Kandinsky Open-Sources S2V 1.0: Speech-Driven Avatar Video From One Still Image
lmoroney · x · 2026-10-16
Kandinsky Lab released Kandinsky S2V 1.0, an open family of speech-to-video models with Lite (2B) and Pro (21B) variants. From a single reference image, a speech waveform, and an optional text prompt, they generate long-form 25 FPS video where lip motion, expression, pose, and gestures follow the speech.
Key details:
- Architecture draws text/audio conditioning from Qwen2.5-Omni plus a reference-aware three-scale visual memory; trained from Kandinsky 5.0 Video with an area-adaptive face loss
- Trained on 1.17M single-speaker clips with English and Russian captions
- In human evaluation, Pro matches or beats the strongest open-source baseline and two proprietary commercial models, and is preferred for lip sync
- π-Flow distilled variants: Pro distilled runs in 16 steps without CFG, a 5.4× end-to-end speedup with nearly no metric loss
Code, checkpoints, and diffusers integration are MIT-licensed for local research and deployment.
Related event: Kandinsky Open-Sources Speech-to-Video Talking-Head Model S2V 1.0(2 posts)→
More from Multimodal
- Krea 3 image model impresses with speed: blending three reference images into one — gizakdag · 2026-10-16
- Open-source ebook2audiobook turns any ebook into a free audiobook, 20.4k GitHub stars — thisguyknowsai · 2026-10-16
- AI-generated Nicole Kidman couture images stun with surreal sculptural silhouettes — OVolosin82152 · 2026-10-16
- Valtz v0.1.0.11 adds FLUX.3 support, local image gen up to 4K on Mac — TgoAI · 2026-10-16
- H3 Character Swap LoRA V1.5 released with open ComfyUI workflow — linoy_tsaban · 2026-10-16
- ForcingBooster halves computation of autoregressive video diffusion with patch-wise denoising — VAST-AI · 2026-10-16