kandinskylab/kandinsky-5

Kandinsky 5.0: A family of diffusion models for Video & Image generation

What it solves

Kandinsky 5.0 provides a family of high-quality diffusion models for generating videos and images from text prompts or input images. It specifically addresses the need for high-fidelity visual generation with strong support for both English and Russian languages and concepts.

How it works

The project utilizes a latent diffusion pipeline with Flow Matching. The core generative backbone is a Diffusion Transformer (DiT) that uses cross-attention to condition on text embeddings provided by Qwen2.5-VL and CLIP. For video processing, it employs the HunyuanVideo 3D VAE to encode and decode video into a latent space.

Who it’s for

  • Researchers and enthusiasts who can use the pretrain models for further fine-tuning.
  • Content creators looking for high-definition video (up to 10 seconds) and image generation.
  • Developers seeking lightweight, efficient models (like the 2B Video Lite version) for faster inference.

Highlights

  • Diverse Model Family: Includes "Pro" (19B parameters) for maximum quality and "Lite" (2B for video, 6B for image) for efficiency.
  • Multimodal Capabilities: Supports Text-to-Video (T2V), Image-to-Video (I2V), and Text-to-Image (T2I).
  • Optimized Performance: Offers CFG-distilled and Diffusion-distilled variants for significantly lower latency (up to 6— faster).
  • Language Support: Strong understanding of Russian concepts alongside English.
  • Controllable Motion: Includes LoRAs for camera control in video generation.