🔬 Kandinsky 6.0 Video: Training Details and Architecture

The new Kandinsky 6.0 Video lineup — the flagship Pro (29B) and the lightweight Lite (3B) — uses a dual-stream architecture: a video stream and an audio stream connected by bidirectional cross-attention.

📚 Training methodology:
• Audio stream: pretrained from scratch on 40 million audio tracks; joint training: 7 million audio-video segments
• Selection criteria: high-quality sound, close-up shots of speaking people
• Objective: precise synchronization of speech and facial expressions

🏗 Architecture:
Two connected streams:
1. Video stream
2. Audio stream
The streams are connected by bidirectional cross-attention for temporal and semantic alignment.

🎯 Key capabilities:
• Lip-sync technology
• Wide audio range (speech, effects, ambient)
• Consumer GPU deployment (block offloading, 16–32 GB VRAM presets)
• Realistic physics modeling

📊 Performance metrics:
• vs Kandinsky 5.0: clear advantage across all criteria (visual quality, prompt adherence, camera movement)
• vs LTX 2.5: preferred in visual quality and speech quality
• vs Veo 3.1 Fast: superior in image animation tasks

🔧 Specifications:
• Resolutions: SD, HD, Full HD
• Audio quality: 44 kHz
• Duration: up to 5 seconds
• License: MIT

🔗 Hugging Face