A new arXiv preprint introduces Kandinsky S2V 1.0, an open family of speech-to-video models aimed at controllable audio-driven human animation. The family includes two variants: Kandinsky S2V 1.0 Lite with 2 billion parameters and Kandinsky S2V 1.0 Pro with 21 billion parameters.
The models are designed to take a single input—likely a speech audio track—and generate animated video of a human figure, with control over the animation driven by the audio. The exact architecture and training details are not covered in the abstract, but the parameter counts suggest a substantial range of computational requirements.
Because the source is a single abstract, there are no conflicting claims to compare. The main takeaway is the release of an open, two-tier model family for speech-driven animation, which may lower barriers for researchers and developers working on avatar generation.