Kandinsky LabKandinsky 3
Kandinsky 3.0
Kandinsky 3.0 expanded its text understanding and image generator. The training data also gave attention to Russian cultural subjects.
The total pipeline is larger than its image generator
The published architecture also includes a 267M MoVQ encoder/decoder. The text encoder and image autoencoder serve different roles from the diffusion U-Net.ai-forever/Kandinsky3.0 — model card and files
Later packages include the base and new trained components
The 3.1 repository packages the base generator, Flash and updated inpainting weights. Flash's use as a refiner is an inference role of that trained checkpoint.Kandinsky 3: base, Flash and inpainting releases
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | ai-forever/Kandinsky3.0ai-forever/kandinsky3-diffuserskandinsky-community/kandinsky-3 |
| Diffusion U-Net | 3B parametersai-forever/Kandinsky3.0 — model card and files |
| Text encoder | Flan-UL2 encoder, 8.6B parametersai-forever/Kandinsky3.0 — model card and files |
| Weight license | Apache 2.0ai-forever/Kandinsky3.0 — model card and files |
Related models
- Kandinsky 3 Flash
Kandinsky Lab · Kandinsky 3
A distilled four-step Kandinsky 3 image model also used as the 3.1 release's refiner.
Launch date not confirmedDownloadable