Yandex presented a prototype of its own neural network for generating images based on user text queries. Beta version 0.1.4 of the Shedevrum app is available on Android and iOS.
Shedevrum application understands Russian and English, and can also take into account special wishes (for example, “photorealism” or “high detail”), imitate famous painters, and work in given art styles.
Shedevrum neural network forms images using the cascaded diffusion method: first it creates images in accordance with the request, and then gradually increases their resolution, saturating them with details.
The first version of the generative model for the beta test was trained on 240 million examples of pictures with text descriptions. Training continues on a dataset of 500 million examples, and the quality will improve incrementally in future updates.
Alexey Gusakov, Head of Machine Intelligence and Research at Yandex said:
”We are still at the very beginning of the journey. At the heart of the Masterpiece is a prototype of a neural network, which has a lot to learn. By the time of the final release, it will be able to generate banners, illustrations, create images for online storefronts and more. The beta test of the application for users is a new experience and a little good mood, for us it is an opportunity to understand a little better in which direction to develop the neural network further”
Finished images could be seen by downloading Shedevrum. To generate images according to your own requests, you need to leave a request and wait for its confirmation. In the future, all registered users will have this opportunity.
