Yandex introduces YandexART 2.0 — it generates text on images and helps users and businesses create logos and advertising banners
Yandex has introduced the next generation of its visual neural network — YandexART 2.0. The model can now create text directly on images, apply multiple styles to a single picture, arrange objects more naturally in general and relative to each other, and incorporate more details from the text prompt during generation. Users can leverage these new features to grow their business, for example, by creating a logo for a brand, a product label, website illustrations, or social media posts. At the same time, companies can use it to boost the effectiveness of their advertising and marketing campaigns.
About the technology
YandexART 2.0 is built on the company’s innovative development — a new hybrid model architecture that combines the strengths of convolutional and transformer neural networks. The convolutional model operates similarly to the human eye, identifying key features in an image, such as edges, textures, and shapes. However, this type of neural network struggles with processing longer prompts with many details, which is where the transformer network steps in. Combining these two models in YandexART 2.0 has enabled it to follow text prompts more accurately. It can now apply multiple styles to a single image. For example, it can generate a photorealistic soda can with an anime character on the label.
YandexART was trained on hundreds of millions of images and their text descriptions. To improve these descriptions and make them more accurate, Yandex used its proprietary VLM model. It analyzed images and described their content in detail. YandexART 2.0 was trained on enhanced data and learned to capture more information from user prompts.
While visual neural networks can generate individual letters, they need additional training on large datasets to start forming words. Yandex expanded the training dataset for YandexART 2.0 with hundreds of thousands of images containing text. As a result, the neural network learned to create captions in Latin letters on images.
Yandex developed a new evaluation system to measure the neural network's performance. The new system focuses on four key parameters: relevance, aesthetics, defects, and complexity, which refers to the level of detail and intricacy in the image. For instance, YandexART 2.0 outperformed Midjourney V6.1 in 66% of cases for complexity and 58% for aesthetics and nearly matched it in terms of relevance to user prompts.
YandexART for business
YandexART 2.0 is now available on the Yandex Cloud platform. You can use the API to integrate image generation into your apps or test it in demo mode to fine-tune your prompts. The neural network better understands user prompts, allowing companies to create realistic images for marketing and advertising campaigns faster and with higher quality. Business owners and creators can generate illustrations for articles and social media, design banners, or develop branding options for clothing.
Yandex Cloud clients are already testing the neural network. For example, using YandexART, the Text.ru service created an Illustration Neuro Assistant that helps creators design materials for their website, blog, channel, group, or marketplace store. The presentsimple.ai service also uses Yandex generative neural networks to automatically create presentations for work or study based on text prompts. The service analyzes and organizes content using YandexGPT, while YandexART generates slide images.
In addition, YandexART 2.0 can already be used to create ads in Yandex Direct. Currently, 11% of advertisers use images generated by YandexART within the service. According to experiments, combining neural-generated ads with custom creatives can boost the effectiveness of advertising campaigns by 10–15%.
YandexART for users
With the Pro option, users can leverage the next-gen YandexART to handle everyday and creative tasks via chat with the virtual assistant Alice.
Alice can generate images and customize and enhance them based on the user’s preferences. Prompts can be refined directly in the chat. For example, you can start with, "Alice, draw a handmade candle," after receiving the result, add, "surround it with pine branches." The virtual assistant can help create a profile picture for social media, an app icon, a logo or a T-shirt print, a funny card for a friend, or an illustration for a post.
Users with the Alice Pro option can generate unlimited images in mobile and desktop versions in various formats, with generation taking just a few seconds.
About YandexART
Yandex AI Rendering Technology (YandexART) is a diffusion-based neural network that generates and enhances images and animations and processes uploaded photos in response to text prompts. It was trained on 850 million image-description pairs, understands the Russian cultural context, and uses a unique text recognition algorithm to grasp user preferences better. YandexART can render intricate details, work in a specified artistic style, and generate photorealistic portraits. The neural network is integrated into Shedevrum and Alice and is used in Yandex Business, Direct, Browser, and Market. Additionally, companies can access the YandexART API via Yandex Cloud.