Kandinsky 2 2 decoder

Kandinsky 2.2 is a free Russian neural network for image generation developed by Sber AI. It operates based on a diffusion model: initially adding noise to images it was trained on, then restoring them through a reverse diffusion process to create new unique images.

Key features of version 2.2:

• Improved image quality. Generated images have resolutions up to 1024 pixels with a high level of detail.
• Aspect ratio selection. Previous versions allowed only square images; this update adds versatility in aspect ratios.
• Sticker generation. Adding the word "sticker" to the prompt automatically removes the background.

Kandinsky 2.2 supports prompts in both Russian and English. It can create images in over 20 artistic styles, blend multiple images, stylize images based on text descriptions, generate images similar to given examples, inpaint missing parts of images, and create artwork in an infinite canvas mode (outpainting).

---

The model is a component of the image generation pipeline, consisting of:

• VQModel: ~67M parameters,
• UNET: ~1.3B parameters.

Total: ~1.3B parameters

---

The 'VRAM requirements' line is calculated based on the size of the weights in the specified precision, with an additional +10% coefficient for inference:
  model_parameters_count * bytes_per_parameter * 1.1 / 1024^3 
where model_parameters_count is number of model parameters, bytes_per_parameter is number of bytes used to store one parameter. For 16 bit = 2, 8 bit = 1, 4 bit = 0.5.

Announce Date: 12.07.2023
Parameters: 1.3B
Context: 77
VRAM requirements: 0.7 GB using 4 bits quantization, 1.4 GB using 8 bits quantization, 2.7 GB using 16 bits quantization
Developer: Sber AI
License: Apache 2.0

Use our pre-built public endpoints for free to test inference and explore Kandinsky 2 2 decoder capabilities. You can obtain an API access token on the token management page after registration and verification.

Rent your own physically dedicated instance with hourly or long-term monthly billing.

We recommend deploying private instances in the following scenarios:

• maximize endpoint performance,
• enable full context for long sequences,
• ensure top-tier security for data processing in an isolated, dedicated environment,
• use custom weights, such as fine-tuned models or LoRA adapters.

Contact our dedicated neural networks support team at nn@immers.cloud or send your request to the sales department at sale@immers.cloud.