Kandinsky 6

<div align="center"> <img src="assets/readme/promo.webp" width="100%"> </div> <br> <div align="center"> <picture> <source media="(prefers-color-scheme: dark)" srcset="assets/readme/Kandinsky_LOGO_6_Horizontal_white.png"> <img src="assets/readme/Kandinsky_LOGO_6_Horizontal_black.png" width="60%"> </picture> </div> <br> <div align="center">

<a href="https://kandinskylab.ai/"><img alt="KandinskyLab" src="https://img.shields.io/badge/KandinskyLab-76E0B7?style=for-the-badge"></a> <a href="https://arxiv.org/abs/2610.05608"><img alt="Report" src="https://img.shields.io/badge/Report-9C2731?style=for-the-badge"></a> <a href="https://huggingface.co/docs/diffusers/main/en/api/pipelines/kandinsky6"><img alt="Diffusers" src="https://img.shields.io/badge/Diffusers-F8D44E?style=for-the-badge"></a> <a href="https://huggingface.co/spaces/kandinskylab/Kandinsky-6.0-Pro-distill-5s"><img alt="HF Demo" src="https://img.shields.io/badge/HF%20Demo-D9622B?style=for-the-badge"></a> <a href="https://registry.comfy.org/nodes/kandinsky6"><img alt="ComfyUI" src="https://img.shields.io/badge/ComfyUI-EBF764?style=for-the-badge"></a> <a href="https://docs.vllm.ai/projects/vllm-omni/en/latest/api/vllm_omni/diffusion/models/kandinsky6/"><img alt="vLL-Omni" src="https://img.shields.io/badge/vLLM Omni-54A0F8?style=for-the-badge"></a>

<!-- <a href="#"><img alt="PyPI" src="https://img.shields.io/badge/PyPI-3171B2?style=for-the-badge"></a> --> <!-- <a href="#"><img alt="SGLang" src="https://img.shields.io/badge/SGLang-C6602D?style=for-the-badge"></a> <a href="#"><img alt="FastVideo" src="https://img.shields.io/badge/FastVideo-436BF6?style=for-the-badge"></a> --> </div> <h1>Kandinsky 6.0: A family of diffusion models for Video + Audio generation</h1>

We present Kandinsky 6.0 Video, a family of foundation diffusion models for synchronized text-to-audio-video generation, comprising Kandinsky 6.0 Video Lite (3B parameters) and Kandinsky 6.0 Video Pro (29B parameters). Both models generate 5-second video clips with synchronized 44 kHz audio, including lip-sync, in text-to-audio-video (T2AV) and image-to-audio-video (TI2AV) modes; a plugin-in super-resolution model raises the output resolution to Full-HD (1920×1080).

Project Updates

• 2026/10/06: We added vLLM-omni support
• 2026/10/06: We added Hugging Face Space for Kandinsky 6.0 Pro Distill
• 2026/10/06: We have open-sourced Kandinsky 6.0

Quick start

An NVIDIA GPU and Python 3.13 or 3.14.

The first run downloads Kandinsky-6.0-Pro-distill-5s into $KANDINSKY_HOME/weights. When $KANDINSKY_HOME is unset, that directory is ~/.cache/kandinsky. The clip is written to $KANDINSKY_HOME/outputs/generate_<YYYY-MM-DDTHH-MM-SS>/generations/output.mp4, with the expanded prompt in expanded_prompt.txt next to launch.json, logs/, and profiles/. The generate command downloads the checkpoint named in the config if it is not already on disk. Other catalog names are pro, pro-pretrain, lite, lite-distill, and lite-pretrain.

Clone the repository

You need uv and just.

git clone https://github.com/kandinskylab/kandinsky-6.git
cd k6_video
just setup

just setup reads the GPU and installs one PyTorch build:

| GPUs | Compute capability | Installed | |---|---|---| | Hopper (H100, H200) | 9.0 | PyTorch for CUDA 13.0, FlashAttention 3 | | Ampere, Ada, consumer Blackwell | 8.0, 8.6, 8.9, 12.0, 12.1 | PyTorch for CUDA 13.0, then SageAttention 2.2.0 |

Ampere, Ada, and consumer Blackwell compile SageAttention, so nvcc has to be on PATH. CUDA_HOME defaults to /usr/local/cuda.

just download pro-distill
just generate "a cat on a mat"

Another GPU uses its preset:

just generate "a cat on a mat" --config kandinsky/configs/devices/rtx-5090.yaml --out clip.mp4

Presets live in kandinsky/configs/devices/.

ComfyUI

For ComfyUI, install kandinsky6 and kandinsky6-sr through ComfyUI Manager, then restart ComfyUI. The comfyui/ directory contains extension source code; no manual copying is needed — see the setup guide.

Performance

Working time (s) for a 5-second clip on the non-distilled model, after warmup. Weight loading and MP4 encoding are excluded.

RTX PRO 6000 (96 GB), A100 80 GB, and H100 80 GB use module offload. The consumer cards use block offload.

RTX 4090, RTX 5060 Ti, RTX 5080, RTX 5090, RTX PRO 6000, and A100 80 GB use SageAttention2++ for visual self-attention. The H100 uses FlashAttention 3.

| | RTX 4090 | RTX 5060 Ti | RTX 5080 | RTX 5090 | RTX PRO 6000 | A100 80 GB | H100 | |---|---:|---:|---:|---:|---:|---:|---:| | Lite SD | 437 | 1310 | 577 | 309 | 242 | 532 | 239 | | Lite HD | 422 | 1328 | 579 | 296 | 274 | 472 | 203 | | Lite Full HD | 578 | 1774 | 770 | 406 | 387 | 664 | 284 | | Pro SD | 936 | 3080 | 1336 | 754 | 621 | 972 | 356 | | Pro HD | 1194 | 2716 | 1189 | 638 | 561 | 791 | 292 | | Pro Full HD | 1247 | 3530 | 1546 | 854 | 765 | 1106 | 402 |