Sber Lets Users Create Digital Avatars for Free
Developers now have access to Kandinsky Avatar 1.0, a family of Russian generative AI models that can create realistic talking avatars from a photograph and an audio recording.
Kandinsky Avatar 1.0 is the first fully Russian-developed model designed to create realistic digital characters. All users need to do is upload an image of a person or fictional character, add a voice recording and provide a brief text description of the scene. The AI then synchronizes the character’s lip movements with the audio and generates facial expressions, head turns and gestures. Users can even specify the character’s emotions: it can appear sad, joyful, happy or angry, depending on the prompt. The model generates videos of up to one minute at 24 frames per second and supports speech in multiple languages.
According to Sber, developers can access the model code and weights for free under the MIT License through GitHub, GitVerse and Hugging Face. The model will also be available through popular tools for running and integrating neural networks, including ComfyUI and Diffusers.
What Kandinsky Avatar 1.0 Can Do
The service could be useful to bloggers and content creators. Given a script and an audio recording, they can produce a finished video featuring a talking presenter. The format could also work well for educational platforms, where a virtual instructor may cost less than a human one. Kandinsky Avatar 1.0 can also be used to animate talking characters and build apps for personalized video greetings and custom avatars.
“Video avatars make it possible to create educational and promotional videos, how-to guides and personalized messages for customers. We are opening up the model so Russian developers have their own technological foundation for building these kinds of solutions. They will be able to integrate it into their products and fine-tune it for specific tasks instead of building the technology from scratch. Feedback from the developer community will help us understand how the model performs in different use cases and what needs to be improved. We will apply what we learn to future versions of Kandinsky and as we bring video avatars into Sber’s services,” says Denis Dimitrov, Managing Director of Data Research at Sberbank.
Kandinsky Avatar comes in two versions. Kandinsky Avatar 1.0 Pro is the flagship model, generating HD video with greater image detail. Kandinsky Avatar 1.0 Lite is a streamlined version designed to run on less powerful hardware, including consumer graphics cards. It produces videos in SD resolution.
The Model Has Held Its Own Against the Competition
The market is already highly competitive, with international models among the contenders. HeyGen, for example, also offers digital avatar creation and multilingual video localization. But testing has shown that Sber’s model has some advantages.
The developers say they compared Kandinsky Avatar 1.0 Pro with other open models offering similar capabilities. Participants watched pairs of videos, rated them against seven criteria and chose the better one without knowing which model had generated each video. Kandinsky Avatar 1.0 Pro was selected more often than each of its competitors: in 63 percent of comparisons with Kling Avatar 2, 60 percent against LongCat-Avatar-1.5 and 54 percent against HeyGen IV. The model’s strongest advantage was in synchronizing lip movements with speech.
A Boost for Russian Developers, With Export Potential
The release of Kandinsky Avatar 1.0 marks a shift in Russian generative AI, from producing individual images to developing specialized tools for creating digital characters and video content. Just as important, the Russian-developed model gives other companies a technological foundation on which to build their own products. Potential users include developers of educational platforms, advertising services, video editors, remote customer service systems and enterprise software.
Open-source code lowers the barrier to entry in the generative video market for Russian IT companies and small studios. Rather than developing a model from scratch, they can adapt an existing technology to meet their needs.
The technology could appeal to a broad range of users, from professionals to hobbyists, including small video production studios, advertising agencies and independent digital content creators. Its ability to work with multiple languages also increases the model’s export potential and lays the groundwork for its further development.
