Kandinsky Avatar: Sber releases neural network models for free

Sber has released Kandinsky Avatar 1.0, a family of neural network models designed to create speaking video avatars based on a single photograph and an audio recording. The source code and model weights have been published under the MIT license on GitHub, GitVerse, and Hugging Face. Developers are free to utilize these models without charge, including for commercial projects.

• Kandinsky Avatar 1.0 generates videos up to one minute at 24 fps.
• Trained on over 1 million video clips across 95 different languages.
• Available in Pro and lightweight Lite versions for standard GPUs.
• Emotions and gestures can be controlled via text prompts.

Technical features and model variations

The model is capable of generating clips lasting up to a minute with a frame rate of 24 frames per second. To create a video, a user simply needs to upload an image, an audio file containing speech, and a brief text description of the scene. The neural network automatically synchronizes lip movements with the speech, adding facial expressions, head tilts, and gestures. Notably, the model does not synthesize or clone the voice itself, but strictly relies on the provided audio track.

Kandinsky Avatar 1.0 comes in two distinct variants. The Pro version generates high-definition HD videos, while the streamlined Lite version is optimized to run on standard home graphics cards, producing SD resolution clips. Training was conducted using an extensive dataset comprising over one million video fragments spanning 95 languages, ensuring broad linguistic adaptability.

Evaluation results and competitor comparison

During blind testing, the Pro version was benchmarked against competing tools such as Kling Avatar 2, LongCat-Avatar-1.5, and HeyGen IV. Kandinsky's videos were chosen over competitors in 63%, 60%, and 54% of comparisons respectively. The model demonstrated its greatest advantage in precise articulation-to-speech synchronization and overall naturalness of movement.

Furthermore, text prompts can be used to assign specific emotions like joy or sadness to the character, increasing creative flexibility. Releasing the project openly significantly boosts its adoption among global developers.

Practical applications and industry impact

Developers can leverage this technology to build educational platforms, virtual assistants, video courses, content localization tools, games, and services that generate personalized video messages. Sber eventually plans to integrate these video avatars directly into its own proprietary products. This open-source release provides a powerful instrument for AI-driven multimedia development.