Kandinsky 5.0 Image Lite
This 6B text-to-image checkpoint supports square and wide image formats. The released model targets image quality and Russian text and concepts.
Generate an image from its description
The release includes separate pretrained and default checkpoints. The native repository name does not assign an extra model to each supported aspect ratio.kandinskylab/Kandinsky-5.0-T2I-Lite — model card and files
Load the corresponding image pipeline
The Diffusers export is identified as supervised fine-tuned. It uses Qwen2.5-VL and CLIP text encoders and a FLUX VAE. The native model table uses the matching SFT configuration.Kandinsky 5.0 T2I Lite SFT — Diffusers model cardkandinskylab/Kandinsky-5.0-T2I-Lite — model card and files
The overview also names an RL training stage
The current image overview discusses RL-fine-tuned and SFT-soup variants. A separately named RL download was not established in this catalog review; the verified export here is labelled SFT.Kandinsky Lab: current image models
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | kandinskylab/Kandinsky-5.0-T2I-Litekandinskylab/Kandinsky-5.0-T2I-Lite-sft-Diffusers |
| Generator size | 6B parameterskandinskylab/Kandinsky-5.0-T2I-Lite — model card and files |
| Weight license | MITkandinskylab/Kandinsky-5.0-T2I-Lite — model card and files |
| Example image sizes | 1024 × 1024 and 1280 × 768kandinskylab/Kandinsky-5.0-T2I-Lite — model card and files |