Kandinsky Neural Network 3.0 users can also create text prompt-based videos in the animation mode. A single prompt generates a four-second 24FPS video with a selected animation effect and a 640×640 resolution.
Sber has released Kandinsky 3.0, the next version of its creative generative model, and presented it during AI Journey, Sber’s artificial intelligence and machine learning conference held last week. Compared to previous versions, the new one is better at recognizing users’ text prompts and can generate even more photorealistic images, full-fledged art paintings and sketch art. The model works with prompts from a wide list of topics and is able to create any fantasy.
Kandinsky 3.0 knows the DNA of the Russian culture better than its predecessor. For example, the quality of generation of Russian and Soviet celebrities and characters, architectural landmarks, cultural objects and elements of Russian folk art, such as Gzhel painting, has been improved significantly. Furthermore, the new model has enhanced inpainting and outpainting modes.
Capable of high-resolution 1024 x 1024 image generation, Kandinsky 3.0 also allows users to choose the aspect ratio. To train the neural network, the developers used an updated dataset of 1.5 billion text-to-image pairs containing data that underwent multi-stage filtering procedures, which resulted in a noticeable improvement in the quality of generation.
Alexander Vedyakhin, First Deputy Chairman of the Executive Board, Sberbank said, “Artificial intelligence can give humans superpowers, and Kandinsky is one of the tools to do that. It is a user-friendly, functional and free creative neural network by Sber. We are improving it relentlessly. The new version understands prompts even better, and has learned to understand the subtleties of Russian culture and folk art. Try and create unique artistic works that can become a gift or be a great addition to your collection.”
It takes an average of approximately 20 seconds to synthesize one second of a video. To extend the capabilities of the basic model, Sber used different types of image animation, which allowed to move objects, zoom in and out, and animate statics in all possible ways. The animation modes have redrawing by text description (image2image) at the core.
Kandinsky 3.0 ingests prompts in 100+ languages, delivering an unlimited number of styles. The neural network was developed and trained by Sber AI researchers with the partner support of AIRI Artificial Intelligence Research Institute on the combined dataset from Sber AI and SberDevices.
