Kandinsky LabKandinsky 2
Kandinsky 2.2
Kandinsky 2.2 changes the CLIP representation used by the image-generation pipeline. The prior was retrained and the image diffusion model was fine-tuned for the new representation.
Use the prior that matches this decoder
Training included several resolutions and aspect ratios. The release describes 1024 × 1024 output and non-square formats. The diffusion stage uses a MoVQ decoder to produce image pixels.kandinsky-community/kandinsky-2-2-decoder — model card and files
Keep generations and benchmarks separate
Practical guidance. A benchmark table about Kandinsky 2.1 is not a measurement of the 2.2 checkpoint. Compare results made with the actual release and settings you plan to use.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | kandinsky-community/kandinsky-2-2-decoder |
| Weight license | Apache 2.0kandinsky-community/kandinsky-2-2-decoder — model card and files |