Showreel
New · Flagship

Kandinsky 6.0 Image Pro

A major update to the Image line, moved to a Mixture‑of‑Experts architecture and focused on editing quality and speed — over 40% faster than Kandinsky 5.0 thanks to more efficient inference parallelization and attention

  • Image RAG — retrieves relevant reference images from a knowledge base to enrich context during both generation and editing, including Russian cultural motifs, characters, and artistic styles
  • Broader editing toolkit — inpainting, object removal, style transfer, photo restoration and colorization, interior/exterior design generation
  • Competitive with Flux 2 Max and outperforms GPT Image 1.5 in side‑by‑side comparisons
  • 3‑reference editing and precision brush controls — combine up to three reference images and paint exactly where an edit should apply
Habr Article
Generated with Kandinsky 6.0 Image Pro
Previous release

The previous generation — open text‑to‑image and instruction‑based editing models, still available and fully documented below

Image

6B · SOTA

A high‑resolution text‑to‑image generation model that achieves state‑of‑the‑art visual quality and prompt alignment. Outperforms leading open‑source models like FLUX.1[dev] and Qwen‑Image in aesthetic realism and compositional accuracy

  • RL‑finetuned model — delivers the highest visual fidelity and realism
  • SFT‑soup model — excels in prompt following and overall visual quality
  • Pretrain checkpoint — for researchers to conduct further fine‑tuning and experimentation
  • Native Cyrillic text rendering — trained on 10M+ Russian‑text images in different styles, from print to handwriting
text‑to‑image ≤1024×1024

Image Editing

Instruct

A specialized variant derived from the base Image model, fine‑tuned on an instructive dataset for precise, context‑aware image editing — inpainting, object replacement, style transfer

  • Instruction‑following editing from a single prompt + source image
  • Preserves identity and composition outside the edited region
image‑to‑image instruct editing ≤1024×1024
Get the code
Generated with Kandinsky Image 5.0
Built for Russian

Trained on 10M+ images containing Russian text in a wide range of styles — from print to handwriting — for accurate recognition across different ways of writing

Read how it was trained

Previous versions

Apr 2024

Kandinsky Flash / 3.1

Flash (11.9B) generates images in just 4 steps — 10× faster than standard diffusion. 3.1 (15.5B) uses Flash as a refiner to boost quality and adds cultural awareness for Russian‑language prompts

Nov 2023

Kandinsky 3.0

11.9B parameters — high‑fidelity image generation with improved text alignment

Jul 2023

Kandinsky 2.2

4.8B parameters — generates photorealistic images at 1024×1024 resolution. Supports ControlNet for precise editing and 4‑second video clips

2022–2023

Kandinsky 2.0 / 2.1

2B / 3.3B parameters — diffusion‑based generation. Supports 101 languages. Enables inpainting, outpainting, and image editing at resolution up to 768×768