We develop open neural networks for image and video generation, publish research, and share code and models with the AI community
A line-up of text-to-image and image editing models that generate HD images from English and Russian text prompts with precise details and accurate text rendering, powered by a 6 billion parameter Mixture‑of‑Experts architecture
Lite — text‑to‑video and image‑to‑video producing up to 10 seconds of SD video with strong consistency. Pro — HD video with rich motion dynamics and precise camera control, English and Russian prompt understanding, Diffusion Transformer scaled to 19 billion parameters
A family of open‑source variational autoencoders for audio, image, and video generation — compressing raw signals into compact continuous latents built for downstream diffusion models
Three specialized image‑to‑video models for autonomous driving, robotics, and complex physical interactions — generating physically plausible continuations of a scene
Three video generation models fine‑tuned on Kandinsky 5.0 Video Lite (2B parameters) for autonomous driving, robotics, and complex physics, with post‑training RL and SOAR. Ranks top‑4 on PAI‑Bench‑G, top‑5 on Physics‑IQ Verified, and top‑6 on RBench. Open source under MIT license.
Read →The new flagship, succeeding Kandinsky 5.0, works up to twice as fast, better understands complex queries, and creates more detailed images. Built on a new Mixture‑of‑Experts architecture, it brings free professional-level editing — object removal, restyling, retouching — with no generation limits.
Read →A 48 kHz audio tokenizer built on a modified DAC architecture with continuous latent representation. At 166.9M parameters and 64 latent channels, it beats larger competing models on text‑to‑audio quality across AudioSet, music, and speech benchmarks. Code and weights under MIT license.
Read →Our research team presented at the International Conference on Machine Learning at COEX, Seoul — sharing posters and talks on the lab’s latest generative modelling work.
Read →Our paper was accepted at the DeLTa Workshop on Deep Generative Models — a Bridge Matching approach for temporally-correlated data such as video, outperforming baselines on frame interpolation, image‑to‑video and video super-resolution.
Read →Two new models in 4×8×8 and 4×16×16 compression, outperforming Wan 2.2 and HunyuanVideo 1.5 on objective metrics. Code and weights open on GitHub.
Read →Three video generation models fine‑tuned on Kandinsky 5.0 Video Lite (2B parameters) for autonomous driving, robotics, and complex physics, with post‑training RL and SOAR. Ranks top‑4 on PAI‑Bench‑G, top‑5 on Physics‑IQ Verified, and top‑6 on RBench. Open source under MIT license.
Read →The new flagship, succeeding Kandinsky 5.0, works up to twice as fast, better understands complex queries, and creates more detailed images. Built on a new Mixture‑of‑Experts architecture, it brings free professional-level editing — object removal, restyling, retouching — with no generation limits.
Read →A 48 kHz audio tokenizer built on a modified DAC architecture with continuous latent representation. At 166.9M parameters and 64 latent channels, it beats larger competing models on text‑to‑audio quality across AudioSet, music, and speech benchmarks. Code and weights under MIT license.
Read →Our research team presented at the International Conference on Machine Learning at COEX, Seoul — sharing posters and talks on the lab’s latest generative modelling work.
Read →Our paper was accepted at the DeLTa Workshop on Deep Generative Models — a Bridge Matching approach for temporally-correlated data such as video, outperforming baselines on frame interpolation, image‑to‑video and video super-resolution.
Read →Two new models in 4×8×8 and 4×16×16 compression, outperforming Wan 2.2 and HunyuanVideo 1.5 on objective metrics. Code and weights open on GitHub.
Read →Three video generation models fine‑tuned on Kandinsky 5.0 Video Lite (2B parameters) for autonomous driving, robotics, and complex physics, with post‑training RL and SOAR. Ranks top‑4 on PAI‑Bench‑G, top‑5 on Physics‑IQ Verified, and top‑6 on RBench. Open source under MIT license.
Read →The new flagship, succeeding Kandinsky 5.0, works up to twice as fast, better understands complex queries, and creates more detailed images. Built on a new Mixture‑of‑Experts architecture, it brings free professional-level editing — object removal, restyling, retouching — with no generation limits.
Read →A 48 kHz audio tokenizer built on a modified DAC architecture with continuous latent representation. At 166.9M parameters and 64 latent channels, it beats larger competing models on text‑to‑audio quality across AudioSet, music, and speech benchmarks. Code and weights under MIT license.
Read →Our research team presented at the International Conference on Machine Learning at COEX, Seoul — sharing posters and talks on the lab’s latest generative modelling work.
Read →Our paper was accepted at the DeLTa Workshop on Deep Generative Models — a Bridge Matching approach for temporally-correlated data such as video, outperforming baselines on frame interpolation, image‑to‑video and video super-resolution.
Read →Two new models in 4×8×8 and 4×16×16 compression, outperforming Wan 2.2 and HunyuanVideo 1.5 on objective metrics. Code and weights open on GitHub.
Read →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
