Custom AI models are not just for the AI giants anymore. Because the 37-person startup Krea is releasing its first generative AI model as the design tools startup repositions itself as a full-fledged AI research lab.
The move is significant for Krea, but it also seems to tease an almost inevitable moment in the rapidly evolving AI market, where smaller players in the industry can make more disruptive bets.
On one hand, Krea can hardly call itself a bootstrapped startup anymore. It’s now raised $83 million through its Series B at a $500 million valuation. On the other, it’s tiny compared to the leading frontier model companies, which constantly raise more money to ensure they have an unlimited war chest to train the next best model: OpenAI and Anthropic, which have raised $180 billion and $72 billion, respectively.
But to Krea’s co-founder, Diego Rodriguez, it’s invigorating to be small, nimble, and, by one significant measure, no less successful than any frontier model company as a core business.
“Until there’s a winner—until OpenAI or someone is profitable—the Olympic Games are on,” he says with a mischievous smile.
The evolution of Krea
Krea launched in 2023 to be something like the Adobe of the AI age, a creative platform designed from scratch to allow you to not just generate media with AI, but to tune those outputs, with controls that feel more like a synthesizer than a drafting table. They were the first to offer real-time AI editing tools and the first to put APIs from other AI models into their own app (a practice that has now become standard). And they were quickly profitable.
But over time, the team has recognized a distinct ceiling to their work: Krea can only be as open-ended as the models it sits upon.
Image models of today are amazing at specific prompts that often go viral, but they can also feel like they are built on rails. Creative phrasings can still lead you down the same old paths, as models fail to reproduce what’s in your mind’s eye.
“The models are trained not to fail and to always give you a good image,” says Krea’s co-founder, Victor Perez. “And I feel like that takes away a lot of the creative uses—breaking the barriers and letting people go off-road, letting [you] make ‘bad’ images, stuff that looks more artistic that a creative might appreciate more.”
Indeed, image models are amazing when it comes to what these companies have been prioritizing: photorealism. But any designer reading this knows that when it comes to graphic design and illustration, you can hit the boundaries faster than you’d think.
In a demo, Krea pulls up comparisons of the prompt “a cat riding a bicycle” between itself and Google’s Nano Banana. In Krea’s case, the first outputs are funky and varied, with some exhibiting a hand-drawn feel. In Google’s, no matter how you adjust the prompt, you get a similar coloring-book-looking image presented in the same way. It’s the difference between eating at McDonald’s or a Michelin burger joint. One will always aim to please, while the other may polarize.
“I think that the kind of stuff that we are interested in is more niche,” says Perez. “It’s a much smaller market, but we’re fine with it.”
Spending 15 minutes prompting Krea’s new image model K2 on my own, and I’m impressed by its breadth. It generates surreal photorealistic scenes, but also grainy VHS-style filtered images and a variety of illustrative techniques (word marks, manga, anime, hand sketching, and sharpie cartoons) well. The examples I saw from Krea were also impressive—and wildly so given the gulf in resources between Krea and the giants. Perez attributes this success to his team’s own taste. They’ve spent the last seven months building their own data set (no, they aren’t disclosing the sources), labeling it by hand, and creating their own unique workflows to train their own generative AI system.
As Perez explains, most big models start the same in development to build a functional neural net, but mid- and post-training steps in particular are what give the model a point of view. I’ve heard from people in the industry that there are only about 200 true post-training experts in the world, which is why the market is so competitive.
“That’s when the artistic direction on the model takes place,” says Perez. “At the end of the day, building a model is almost like crafting a sculpture.”
That final layer of training, where a model develops its visual or verbal voice, is where taste comes in. Making the AI do one thing better can often make it do another thing worse, and balancing those priorities is particularly tricky when trying to build a model that makes cool, personally expressive stuff.
“This is like the nemesis of an AI researcher, because what researchers are really good at optimizing for [is] metrics,” says Perez. “But what is this metric that we are optimizing for? Like, it’s something so subjective.”
The user interface
K2, Krea’s new model, seems impressive on its own. But what makes it so attractive is how Krea will let you use it.
On the baseline, Krea promises that just describing what you want will get you better results with K2 than its competitors. Then Krea’s user interface lets you really get your hands dirty in tuning the output. You can drag one or multiple images you want into the prompt bar, to use that to influence the style it generates. Then you can drag a slider up or down on those images, to signal how much you want them to influence the visual style. You can even build a mood board to inform the aesthetic that you’re after. (After generating some images, Krea will proactively produce a sort of personalized Pinterest board with more images it thinks you’ll like.)
Because this system is built for creatives, Krea is also being careful with how it frames up IP. As you ostensibly train your own model inside Krea, you can remove that from Krea’s own model training. And all IP generated is your own. So if you are an oil painter who has a very particular style that you want to use within gen AI media, you can upload your work to reproduce it without worrying that Krea is about to sell that as a filter to someone else.
Longer term, Krea is considering if there are ways to credit artists whose IP measurably influences a piece of media, and they’re experimenting with using AI to do just that to create a more sustainable royalties system.
Rodriguez admits some confusion as to why, in an industry dominated by OpenAI, Anthropic, and Google, smaller AI companies aren’t banding together in order to build bigger ideas and share the wealth. Originally, Krea tried partnering with a model company that refused to offer even a small split of revenue, which led them to develop the technology completely in-house.
But now, I can’t help but notice how much Krea’s ambitions have grown. Perez declares that this launch product, K2, is “conservative.” The GPU cluster Krea is using for a year, over which time it will have trained K2 and two future Krea models, will cost the company $20 million. Krea couldn’t afford to faceplant with an experimental approach that might not work. However, with a success under their belts, they feel more confident to take more risks and challenge training norms.
“We just wanted to make it work,” says Perez. “It worked way better than we expected, but this was an extremely risky bet. We’d never trained a model before. We didn’t know how hard it would be. And it was it was fucking hard, but at the end of the day we figured it out. And now we know so many things—because there’s so many things about training a model that you can only learn through training a model.”
