Kandinsky Lab open-sources S2V 1.0: audio-driven avatar video from one still, MIT licensed
lmoroney · x · 2026-10-16
Kandinsky Lab released Kandinsky S2V 1.0 (Lite 2B / Pro 21B), an open speech-to-video family that turns a reference image, speech waveform and optional prompt into 25 FPS video with speech-aligned lips, expression, pose and gesture. Text/audio conditioning comes from one Qwen2.5-Omni model with a three-scale visual memory across 5-second chunks; training used 1.17M single-speaker clips with English and Russian captions. The π-Flow distilled Pro runs in 16 sampling steps for a 5.4× end-to-end speedup, and human evals rank Pro at or above top open and proprietary baselines, especially on lip sync. Code, checkpoints and Diffusers integration ship under MIT.
Related event: Kandinsky Open-Sources Speech-to-Video Talking-Head Model S2V 1.0(2 posts)→
More from Multimodal
- CoWorld Reasoner uses an evolving video world model as a visual reasoning engine — ModalityDance · 2026-10-16
- PixelLab tackles ComfyUI pain points: model dedup, image triage, workflow repair — SmartRick137 · 2026-10-16
- Everything is instantly a playable game: browser Roadster Rally demo — bennash · 2026-10-16
- AI recreates Star Wars scenes with stunning neon-smoke cinematic lighting — azed_ai · 2026-10-16
- Stunning 3D Gaussian Splatting Scan of Dandelion Seeds Made with LichtFeld Studio — janusch_patas · 2026-10-16
- Grok writes its own Imagine prompts and edits a full 'grandmageddon' video solo — techhalla · 2026-10-16