Kandinsky 3
Text-to-image model by Sber AI with latent diffusion architecture.
About
Sber AI's Kandinsky 3 is a large scale text-to-image model based on latent diffusion. The pipeline pairs an 8.6 billion parameter Flan-UL2 text encoder with a 3 billion parameter diffusion U-Net and a MoVQ image encoder and decoder, giving it stronger prompt understanding than earlier Kandinsky releases. Beyond text-to-image generation the repository covers inpainting, outpainting, image fusion, and image variations, while a distilled Kandinsky Flash variant samples in four steps for roughly three times faster inference. Extensions include an HED based ControlNet, an IP-Adapter for image conditioning, and the KandiSuperRes upscaler. Training paid particular attention to Russian cultural content, and the model handles both Russian and English prompts. Code and checkpoints are openly released, with demos on Hugging Face and the FusionBrain web service. Researchers and developers seeking an alternative to Stable Diffusion lineage models are the typical users. Apache 2.0 license.
Should you use Kandinsky 3?
Pick it when
Pick Kandinsky 3 when prompts are in Russian or need Russian cultural content, or when you want an Apache-2.0 model outside the Stable Diffusion lineage that bundles inpainting, outpainting, image fusion, and a 4-step Flash variant.
Look elsewhere when
Skip it if you depend on community LoRAs, checkpoints, and UI support, which center on SDXL and FLUX. For Apache-2.0 with wider adoption, FLUX.1 schnell or Chroma are better bets, and Kandinsky 5 is the newer family.
Alternatives to Kandinsky 3
- Kandinsky 5.0
The newer family with a 6B image model and editing variant plus video models under MIT, at a roughly 12 GB floor for the smaller models.
- Stable Diffusion XL
Similar latent diffusion class with a far larger ecosystem of fine-tunes and ControlNets, but no Russian-focused training and a RAIL-family license.
- FLUX.1
Schnell is Apache-2.0 too, with much broader frontend support and companion editions, but needs 12 GB or more.
- Kolors
The analogous pick for Chinese rather than Russian prompts, with a similar add-on set of IP-Adapter, ControlNet, and inpainting.
Reviews (0)
Leave a Review
No reviews yet. Be the first to review!
Details
- Category
- Image Generation
- Price
- Free
- Platform
- Local/Desktop
- Difficulty
- Intermediate (3/5)
- License
- Apache-2.0
- Minimum VRAM
- 8 GB
- Added
- Apr 3, 2026