>
K
Introduction
We are releasing Kandinsky 5.0 Video Lite, the first model in the new Kandinsky 5 series. The model runs at 768x512 resolution and, despite its compact size of just 2 billion parameters, demonstrates quality superior to previous versions of Kandinsky and most current open-source state-of-the-art solutions. A key focus is efficiency: the model is compact, requires fewer resources, and generates faster. This result was achieved through a comprehensive approach—from data collection and preparation to pretraining and fine-tuning. We explored modern architecture optimization methods and applied our own developments to balance quality and speed.
Training
Training Code Accessibility
Apache 2.0
https://github.com/ai-forever/Kandinsky-5
https://huggingface.co/ai-forever/Kandinsky-5.0-T2V-Lite-pretrain-5sSize Notes: To create the pretrain dataset, we collected a massive dataset of 6 billion images and 35 million videos, which we then sliced (using the pyscenedetect scene change detector) into 1.5 billion short scenes ranging from 2 to 60 seconds. We then filtered out samples that: were too low-resolution: up to 256 pixels on the shortest side; were duplicates and very similar; were watermarked; were overloaded with text (document photos, etc.); were not dynamic enough. From the remaining data, we selected 124 million scenes and 520 million images that were the most aesthetically pleasing and technically sound.
Parameters
Parameters
2000000000Notes: 2B
Authors
Project Leader: Denis Dimitrov
Team Leads: Vladimir Arkhipkin, Vladimir Korviakov, Nikolai Gerasimenko, Denis Parkhomenko
Core Contributors: Alexey Letunovskiy, Maria Kovaleva, Ivan Kirillov, Lev Novitskiy, Denis Koposov, Dmitrii Mikhailov, Anna Averchenkova, Andrey Shutkin, Julia Agafonova, Olga Kim, Anastasiia Kargapoltseva, Nikita Kiselev
Contributors: Anna Dmitrienko, Anastasia Maltseva, Kirill Chernyshev, Ilia Vasiliev, Viacheslav Vasilev, Vladimir Polovnikov, Yury Kolabushin, Alexander Belykh, Mikhail Mamaev, Anastasia Aliaskina, Tatiana Nikulina, Polina Gavrilova
Language
Image generation