Kandinsky 6.0 Video released as open audio-video model
Kandinsky 6.0 Video generates five-second clips with matching sound, including lip-synced speech, from text or an image. The Lite and Pro models, code and weights are released under the MIT license.
What it does
Kandinsky 6.0 Video is a family of two open models, Lite (3B parameters) and Pro (29B parameters), that make video and sound together from a text prompt or a starting image. Each clip is 5 seconds with 44 kHz audio, including lip-synced speech.
A video stream and a separately trained audio stream are linked so picture and sound stay in step. A built-in super-resolution step raises the output to Full HD (1920×1080).
Results and limits
In the authors' side-by-side human tests, the Pro model clearly beats Kandinsky 5.0 Video Pro and stays competitive with leading audio-video models, especially on speech quality. Reinforcement learning cut the word error rate of generated speech by 47% for Pro.
The paper lists limits plainly: clips top out at 5 seconds, base generation runs at SD resolution with Full HD reached by upscaling, and a gap remains to the strongest closed systems in visual and overall audio quality.
Worth knowing
Code, weights and diffusers integration are released under the MIT license. The paper includes memory presets for 32, 24 and 16 GB GPUs.
Try it yourself
[Shot type] of [character] in [location], [lighting]. The character says: "[one short line of dialogue]". Sound: [room tone or ambience], [one sound effect].