Sber released code and weights for Kandinsky Avatar 1.0

Sber's development team has officially introduced Kandinsky Avatar 1.0, a model designed to generate animated talking characters from a single image and an audio speech recording. The model's source code and weights have been made completely open under the MIT license.

• The model synchronizes lip movements, adds facial expressions, and head turns based on text and voice.
• Available in two versions: Lite with 2 billion parameters and Pro with 21 billion parameters.
• Code and weights are freely accessible via GitHub, GitVerse, and Hugging Face platforms.

Features and Technical Specifications

Kandinsky Avatar 1.0 allows users to specify speaking styles and target emotions using simple text prompts. Avatars can be created based on real humans, cartoon characters, or animals. The model was trained on the Kandinsky 5.0 Video dataset, which includes 1.17 million video clips featuring a single speaking person across 95 languages.

Running these models requires high-end Nvidia GPUs: 48 GB of VRAM is recommended for the Lite version, and 80 GB for the Pro version. However, fragment-based frame decoding can lower peak memory consumption to 28 GB, though this approach may introduce visible boundary lines between segments in the generated video.

Benchmark Results and Competitor Comparison

Kandinsky Avatar 1.0 Pro was compared against leading open and proprietary models in blind tests. Participants evaluated pairs of videos across seven criteria without knowing their origin. Sber's model outperformed competitors, being chosen in 63% of cases against Kling Avatar 2, 60% against LongCat-Avatar-1.5, and 54% against HeyGen IV.

Impact on Developers and Business

Releasing the model opens up technological foundations for creators to build educational videos, promotional clips, and personalized customer messages without building systems from scratch. Developers can integrate the technology into their products and fine-tune it using popular tools like ComfyUI and Diffusers.