Captured photo
Kandinskiy 2.2
60

About

Kandinsky 2.2 is an open-source text-to-image model that produces high-resolution, photorealistic and stylized images from natural language prompts while also supporting guided image editing and blending. It delivers detailed textures, realistic lighting and fine detail up to 1024×1024, making it suitable for portraits, landscapes, cityscapes, fantasy concepts and abstract art. Improved text–image alignment helps the model follow complex instructions and yield consistent results across varied prompts. Beyond standard text-to-image generation, Kandinsky 2.2 offers text-guided image manipulation: you can inpaint, extend, or alter parts of an existing photo by describing changes in words. The model also supports image interpolation and morphing between sources, letting you create hybrid visuals or smooth transitions guided by text. Integration with ControlNet gives precise control over composition—condition generation on sketches, edge maps or other structural inputs—to produce repeatable, layout-driven outputs that are valuable for concept artists, illustrators and designers. Practically, Kandinsky 2.2 is useful for rapid prototyping in advertising, game and film concepting, portrait and landscape generation, and image restoration or creative recomposition. Its open-source, permissive licensing enables customization and integration into research and commercial pipelines. Note that high-quality and high-resolution outputs require significant GPU resources, and the model can reflect dataset biases or occasionally produce artifacts. Overall, Kandinsky 2.2 balances realism, controllability and versatility, offering creatives and researchers a powerful tool for generating and editing images from text prompts.

Percs

High quality
Fine control

Settings

Negative Prompt-  Things you don't want to see in the output
Width
Height-  undefined
Inference Steps-  Number of denoising steps
Inference Steps Prior-  Number of inference steps for prior