Is Kandinsky S2V 1.0 Lite state of the art?Pending benchmarks.Announced Oct 13, 2026; no independent results yet.
The 2B size of Kandinsky Lab's Kandinsky S2V 1.0 speech-to-video family: it animates a person from a single reference image, a speech track and a text prompt, with lip motion, expression, pose and gesture following the speech. It does not generate audio. Output is 25 fps in SD, with long speech generated in 5-second chunks. Initialized from Kandinsky 5.0 Video. Code and weights under MIT.
The details so far
- Lab
- Kandinsky
- Announced
- Oct 13, 2026
- API
- Not yet
- Open weights
- Published
- Parameters
- 2B
- Task
- Speech-to-video (audio-driven avatar animation) from one reference image, a speech waveform and a text prompt; audio is a condition only
- Output
- 25 fps video in SD at the reference frame's aspect ratio; long speech in 5-second chunks
- Architecture
- 3D Diffusion Transformer with flow matching in a frozen Hunyuan video VAE latent; Qwen2.5-Omni text and audio front end; three-scale FramePack visual memory
- Training data
- 1.17M single-speaker clips, captions in English and Russian
- Checkpoints
- Base, pi-Flow distilled (16 steps instead of 50) and Diffusers checkpoints on Hugging Face; demo Space
- License
- MIT (code and weights)
- Related models
- Separate from Kandinsky S2V 1.0 Pro and from Kandinsky 6.0 Video
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: in a side-by-side human evaluation, Kandinsky S2V 1.0 Pro performs comparably to or better than the strongest open-source baseline and two proprietary commercial models on average, and is preferred for lip sync.
Lab claim: the distilled Pro model runs in 16 steps without classifier-free guidance, a 5.4x end-to-end speedup over the base model.
Sources
PAPERKandinsky S2V 1.0: Audio-Driven Avatar Video Generation (arXiv)arxiv.org/abs/2610.19290 · Oct 13, 2026REPOkandinskylab/Kandinsky-S2V-1.0-Lite model card (Hugging Face)huggingface.co/kandinskylab/Kandinsky-S2V-1.0-LiteREPOkandinskylab/kandinsky-s2v (GitHub)github.com/kandinskylab/kandinsky-s2vLABKandinsky S2V 1.0 project pagekandinskyavatar.github.io/Tracking continues until an independent board lists it.