Kandinsky 3.0

Kandinsky-3 is a diffusion model for text-to-image generation, developed based on previous versions of the Kandinsky2-x family. It has been improved through expanded data volume, including information related to Russian culture, enabling the generation of images reflecting this theme. The model also demonstrates enhanced text understanding and improved visual quality due to larger text encoder and Diffusion U-Net model sizes.

The model comprises three components:

• Text Encoder based on Flan-UL2 (part of the encoder) with 8.6 billion parameters.
• Latent Diffusion U-Net with 3 billion parameters.
• MoVQ Encoder/Decoder with 267 million parameters.

The base model was trained over 2 million steps using 400 NVIDIA A100 GPUs. The inpainting model initialized from the base model and further trained over 250,000 steps using 300 NVIDIA A100 GPUs.

Possible generations include:

• Artistic styles of specific artists (e.g., Alfons Mucha, Claude Monet).
• Detailed scenes (e.g., "dragon fruit in a realistic style").
• Fantastical or surreal scenes (e.g., "a palace against the Milky Way backdrop")

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

• VQModel: ~271M parameters,
• T5 text encoder: ~8.7B parameters,
• UNET: ~3B parameters.

Total: ~12B parameters

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 support@immers.cloud or send your request to the sales department at sale@immers.cloud.