Exactly.ai — A Platform Where Artists Can Train AI on Their Artworks

The new hero of 'Hi, AI! Media' is serial entrepreneur Tonya Samsonova:

• Education: School of Sociology Graduate, Higher School of Economics
• Career: Former journalist - host at 'Echo of Moscow' radio station and 'Dozhd' TV channel
• Entrepreneurial Experience: Founder of the question and answer service TheQuestion, which was bought by 'Yandex' for 10 million dollars
• Founder of the startup Exactly.ai

About a year ago, in ancient times by AI standards, developers trained their models on any data they could find in the public domain.

This method gave rise to many of today's leading neural networks.

However, this approach has a major flaw: eventually, content creators such as artists, writers, and designers began to restrict access to their datasets, realizing the potential of AI to make their roles obsolete. This led to a slowdown in AI development, often entangled in legal disputes.

Tonya Samsonova, the founder of Exactly, addresses this issue. Her previous venture, a Q&A service called TheQuestion, was acquired by Yandex for $10 million.

Exactly.ai offers a unique platform for creating bespoke AI models. Here, artists, designers, and illustrators can develop their own AI models using just 10 of their images in a consistent style. They can then monetize these models by granting access to well-known brands. This enables brands to legally generate and utilize unique images in the artist's style.
How did the idea for Exactly.ai come about?

In August 2022, I began thinking about creating a platform where people could publish their own AI models. Generative AI models had existed for a long time. That is, seriously considering them as a product or an application intended for people was not feasible. At some point, there was a technological breakthrough in image generation. It immediately became clear that if there are only one, two, or three AI models in the market, they would create images in a certain aesthetic. In reality, the world is much more diverse, and uniqueness and an authorial approach are valued. If we develop technology trained only on one segment of data, we miss out on the values that exist in everything else.

I decided to create a platform where every person with a unique style and data could create their own AI model.

There are 200 million visual content creators in the market. There's a high demand from brands for their work. Annually, companies invest $11 trillion in image production. Many of these skilled creators are skeptical of 'all these computers and new abilities'. I saw an opportunity in this.

I decided that if Midjourney is one algorithm, I would have 500,000 different algorithms. Data is a key ingredient in artificial intelligence. All companies developing AI focus on their own developments. But in open-source, everything happens faster. Scientific achievements develop faster than in any single company. Therefore, it's not enough to just have access to technologies and engineering talents; access to unique data is necessary. Consequently, platforms need to be created that allow these data to be applied. This is how the idea came about, which later took a comprehensive form: a business form, an investment form, and proprietary technology.

To create personalized AI models, do you fine-tune Stable Diffusion with the data uploaded by creators, artists, and platform users?

Not exactly. It was a great joy when Stable Diffusion appeared, because before that, the idea was to train our own model. We work on the Stable Diffusion architecture, into which our development is integrated, containing three key components.

The first component is a model that is responsible for the quality of image generation. Professionals have very high requirements: for video transmission, light composition, shadows, and other aspects. Each artist has their own requirements, so we have high sensitivity to textual description. If an artist communicates with the model in a professional language, we must be able to understand it well.

The second component is a technological solution that allows for a precise understanding of an artist's style, even from a small number of their works. In this, we are currently the best in the market. This is especially true for styles related to abstract painting or abstract patterns. You take 10 images of anything. They must be in one style. Send them to us. And our task is to replicate this style. This creates an effect of magic. This was one of the huge drivers of growth for Exactly, because artists look at this and think, how is this possible?!"

The third part of our technology is the infrastructure and user interface that provide integration of all the necessary tools for working with images in one place. For example, tools for editing, finishing, and so on. We aim to provide the best tools that combine knowledge about each user's personal model.

Can an artist publish their unique style on your platform? How do you verify that the artworks are indeed theirs?

Our goal in creating a marketplace is to give artists the opportunity to publish their artificial intelligence models. We also aimed to build a platform that enables businesses to acquire access to these models, the ability to generate images, or to purchase licenses for using the entire model.

We have a multi-tiered model of verification. When you upload your images, you claim they are yours. If it turns out to be untrue, we side with the rights holder. In public models, we publish the dataset. We also conduct both manual and automatic checks for authorship. So, there is a legal layer of verification, a public layer, and a technological layer. This is a challenge we solve every day: matching the name of the rights holder with the data.

Is personalization of models a trend in AI development?

Yes, we believe that AI development will move towards the personalization of models. Everyone will be hunting for quality data: not just from artists and writers, but also from highly skilled lawyers, McKinsey employees, doctors, and all other professionals. Less data will be needed to fine-tune the model, but it must be of high quality. Data will be golden, and it will be important for every person, every professional, to know whether models are trained on their data or not.

Today, it's like the 'Wild West'. If I want to fine-tune my model with anyone's data, I copy it from the internet, and the person doesn't find out. Legally, this will change. For example, we see that after a series of strikes, American screenwriters signed an agreement with studios allowing them to check what models used for writing scripts are trained on. We see that European legal system is moving in this direction too. We plan to be leaders in helping professionals protect their data and use it ethically and legally.

You mentioned lawyers and McKinsey specialists, so are you referring not only to image creation models, but also, for example, to an AI lawyer trained on the documents and work of a specific well-known lawyer?

These are verticals that could emerge. However, for the next few years, our focus is the market for business illustrations. Take a look at Uber app, Yandex home page, etc, and you'll find many illustrations in any product. There's a huge demand; trillions of dollars are spent on images worldwide every year. I have always been on the client's side. I was the person who needed these illustrations to grow my business, launch a marketing campaign, and so on. Now, businesses can upload their illustration guidebooks to Exactly, not worry about the security of their data, and generate pictures for the company across marketing, SMM, and other departments. Meanwhile, artists can display their styles on the marketplace for businesses to purchase.

So, the client is the brand. For example, I can generate images in Midjourney, but Coca-Cola cannot do that?

Coca-Cola cannot do that for two reasons. First, because they would face a lawsuit. The second reason is that now, with the advent of generative networks, the competition for attention has only increased. Now every blogger, every company can create an image, and accordingly, for top brands that are building their communication, it's even harder to break through this noise in the feeds that people read.

The demand for uniqueness and artistic talent will only grow. It seems that now a vase can be drawn, say, in DALL-E 3, but in reality, it cannot. In reality, if the world was such that you could put any picture on an advertisement, and it would increase your sales a thousand times, everything would be great.

However, we are aware that for an image to truly resonate with a brand and drive revenue, it necessitates the collaborative efforts of a diverse team, including strategists, researchers, brand managers, marketing professionals, art directors, and other key players. Therefore, the skill to draw itself is not valuable. The valuable skill is to turn a marketing brief into a creative brief, that is, the talent and individuality of the artist who can be different and relevant to the brand is the most valuable.

Can you give an example from real life?

I needed to order illustrations from Jean Jullien for TheQuestion. I wrote a letter to his agency. They replied that he would be available in 8 weeks, and one picture would cost 5000 euros. I thought: okay, 5000 euros, I understand he's great, and I need this, but I can't wait 8 weeks for him to draw one picture. And I don't need just one picture, but 4000 a day, because I have a constant experiment. This road led to nowhere. But now, thanks to AI, what can we do? I can buy the style of Jean Jullien and generate 4000 images a day. He has 2 million followers on Instagram. He can work individually with someone, but for everyone else, he can publish models on Exactly. He gets additional income. As a brand, I get the opportunity to stand out among competitors and work with the coolest people in the market.

We already have many famous artists and illustrators who work with British, Asian, American superbrands. The quality of the content is very important, so we are working on attracting the best of the best in the art market and among illustrators.

Interviewed especially for Hi, AI! media by Anatoly Buzinsky

@gpt4telegrambot