Kandinsky Lab
Kandinsky Lab publishes the Kandinsky model family and related research. Its history runs from Russian text-to-image models to synchronized video and sound. Earlier releases credit Sber AI and AIRI; the current lab presents itself as a research team.
Company & leadership
- Organization
- Research team and model-development project.
- Project origins
- The published history begins with Malevich and Kandinsky 1.0 in 2021–2022.
- Developer lineage
- Sber AI and AIRI are the credited research affiliations for Kandinsky 3.0.
- Project supervisor
- Denis Dimitrov, as credited in the Kandinsky 5.0 release.
- Research track leaders
- Vladimir Arkhipkin, Vladimir Korviakov, Nikolai Gerasimenko and Denis Parkhomenko, in the 5.0 credits.
- Website operator
- Denis Valeryevich Dimitrov is the named personal-data operator for kandinskylab.ai.
A research project with several contributing organizations
The 3.0 research page lists Sber AI as the authors' main affiliation. Andrey Kuznetsov and Denis Dimitrov also list AIRI. Its acknowledgements credit SberDevices. These credits establish model-development roles; they do not establish a separately incorporated company called Kandinsky Lab.
The early Russian text-to-image work had a wider team
The ruDALL-E release credits Sber AI, SberDevices, Samara University, AIRI and SberCloud. It reports training both Malevich and the larger Kandinsky model on the Christofari cluster. These are developer credits for the early project, not evidence that one current lab owns every contributor.
The architecture changed across generations
The official history starts with autoregressive image generation in 2021–2022. Diffusion replaced that approach in the 2.x family. Image and video branches followed in 2023. The 2025 generation added distinct image, editing and video models.
Release credits name the people responsible
The 5.0 repository separates video, image and editing, data, and research work. Nikolai Vaulin, Nikita Kiselev and Alexander Varlamov receive the image and editing credits. Denis Dimitrov is project supervisor. This profile uses those stated roles without assigning a CEO or legal owner.
Current releases cover generation and its supporting models
The 6.0 video release introduced synchronized picture and sound on 6 October 2026. Later inference additions include first-and-last-frame conditioning and audio/video conversion modes. These modes use the released model family; they are not counted as extra trained checkpoints.
Image Pro moved to a shared generation and editing model
Sber announced Kandinsky 6.0 Image Pro on 28 April 2026. It combined generation and editing in a mixture-of-experts model and offered it through GigaChat. The release also described retrieval of reference images to add visual context.
Checkpoint roles remain separate from download formats
Practical guidance. The guides distinguish base models, fine-tuned models, distilled models and trained adapters. Diffusers packages and precision formats stay with their matching release. Autoencoders and classifiers are identified as components. Older product versions with no confirmed current access are marked accordingly.