Kandinsky S2V 1.0: Audio-Driven Avatar Video Generation

In plain English

An open speech-to-video model family that turns a single reference image plus a speech track and a text prompt into long-form 25 FPS video where lips, expression, pose and gesture follow the audio. It comes in a 2B-parameter Lite and a 21B-parameter Pro size, and the authors report that Pro matches or beats the strongest open-source baseline and two commercial systems in human side-by-side comparisons, with a clear preference on lip sync. Distilled variants run in 16 sampling steps without classifier-free guidance for a reported 5.4x end-to-end speedup, which is the number to watch if you are costing out avatar video at volume.