Startup wants to train AI to generate art on licensed pictures

AI, especially text-to-image AI, is attracting as many lawsuits as venture dollars.

Two companies developing popular artificial intelligence art tools, Midjourney and Stability AI, are embroiled in a lawsuit alleging they violated the rights of millions of artists by training their tools on images from the internet. Additionally, stock image provider Getty Images is suing Stability AI for using images from its site without permission to train Stable Diffusion, an art-generating artificial intelligence.

The disadvantages of generative AI are the tendency to repeat the data it was trained on, and, accordingly, the composition of the training data continues to put it under the law. However, new startup Bria claims to minimize risk by training AI that generates images, and soon videos, in an “ethical” way.

“Our goal is to empower developers and creators while ensuring the legal and ethical security of our platform,” Yair Adato, co-founder of Bria, told TechCrunch via email. “We have combined the best of visual generative AI technology and responsible AI practices to create a sustainable model that prioritizes these considerations.”

Adato co-founded Bria in 2020 when the pandemic hit. Another co-founder of the company, Assa Eldar, joined the company in 2022. While working on his doctoral thesis at Ben-Gurion University of the Negev, Adato said he developed a passion for computer vision and the potential to “enhance” communication through generative artificial intelligence.

“I realized there was a real business opportunity for this,” says Adato. “The visualization process is complex, manual and often requires specialized skills. Bria was created to solve this problem by providing an enterprise-friendly visual generative AI platform that digitizes and automates the entire process.”

Thanks to recent advances in AI, both in the commercial and research fields (open source models, reduced cost of computing, etc.), there is no shortage of platforms offering AI tools for text and image creation (Midjourney, DeviantArt etc.). But Adato argues that Bria is different in that it is (1) purely enterprise-focused and (2) built with ethical considerations in mind from the start.

The Bria platform enables businesses to create visuals for social media posts, ads, and e-commerce listings using image-generating AI. Through a web app (API under development) and Nvidia’s Picasso cloud AI service, customers can create, modify or upload visuals, and enable a “brand guardian” feature that attempts to ensure visuals meet brand guidelines.

The AI ​​in question is trained on “authorized” datasets containing content that Bria licenses from partners, including individual photographers and artists, as well as media companies and stock image repositories, who receive a cut of the startup’s income.

Bria is not the only enterprise exploring a revenue-sharing business model for generative AI. Shutterstock’s recently launched Contributors Fund compensates contributors whose work is used to train AI art models, and OpenAI has licensed part of the Shutterstock library to train DALL-E 2, its image generation tool. Adobe, meanwhile, says it is developing a compensation model for members of Adobe Stock, its stock content library, that will allow them to “monetize their talent” and profit from any revenue generated by its Firefly generative AI technology.

But Bria’s approach is broader, Adato says. The company’s revenue-sharing model rewards data owners based on their contribution, allowing artists to set prices for each AI training session.

Adato explains: “Each time an image is created using the Bria generative platform, we track the visual images in the training set that contributed the most to [созданное искусство], and with the help of our technology we distribute income among the creators. This approach allows us to have multiple licensed sources in the training set, including artists, and avoid any copyright infringement issues.”

Bria also clearly watermarks all generated images on its platform and provides free access – at least it claims so – to non-profits and academics who are “working to democratize creativity, prevent deep fakes, or encourage diversity.”

In the coming months, Bria plans to take it one step further by offering an open source generative AI art model with a built-in attribution mechanism. Such attempts have already been made, for example, Have I Been Trained? and Stable Attribution, sites that go to great lengths to determine which artwork contributed to a particular AI-generated visual. But the Bria model will allow other platforms to enter into similar revenue-sharing agreements with creators, Adato says.

It’s hard to overstate Bria’s technology given the nascent generative AI industry. It’s unclear how, for example, Bria “tracks” the visuals in the training sets and uses that data to distribute revenue. How will Bria handle complaints from creators who claim they are underpaid? Will bugs in the system cause some authors to get paid more? Time will show.

Adato exudes the confidence you would expect from a founder. Despite the obscurity, he claims that the Bria platform ensures that every contributor to AI training databases gets their fair share based on “real impact”.

“We believe that the most effective way to solve [проблем, связанных с генеративным ИИ] – at the training set level, by using a high quality, enterprise, balanced and secure training set,” said Adato. “When it comes to adopting generative AI, companies need to consider the ethical and legal implications to ensure that the technology is used responsibly and safely. However, by working with Bria, companies can be confident that these issues will be resolved.”

This is an open question. And he’s not the only one.

What if the creator wants to abandon the Bria platform? Will it be possible to do this? Adato assured that the authors will be able to do this. But Bria uses its own opt-out mechanism, rather than an accepted standard like DeviantArt or the artist advocacy group Spawning, which offers a website where artists can remove their work from one of the most popular generative art training datasets.

This adds to the burden on content creators, who will now potentially have to worry about taking action to have their art removed from the next generative AI platform (unless, of course, they use a “cloaking” tool like Glaze that makes their art unlearnable). Adato doesn’t think so.

“We’ve made it a priority to focus on secure, high-quality enterprise data collection when building our training sets to avoid biased or toxic data and copyright infringement,” he said. “Overall, our commitment to the ethical and responsible training of AI models sets us apart from our competitors.”

These competitors include incumbents such as OpenAI, Midjourney and Stability AI, as well as Jasper, whose generative art tool Jasper Art is also aimed at enterprise clients. The hefty competition – and open ethical questions – do not seem to have deterred investors: Bria has raised $10 million in venture capital to date from Entrée Capital, IN Venture, Getty Images and a group of Israeli angel investors.

Adato said that Bria currently serves a “wide range” of clients, including marketing agencies, visual asset repositories, and technology and marketing companies. “We intend to continue to expand our customer base and provide them with innovative solutions for their visual communication needs,” he added.

If Bria succeeds, part of me wonders if it will spawn a new crop of generative AI companies that are more limited in scope than the big players today – and therefore less susceptible to legal issues. As funding for generative AI begins to cool, due in part to high levels of competition and questions of liability, narrower generative AI startups may get a chance to cut through the noise – and avoid lawsuits in the process.

We just have to wait and see.