FusionFrames: efficient architectural aspects for text-to-video
generation pipeline
arxiv, 2023
arxiv, 2023
Read more
Kandinsky: an improved text-to-image synthesis with image prior and
latent diffusion
EMNLP 2023
EMNLP 2023
Read more
Kandinsky 3.0 technical report
arxiv, 2023
arxiv, 2023
Read more
CRAFT: Cultural Russian-oriented dataset adaptation for focused
text-to-image generation
Doklady Mathematics, 2024
Doklady Mathematics, 2024
Read more
Kandinsky 3: Text-to-image synthesis for multifunctional generative
framework
EMNLP 2024
EMNLP 2024
Read more
Improveyourvideos: Architectural improvements for text-to-video
generation pipeline
IEEE Access, 2024
IEEE Access, 2024
Read more
RusCode: Russian Cultural Code Benchmark for Text-to-Image
Generation
NAACL 2025
NAACL 2025
Read more
VIVAT: Virtuous Improving VAE Training through Artifact Mitigation
arxiv, 2025
arxiv, 2025
Read more
NABLA: Neighborhood Adaptive Block-Level Attention
arxiv, 2025
arxiv, 2025
Read more
Publications
Here we share the results of our research and development. We are
convinced that openness
and collaboration are the foundation of progress in the field of artificial intelligence. This section contains publications that describe in detail the architectural solutions, teaching methods, and stages
of creating Kandinsky.
and collaboration are the foundation of progress in the field of artificial intelligence. This section contains publications that describe in detail the architectural solutions, teaching methods, and stages
of creating Kandinsky.
Read more
A line-up of Text-to-Image
and Image Editing models that generate HD images (1280×768, 1024×1024) from English and Russian text prompts with precise details and accurate text rendering, powered by a 6 billion parameter architecture.
and Image Editing models that generate HD images (1280×768, 1024×1024) from English and Russian text prompts with precise details and accurate text rendering, powered by a 6 billion parameter architecture.
Image Lite
A line-up of Text-to-Video and Image-to-Video models that produce up to
10 seconds of high-quality SD video (768x512, 512x512, 24 fps) with
strong consistency, powered by 2 billion parameters Diffusion
Transformer
for fast, efficient generation.
for fast, efficient generation.
Video Lite
A line-up of Text-to-Video and Image-to-Video models designed for
high-quality HD video (1280x768, 24 fps) synthesis, delivering rich
motion dynamics and precise camera control, English and Russian prompts
understanding. Diffusion Transformer is scaled to 19 billion parameters
for superior visual quality.
Video Pro
2021-2022
Malevich/Kandinsky 1.0
1.3B/12B parameters
Text-to-image generation at 256x256 resolution. Autoregressive approach, trained on Russian-language data.
1.3B/12B parameters
Text-to-image generation at 256x256 resolution. Autoregressive approach, trained on Russian-language data.
July 2023
Kandinsky 2.2
4.8B parameters
Generates photorealistic images at 1024x1024 resolution. Supports ControlNet for precise editing and 4-second video clips.
4.8B parameters
Generates photorealistic images at 1024x1024 resolution. Supports ControlNet for precise editing and 4-second video clips.
November 2023
Kandinsky Video 1.0
17.5B parameters
Generates 8-second videos from text. Uses keyframe interpolation at 512x512 resolution.
17.5B parameters
Generates 8-second videos from text. Uses keyframe interpolation at 512x512 resolution.
May 2024
Kandinsky Video 1.1
17.5B parameters
Generates 5.5-second videos. Trained on 4.6M text-video pairs, with captions generated by LLaVA-1.5.
17.5B parameters
Generates 5.5-second videos. Trained on 4.6M text-video pairs, with captions generated by LLaVA-1.5.
June 2025
Kandinsky 4.1
14B parameters
Refined image generation with improved aesthetics, guided by expert artist feedback.
14B parameters
Refined image generation with improved aesthetics, guided by expert artist feedback.
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 768x768.
2B/3.3B parameters
Diffusion-based generation. Supports 101 languages. Enables inpainting, outpainting, and image editing at resolution up to 768x768.
November 2023
Kandinsky 3.0
11.9B parameters
High-fidelity image generation with improved text alignment.
11.9B parameters
High-fidelity image generation with improved text alignment.
April 2024
Kandinsky Flash/3.1
11.9B/15.5B parameters
Flash: Generates images in just 4 steps — 10x faster than standard diffusion.
3.1: Uses Flash as a refiner to boost quality. Adds cultural awareness for Russian-language prompts.
11.9B/15.5B parameters
Flash: Generates images in just 4 steps — 10x faster than standard diffusion.
3.1: Uses Flash as a refiner to boost quality. Adds cultural awareness for Russian-language prompts.
December 2024
Kandinsky 4.0
17.5B parameters
Generates images and 12-second videos from text. Supports audio generation conditioned on video input.
17.5B parameters
Generates images and 12-second videos from text. Supports audio generation conditioned on video input.
November 2025
Kandinsky 5.0
6B/2B/19B parameters
New family based on Flow Matching.
Includes:
6B/2B/19B parameters
New family based on Flow Matching.
Includes:
- Image Lite (6B) — generates high quality HD images and has strong text writing capabilities and Russian concept understanding. Variant of this model is capable to edit input images given text prompt.
- Video Lite (2B) — is a lightweight video generation model that ranks #1 among open-source models in its class.
- Video Pro (19B) — generates high quality HD images and has strong text writing capabilities and Russian concept understanding. Variant of this model is capable to edit input images given text prompt.
History
Our
Mission
Our mission is to move the industry forward by creating models and
approaches that open up new formats
of creativity, accelerate scientific discoveries, and unlock the potential of generative AI.
of creativity, accelerate scientific discoveries, and unlock the potential of generative AI.
About
Kandinsky Lab
