Kandinsky 6.0 in a pipeline: offload, the untimed default and what to measure
Say your product manager wants "generate a short video with sound" in the app. Kandinsky 6.0, which Kandinsky Lab put on GitHub on 6 October under MIT for both code and weights, is a candidate: text, or text plus an image, in; a 5-second clip with a matching soundtrack out. It comes in two sizes, Lite with 3B parameters and Pro with 29B (parameters are the learned numbers inside a model, a rough…
Kandinsky 6.0, released by Kandinsky Lab on GitHub, is a model that generates short videos with sound, either from text alone or text combined with images. The model comes in two sizes: Lite, with 3 billion parameters, and Pro, with 29 billion parameters. However, there is no information about the quality of the model's output. The model's timing, scheduling, and hardware requirements are not provided in detail, leaving it as a batch job.
The Lite model requires significant memory, and the Pro model, despite having 16 more GB of GPU memory, takes 561 seconds to generate a 5-second clip with a matching soundtrack. The offload feature, which is essential for serving the model, is not explicitly mentioned in the CLI flags, but it is required for 80 GB cards. The default model, a distilled version of Pro, was not included in the timing table.
The full story on Kandinsky 6.0 video with sound can be found in the InferenceKey article on Kandinsky 6.0 video with sound.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.