Kandinsky 6.0 Video goes open source: 3B/29B models generate clips with synced audio
lmoroney · x · 2026-10-13
Kandinsky Lab has released the Kandinsky 6.0 Video family on Hugging Face under MIT, focused on generating video and sound in a single pass.
- Two variants: Lite (3B params) and Pro (29B), both supporting text-to-audio-video (T2AV) and image-to-audio-video (TI2AV)
- Output: 5-second clips with synchronized 44 kHz audio including lip-sync; a plug-in super-resolution model scales output to Full HD (1920×1080)
- Architecture: video and audio streams coupled via bidirectional cross-attention, trained jointly so sync is native rather than post-dubbed
- Release: code, checkpoints, and Diffusers pipelines all open (arXiv 2610.05608); Lite supports CPU offload for consumer hardware
The author also suggests a practical test: bake audio cues with timestamps into the prompt, generate, then mute/unmute while watching — the fastest way to gauge how well sync holds.
Related event: Kandinsky 6.0 Open-Sourced with Synchronized Audio-to-Video Generation(2 posts)→
More from Multimodal
- Free Krea2 Character & Style Catalog v0.9: 3,919 samples with ComfyUI workflows — Savings_Prune_8792 · 2026-10-13
- Dev trains two Anima ControlNet-LLLite models for 3D poses and color regions — toyxyz · 2026-10-13
- Midjourney-generated virtual influencer posts lifelike morning portrait — tisch_eins · 2026-10-13
- Training a DLSSNR cartoon NPR rendering style: new lines data visibly closer to target — bdsqlsz · 2026-10-13
- Dev wires Claude/Codex into a 3D Gaussian splat editor via MCP — y0un51n · 2026-10-13
- Krea 3 image model is 'very very close' as fal.ai posts prompting guide — Oren_Lester · 2026-10-13