Sber has released Kandinsky Avatar 1.0, a suite of neural network models designed to generate talking video avatars from a single photo and an audio clip. The source code and model weights are available under the MIT license on GitHub, GitVerse, and Hugging Face, allowing for free commercial use.
The models produce videos up to one minute long at 24 frames per second. Users provide an image, an audio file, and a text prompt to define the scene and emotions. The system synchronizes lip movements with speech and adds gestures, head turns, and facial expressions, while relying on the user-provided audio rather than synthesizing a voice.
The release includes two versions: Pro for HD resolution and Lite for SD, which is optimized for standard consumer graphics cards. The models were trained on over 1 million video clips covering 95 languages.
In blind tests, the Pro version outperformed Kling Avatar 2, LongCat-Avatar-1.5, and HeyGen IV, being preferred in 63%, 60%, and 54% of comparisons respectively. The model showed particular strength in lip-syncing and natural movement.
Developers can apply this technology to educational platforms, virtual assistants, video courses, content localization, and gaming. Sber plans to integrate these video avatars into its own product ecosystem.