Kandinsky 6.0 pipeline: offload, the untimed distilled default and what to measure

Kandinsky Lab published Kandinsky 6.0 on GitHub on 6 October under MIT for code and weights: text or text-plus-image in, a 5-second clip with soundtrack out, in Lite (3B) and Pro (29B) sizes. A Pro HD clip takes 292 s on an H100, 1,194 s on an RTX 4090 and 2,716 s on an RTX 5060 Ti; every GPU preset offloads weights, and the distilled Pro loaded by default has no published timings.

DEV Community

Open source

Kandinsky 6.0 pipeline: offload, the untimed distilled default and what to measure

Pro weighs about 90 GB, so every GPU preset offloads weights to host RAM
Pro HD clip: 292 s on H100, 1,194 s on RTX 4090, 2,716 s on RTX 5060 Ti
Presets and the demo load a distilled Pro at 10 passes instead of 50, with no timings
Full HD upscale on H100: 41 s for 2x and 118 s for 4x; audio must be reattached

#Kandinsky#Nvidia#HuggingFace

Kandinsky 6.0 pipeline: offload, the untimed distilled default and what to measure

Kandinsky 6.0 on a 16 GB GPU: official recipe OOMs, only the 29B model passed the Spanish test

Kandinsky 6.0 Video: open model generates video with synced audio

Study links AI use to lower critical thinking via cognitive offloading

Post-training digest: distillation leaks teacher behavior, reward models surpass teachers