🔬Kandinsky 6.0 Video: 29B Model, Synced Audio, MIT License
TL;DR
Team Kandinsky released two diffusion models that generate video and synchronized 44 kHz audio, including lip-sync: a 3B Lite and a 29B Pro. Code, checkpoints and diffusers integration ship under the MIT license.
Team Kandinsky released two diffusion models that generate video and synchronized 44 kHz audio, including lip-sync: a 3B Lite and a 29B Pro. Code, checkpoints and diffusers integration ship under the MIT license.
Key Points
Two sizes: Video Lite at 3B parameters and Video Pro at 29B
Clips are 5 seconds long with a built-in super-resolution step to 1920x1080
Dual-stream design: a pretrained video stream plus a new audio stream joined by bidirectional cross-attention
Human side-by-side tests show Pro clearly beats Kandinsky 5.0 Video Pro and competes on speech quality
Training adds supervised fine-tuning, RL post-training and distillation
Why It Matters
An MIT-licensed audio-video model at 29B gives builders a self-hostable alternative to closed generators. The 5-second cap is the catch.
Quick Facts
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.