Researchers from the Kandinsky team announced two new speech‑to‑video models on 15 October 2026. The models, called Kandinsky S2V 1.0 Lite and Kandinsky S2V 1.0 Pro, generate full‑length video from a single reference image, an audio waveform, and an optional text prompt. They run at 25 frames per second and aim to sync lips, facial expression, pose, and hand gestures with the spoken words.
What the models can do
Both versions use a Qwen2.5‑Omni language model for text and audio conditioning and a three‑scale visual memory that keeps the reference image in view. The Pro model has 21 billion parameters, while the Lite version is a smaller 2 billion‑parameter model. Training used 1.17 million single‑speaker video clips with English and Russian captions.
Performance and availability
In a side‑by‑side human study, the Pro model performed as well as or better than the strongest open‑source baseline and two commercial systems, especially for lip synchronization. On standard speech‑to‑video benchmarks the Lite and Pro models rank among the top five across five metrics. A distilled variant called π‑Flow runs 5.4 × faster, producing video in 16 sampling steps without extra guidance, while matching or exceeding the base model on most scores. The authors released the code, model checkpoints, and a Diffusers integration under the MIT license, so anyone can download and run the models.
Why it matters
For creators who want to add a talking avatar to a video, the open‑source models let them generate realistic speech‑driven animation without paying for a commercial API. The faster distilled version makes it feasible to run the process on a single high‑end GPU, opening low‑cost options for indie developers and educators.
