How Sivi is using Gen AI to turn text into visual designs

Creating compelling visuals isn’t everyone’s forte, yet the aesthetic identity of your business can make or break a potential customer’s decision to engage. For example, a poorly designed advertisement may fail to convey its message effectively and lead to missed sales opportunities.

This principle extends beyond advertisements, and holds true for all graphical communications, whether it is a business card or marketing collateral. 

Streamlining design processes automatically is AI platform Sivi. Started by Sona J and Ram Ganesan in 2019, the Bengaluru-based startup lets users input their message to generate multiple AI-designed options. These designs can be used as-is or fine-tuned manually to enhance the final product.

Sivi's product lets users instantly generate ads, social posts, product banners, profile covers, and more in multiple languages.

Building the product

A computer science engineer, Ganesan started his AI journey at Samsung where he worked on NLP and machine learning projects, and filed three patents in the US during his tenure. From there, he went on to build MineWhat Inc, a sales and intelligence and automation platform.

At MineWhat, Ganesan Sona, a domain expert, who later joined him to start Sivi. A BTech Gold medalist with more than 14 years of experience in design, Sona has worked at Bridgei2i as a chief designer.

Sona and Ganesan started Sivi after realising the wide gap in AI models for making customisable designs. AI is a relatively new tool and getting the desired design may not be possible in the first try. On top of that, these tools cannot reliably output text, a major component of business communication.

“Abundant data exists online for text and images. There is a lack of structured data for designs as it is multifaceted and multi-layered in nature. At Sivi, we built custom foundational models for layered design generation and prepared unique structured design datasets over a period of three years (beyond image and text datasets) to solve customisable design generation,” says Ganesan.

The founders created a model trained on objects such as text, shapes, and images. They collected and created over 30 million layered and structured design data points to train an in-house model.

After fine-tuning the model to achieve the desired results, the next step was to build a product around it. 

“We built Sivi on two key pillars—the Atomic Design methodology and a Human-in-the-Loop approach. Users provide the building blocks—assets, text, brand, and style preferences, and Sivi analyses and understands these elements, gradually composing the designs from scratch, just like a designer would,” says Sona.

Unlike other popular image generation AI models such as Midjourney or Dall-E, Sivi’s model is trained on design compositions, allowing users to select any object, such as a shape or text, within a design and manipulate it as desired. 

However, the company faced challenges in achieving this and had to create an AI model from scratch as there was limited structured data available for such projects. 

Sivi also needed to hire skilled AI engineers and train models on expensive cloud infrastructure, all while operating within a limited budget.

Standing out from competition

With its product, Sivi aims to compete with the likes of Sydney-based Canva. Microsoft has also developed a generative AI model that outputs layered designs, but it is currently hard to access and available only to researchers.

While both Sivi and Canva operate through web portals, they have different approaches to design. Canva relies on a vast repository of objects, fonts, and images, along with powerful editing tools.

Sivi, on the other hand, creates objects on the go, and ensures no two layouts are the same. Users can easily change any aspect of the layout they dislike.

Early this year, the startup, with its 12-member team, launched Sivi Gen 2 to make designs much catchier. 

“Today, our design quality matches up to 65% of professionally designed templates. Our plan is to develop Sivi Gen 3, a more powerful model that can achieve up to 90% of professional design standards,” adds Sona.

Customers and market

Sivi’s offerings can be useful for generating designs at scale. The same message can have different iterations while having the brand image consistently maintained throughout the design. With just a click, the templates can be translated into any of the 72+ languages Sivi supports, including 17 Indian languages.

“We have users from 140+ countries who have generated over a quarter of a million designs using Sivi. Most of our users are from the US and Europe,” says Sona.

While the startup is targeting various kinds of users, including ecommerce firms, agencies, freelancers, SaaS, education, and Tours and Travels, Sona says that right now the tool makes most sense for digital advertising agencies who need to quickly reiterate a design.

“We have over 24,000 users, and some of the ecommerce and agencies as early customers,” says Ganesan. 

“We are operating in a large graphic design market worth $45 billion. Nearly 200 million business users need visual content. With Sivi Gen-2, we are targeting 60 million users,” he adds.

Right now, Sivi works on a token-based pricing model where users need to buy credit in order to generate fresh designs. While it is currently offering six designs worth of tokens for free, its plans can go up to $160 per month with unlimited tokens and 16 GB worth of free online storage for the designs.

So far, Sivi has raised $650K from institutional and angel investors. It is now looking to raise more funding to improve its tech stack, build larger models and marketing the product to a wider user base.

The startup was part of TechSparks 2023, YourStory’s flagship tech startup event held last year in Bengaluru, and made it to YourStory’s Tech30 cohort as one of the 30 most promising Indian startups of 2023.