Kandinsky 6.0 Video

A line‑up of lightweight (3B parameters), high‑speed models for text‑to‑audio‑video and image‑to‑audio‑video generation. Produces 5‑second clips with synchronized 44 kHz audio, including lip‑sync, at SD resolution, upscalable to Full HD through a built‑in super‑resolution model

• SFT + RL models — two‑stage fine‑tuning with model soup across 11 domains, followed by OmniNFT‑based reinforcement learning that improves speech intelligibility and audio‑video synchronization
• Distilled models — π‑Flow trajectory distillation to 10 NFE, followed by Sim‑LADD adversarial refinement for perceptual fidelity
• Pretrain checkpoints — for further fine‑tuning and research

A line‑up of high‑capacity (29B parameters) models for text‑to‑audio‑video and image‑to‑audio‑video generation. Delivers state‑of‑the‑art visual fidelity, cinematic motion dynamics, precise camera control, and synchronized 44 kHz audio with accurate lip‑sync

• SFT + RL models — two‑stage fine‑tuning with model soup across 11 domains, followed by OmniNFT‑based reinforcement learning that improves speech intelligibility and audio‑video synchronization
• Distilled models — π‑Flow trajectory distillation to 10 NFE, followed by Sim‑LADD adversarial refinement for perceptual fidelity
• Pretrain checkpoints — for further fine‑tuning and research