Kandinsky LabKandinsky Video
Kandinsky Video 1.0
The original video model generates sparse keyframes and then interpolates frames between them. Its FusionFrames design separates appearance generation from temporal interpolation.
The release includes trained temporal components
The keyframe output passes to a latent interpolation model that predicts three frames between keyframes. The package contains separate generator and interpolation weight files.ai-forever/KandinskyVideo — model card and files
Keep the original pipeline components together
Practical guidance. Changing the keyframe model or interpolator changes the generated sequence. Compare the complete pipeline when reproducing an older result.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | ai-forever/KandinskyVideo |
| Weight license | Apache 2.0ai-forever/KandinskyVideo — model card and files |
| Keyframe U-Net | 4.0B parametersai-forever/KandinskyVideo — model card and files |
Related models
- Kandinsky 3.0
Kandinsky Lab · Kandinsky 3
An image diffusion model with a larger text encoder and U-Net, trained with attention to Russian cultural subjects.
Launch date not confirmedDownloadable