Kandinsky 6.0 Video targets synchronized video and audio. Public Lite and Pro artifacts make a focused trial possible; parameter count and an offload example do not establish affordable, fast generation on your own GPU.
A video-and-audio family rather than one model
The research describes synchronized video and sound generation, with Lite and Pro variants. The paper presents a joint audiovisual approach rather than a workflow where an unrelated sound generator is appended afterward. For creators, synchronization is the question to evaluate: a plausible soundtrack can still fail a shot if an impact, spoken sound or visible movement arrives at the wrong moment.
Lite, Pro and the setup evidence
The official repository identifies Lite as 3B and Pro as 29B, with five-second clip examples and 44 kHz audio. It links public checkpoints and integration paths and marks code MIT. Its Pro example uses CPU offload with an 80GB GPU. That configuration is not a stated minimum and does not prove every consumer setup works. Dependency and checkpoint terms still deserve their own review. No FyreLinkz runtime benchmark has been performed.
Choose the first shot before choosing the larger model
Our suggested trial uses one simple subject and one visible sound-producing action. A hand placing a cup is easier to assess than a crowd scene with music and dialogue. Write the intended action, framing, duration and audio cue before sampling. Compare Lite and Pro against that same brief if your resources allow. Score whether the shot is usable, not merely whether a larger model creates more texture. Record rejection reasons such as object deformation, ambiguous timing or an audio event with no visible cause.
Budget accepted shots rather than raw generations
A useful cost sheet starts when the machine is provisioned and ends when an accepted clip is exported. Include downloads, startup, retries, decoding and human review. Separate cold-start and already-loaded runs so an occasional user does not inherit an unrealistic warm-worker estimate. If a cloud instance bills in a different currency, preserve the original quoted rate and identify taxes separately. The relevant denominator is an approved shot with usable audio, not a model’s advertised duration. A cheaper failed sample is still part of the production bill.
| Proposed check | Keep fixed or record | Question it answers |
|---|---|---|
| Picture | Same subject and brief in each candidate | Continuity and action completion |
| Audio | One visible sound-producing event | Timing, plausibility and synchronization |
| Delivery | Review every candidate including rejections | Resource time per accepted clip |
Inspect picture and sound together
Our proposed review alternates muted and unmuted playback. Muted review exposes motion and continuity problems; restored sound exposes timing and distractingly mismatched atmosphere. Look at cut boundaries if several generated shots will share a timeline. Keep output resolution, seed, version and offload settings attached to each candidate. The video comparison checklist provides a way to frame alternatives, but no table can replace an exact test. Start with a small reproducible brief before committing client delivery dates or purchasing hardware.
Frequently asked questions
What are the Kandinsky 6 Lite and Pro sizes?
The official repository describes 3B Lite and 29B Pro variants.
Does the 80GB Pro example define a minimum?
No. It is a specific offloaded example, not a universal minimum specification.
Are the results in this article independently measured?
No. We review official documentation and propose a controlled creator trial.
Sources & further reading
Primary sources checked Oct 11, 2026. Vendor statements are attributed; editorial advice is our own.
- 1Official Kandinsky 6 repository ↗Kandinsky Lab
- 2Kandinsky 6.0 Video paper ↗Kandinsky authors
Help us keep this useful. Check contact availability →
