A novel production-grade text-to-image diffusion model aligned with to human preferences using Reinforcement Learning from Human Feedback (RLHF). Our safe-to-use AI system is remarkable for its creativity and strict prompt following. Users consistently prefer YaART’s high expressive capabilities over alternatives *

*
According to a user study on YaBasket and DrawBench benchmarks reported in the paper

AI models learn to generate images from textual descriptions using the examples provided to them during training. Completeness of the concept representation in data and the level of text-to-image relevance are common factors limiting the performance of generative models.

Understanding that, we ensure completeness in our training set by carefully filtering our initial 110B image-text pool, allowing YandexART to learn more than any other generative system. After pre-training on the generic data, we fine-tuned our system to a small set of carefully selected samples with remarkable image quality and image-text relevance. As a result, YandexART shines with precise prompt following and exceptional generation variability, including its ability to generate art and photorealistic images.

Though beauty lies in the eye of the beholder, there exists an inherent splendor that calls to us all, stirring a collective agreement in its presence.

RLHF is a powerful tool known to empower generative language models. As the alignment of LLMs makes their generations more natural and logical, the alignment of our diffusion model makes generated images more aesthetic and consistent. Empowered by RLHF, YandexART can better understand human aesthetic preferences, allowing it to generate visually appealing pictures without tedious re-generation and prompt engineering.

YandexART takes the safety and security of AI very seriously. We strive to create a tool that perfectly suits both work environments and home use. For this purpose, we have implemented advanced image content filtering technology, which filters out potentially harmful content from the image data before adding it to our AI system’s training set. Thanks to content filtering, you can always trust YandexART to deliver safe and appropriate images.

We are committed to providing a high-quality AI system that is powerful, efficient, and safe for everyone. Our image content filtering technology is constantly updated to meet the latest standards, ensuring that YandexART remains reliable and secure.

We acknowledge the effort of those who helped with and provided feedback on this release: Artur Vasilov, Nikolai Gavrilov, Timur Nurutdinov, Vadim Petrov, Andrei Nechaev, Evgenii Bogdanov, Daniil Aksenov, Grigorii Kosarev, Maxim Govorov, Ivan Degtev, Konstantin Lakhman, Alexey Gusakov.

© 2024 Yandex