/B test it with our users. Within 24 hours, we can determine if the new model is better or worse than the previous one. This is a huge advantage for us because we can quickly see how many people download the images they create and evaluate the ratio of downloads to gauge the model's performance. We're talking about small improvements or deteriorations — around 2-3% each time. Based on this, we decide whether the model is good to go or not. We do this every week, and I think we are one of the few players with enough volume to iterate this quickly.

We also have a R&D AI team and an AI product team at Freepik, and in the past few months, they have started to work closer together. The idea is to have them operate as a single team. This way, the people in the lab understand the most important needs of the product team, and the product team understands what is possible and what is not feasible with the current technology.

Initially, the R&D team was working on more abstract, exploratory projects. But as the product became more concrete, it became easier to identify the most important areas to focus on. This collaboration has helped align our efforts and ensure that both teams are working towards the same goals.

What has been the biggest challenge in building Freepik thus far?

The good thing about making the model faster is that it also becomes cheaper. When we started, we could generate 30 images for $1, which was too expensive for us. Our goal with Freepik is to make design affordable, so we offer a very affordable subscription that is a fraction of the price of our competitors and it's unlimited. We want to keep it that way and provide as much as our users need. Initially, the cost of AI-generated images was too high, but with optimization, we went from generating 30 images per dollar to 3,000 images per dollar in just a few months.

I believe that soon we will think about the cost of image generation like we think about the cost of electricity—it's a fixed cost that you don't worry about too much. Some people will use more, some less, but it won't be a significant concern. That's why we are committed to an unlimited model as much as possible. To achieve this, we had to move out of the cloud to make it cheaper. Today, we use GPUs on a bare-metal provider, pushing the limits to ensure we can offer this service to all our subscribers at no extra cost.

How are you planning to incorporate recent advances in multimodal technology into Freepik’s core platform, given that it is so relevant to your user experience?

What we see is that these models act as building blocks in workflows. People have more complex needs than just generating an image or video. They might want to specify with precision the style they want or insert a particular object without AI altering it. This is crucial for various fields like marketing and video games, where precise control is needed.

It's not just about generating the image; it's about giving people more control over what they create. The cost of generating an image won't be the main factor in pricing for users, just like the cost of rasterizing a 3D image isn't relevant to the cost of the tool you use. What excites me is the potential value we can provide to creative people who are currently very constrained. We aim to enable them to do much more.

How have you embraced this new phase of growth for Freepik, and what changes did you have to make to the team in order to prepare it?

Specifically in AI, we started from scratch. Generally, we had a culture of hiring new people who might not have a lot of experience but were enthusiastic. Attitude is very important because it's difficult to change. If someone has the right attitude, the rest can be developed. So we worked with talented individuals with the right attitude, even if they lacked extensive experience. This was partly because of our location in the south of Spain, where the startup ecosystem isn't as developed as in places like San Francisco. We had to work within that context.

However, with AI, we took a different approach. We hired people with deep expertise. This was more expensive, but it also allowed us to get results faster. The AI team members all had previous experience in AI, and we knew they were very good and could be productive from day one. While the culture of the AI team was slightly different, the overall culture of the company remains very strong and consistent with what it has always been.

Lastly, how do you feel about the European efforts in AI? Do you think these are net-beneficial or harmful to Europe’s position in the world of tech?

My feeling is that there is plenty of talent in Europe, and many Europeans are very successful in the US. We have a good education system, and people are eager to innovate. However, there's something—whether it's legislation, culture, or something else—that's holding us back from being as successful as we could be.

I would like European policies to focus more on fostering successful companies, creating bigger enterprises, and helping people succeed, rather than just imposing limitations for theoretical benefits. Sometimes, it feels like we are giving people rights they aren't even asking for. The dangers of AI, for instance, seem overblown to me, and it's still too early to legislate effectively on AI. Yet, that's what Europe is focusing on.

Instead, we should take a deeper look at what makes the US, China, and other countries successful and learn from that because we certainly have the right people and the talent to succeed.

Conclusion

To stay up to date on the latest with Freepik, follow them on X and learn more about them at Freepik.

Read our past few Deep Dives below:

• Zilliz Cloud is powering vector search for the enterprise 🔋

• Kolena aims to redefine quality standards in ML 🎛

• Pinecone Serverless is turbo-charging your time-to-prod ⚡️

• Cleanlab is automating data curation at scale 🌐

If you would like us to ‘Deep Dive’ a founder, team or product launch, please reply to this email (newsletter@cerebralvalley.ai) or DM us on Twitter or LinkedIn.

Join Slack | All Events | Jobs