Can Recraft's foundational model for graphic design swerve the AI controversy?


Controversy continues to surround the world of AI-generated imagery, and even as AI-generated images used in elections became a source of concern this week at the World Economic Forum, startups continue to plow the new furrow of AI tools for creators.

The latest is Recraft, an AI graphic design generator aimed at professionals, which has raised a $12 million Series A round led by Khosla Ventures in Silicon Valley, together with former GitHub CEO, Nat Friedman. Also participating were RTP Global, Abstract VC, Basis Set Ventures, Elad Gil, and various other angel investors.

Admittedly there are now myriad Generative AI design tools out there, such as Jasper, Adobe Sensei, Let’s Enhance, and many others. However, Recraft claims to be among the first to be a ‘foundational’ tool, in the sense that it is building its own Foundation Model — a pre-trained, deep learning algorithm — to generate consistent design elements, such as icons and images, that can be tweaked and used within a brand’s particular style controls. It also claims to have amassed over 300,000 since its launch eight months ago.

But this is not designed as a tool for spitting out funny images of cowboy dogs riding horses, or similar. As well as being able to produce these ‘raster images’ it can also generate vector images which are infinitely scalable and used in professional graphic design spheres, unlike platforms where the quality of images can often be limited.

Perhaps this is to be expected, since founder, Anna Veronika Dorogush is not exactly a non-technical founder. She created CatBoost, a high-performance open-source library for gradient boosting on decision trees, for instance. A former head of machine learning systems with the search engine Yandex in Moscow, Dorogush and her 13-strong team are these days based in London.

She told me over a call that the fundraising would power its attempts to build its own foundational model: “You have to provide a lot of control over the outputs to users… over the style so, that you can get consistent images, and control over things like brand colors or level of detail, and also the ability to iterate on the resulting image.”

“But if the model initially cannot generate a pool player or a ballet dancer, then that means that just using fine tuning is not enough to provide high-quality generation, as of now. So we have to build our own model and we are working on that right now,” she added.

She says that Recraft is very much aimed at professionals: “What is unique about what we are doing is that we are not only providing image generation. We are also providing style control: the ability to create your own style and then generate images in your own style… This is important if you want to create a brand and grow it, create marketing materials, creating consistent adverts.”

I asked her what was the likelihood that they might be sued by an artist whose style had been used on the Recraft platform.

“We do state in our terms that if an artist uploads something to the system as a sterile reference — like the image for which they will be able to generate new images — we are not using those to train our model,” says Dorogush.

“We are not using those to train our model.”

She added that Recraft effectively adds generic styles to images uploaded by the artists: “You provide the style to the model. So you have a style reference. And then you generate images using this style of reference. So it is the action of the user to provide the style.”

Dorogush also says that rather than displacing graphic designers, tools like Recraft are more likely to simply make graphic design more accessible in arenas where currently uninspiring stock imagery might have been used: “A year ago, if you were writing a children’s book, you got stock imagery. Now, the same people are able to have much more beautiful illustration and imagery in the same books.”

Of course, Recraft isn’t the only startup playing in this space, and last year I covered how Berlin-based Kittl had raised a €10.8 million ($11.6 million) Series A for its design platform to allow users to turn ideas into graphic products to create professional-grade designs, fast.

This week CEO Nicolas Heymann told me over a call that products like Recraft are part of an increasingly sophisticated set of tools providing alternatives for designers: “I think if Recraft are able to produce design assets successfully — similar to how image generation models can be a good replacement nowadays for stock photo libraries — then they definitely may take [some market] share of those icon libraries.”

Dorogush says that there is space for more than one player right now and that Kittl has a different use case to Recraft: “We do have a unique technology for building style consistent imagery, and that is our main focus. We’re focusing on graphic designers and marketers who need style consistency, while Kittl focuses on merchandise, and I have huge respect for them. They are a great company.”

Whatever the case, it would appear investors are hyped by the possibility of avoiding the controversy some AI and design imaging platforms have attracted, in favor of backing these business-focused AI tools. 

In a statement, Nikita Shamgunov, Partner at Khosla Ventures commented: “We are seeing rapid and significant transformation of the design space driven by generative AI. But to date, many of the generative AI design solutions have been targeting consumers, rather than professionals that require high degrees of control. Recraft delivers on professional workflows such as vector images, style controls and end-to-end content production, all powered by in-house built foundation models.”



Source link